1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794 4795 4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807 4808 4809 4810 4811 4812 4813 4814 4815 4816 4817 4818 4819 4820 4821 4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843 4844 4845 4846 4847 4848 4849 4850 4851 4852 4853 4854 4855 4856 4857 4858 4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872 4873 4874 4875 4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925 4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976 4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987 4988 4989 4990 4991 4992 4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050 5051 5052 5053 5054 5055 5056 5057 5058 5059 5060 5061 5062 5063 5064 5065 5066 5067 5068 5069 5070 5071 5072 5073 5074 5075 5076 5077 5078 5079 5080 5081 5082 5083 5084 5085 5086 5087 5088 5089 5090 5091 5092 5093 5094 5095 5096 5097 5098 5099 5100 5101 5102 5103 5104 5105 5106 5107 5108 5109 5110 5111 5112 5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123 5124 5125 5126 5127 5128 5129 5130 5131 5132 5133 5134 5135 5136 5137 5138 5139 5140 5141 5142 5143 5144 5145 5146 5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209 5210 5211 5212 5213 5214 5215 5216 5217 5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230 5231 5232 5233 5234 5235 5236 5237 5238 5239 5240 5241 5242 5243 5244 5245 5246 5247 5248 5249 5250 5251 5252 5253 5254 5255 5256 5257 5258 5259 5260 5261 5262 5263 5264 5265 5266 5267 5268 5269 5270 5271 5272 5273 5274 5275 5276 5277 5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292 5293 5294 5295 5296 5297 5298 5299 5300 5301 5302 5303 5304 5305 5306 5307 5308 5309 5310 5311 5312 5313 5314 5315 5316 5317 5318 5319 5320 5321 5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340 5341 5342 5343 5344 5345 5346 5347 5348 5349 5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436 5437 5438 5439 5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480 5481 5482 5483 5484 5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 5512 5513 5514 5515 5516 5517 5518 5519 5520 5521 5522 5523 5524 5525 5526 5527 5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557 5558 5559 5560 5561 5562 5563 5564 5565 5566 5567 5568 5569 5570 5571 5572 5573 5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584 5585 5586 5587 5588 5589 5590 5591 5592 5593 5594 5595 5596 5597 5598 5599 5600 5601 5602 5603 5604 5605 5606 5607 5608 5609 5610 5611 5612 5613 5614 5615 5616 5617 5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632 5633 5634 5635 5636 5637 5638 5639 5640 5641 5642 5643 5644 5645 5646 5647 5648 5649 5650 5651 5652 5653 5654 5655 5656 5657 5658 5659 5660 5661 5662 5663 5664 5665 5666 5667 5668 5669 5670 5671 5672 5673 5674 5675 5676 5677 5678 5679 5680 5681 5682 5683 5684 5685 5686 5687 5688 5689 5690 5691 5692 5693 5694 5695 5696 5697 5698 5699 5700 5701 5702 5703 5704 5705 5706 5707 5708 5709 5710 5711 5712 5713 5714 5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735 5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751 5752 5753 5754 5755 5756 5757 5758 5759 5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771 5772 5773 5774 5775 5776 5777 5778 5779 5780 5781 5782 5783 5784 5785 5786 5787 5788 5789 5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808 5809 5810 5811 5812 5813 5814 5815 5816 5817 5818 5819 5820 5821 5822 5823 5824 5825 5826 5827 5828 5829 5830 5831 5832 5833 5834 5835 5836 5837 5838 5839 5840 5841 5842 5843 5844 5845 5846 5847 5848 5849 5850 5851 5852 5853 5854 5855 5856 5857 5858 5859 5860 5861 5862 5863 5864 5865 5866 5867 5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880 5881 5882 5883 5884 5885 5886 5887 5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905 5906 5907 5908 5909 5910 5911 5912 5913 5914 5915 5916 5917 5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929 5930 5931 5932 5933 5934 5935 5936 5937 5938 5939 5940 5941 5942 5943 5944 5945 5946 5947 5948 5949 5950 5951 5952 5953 5954 5955 5956 5957 5958 5959 5960 5961 5962 5963 5964 5965 5966 5967 5968 5969 5970 5971 5972 5973 5974 5975 5976 5977 5978 5979 5980 5981 5982 5983 5984 5985 5986 5987 5988 5989 5990 5991 5992 5993 5994 5995 5996 5997 5998 5999 6000 6001 6002 6003 6004 6005 6006 6007 6008 6009 6010 6011 6012 6013 6014 6015 6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030 6031 6032 6033 6034 6035 6036 6037 6038 6039 6040 6041 6042 6043 6044 6045 6046 6047 6048 6049 6050 6051 6052 6053 6054 6055 6056 6057 6058 6059 6060 6061 6062 6063 6064 6065 6066 6067 6068 6069 6070 6071 6072 6073 6074 6075 6076 6077 6078 6079 6080 6081 6082 6083 6084 6085 6086 6087 6088 6089 6090 6091 6092 6093 6094 6095 6096 6097 6098 6099 6100 6101 6102 6103 6104 6105 6106 6107 6108 6109 6110 6111 6112 6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126 6127 6128 6129 6130 6131 6132 6133 6134 6135 6136 6137 6138 6139 6140 6141 6142 6143 6144 6145 6146 6147 6148 6149 6150 6151 6152 6153 6154 6155 6156 6157 6158 6159 6160 6161 6162 6163 6164 6165 6166 6167 6168 6169 6170 6171 6172 6173 6174 6175 6176 6177 6178 6179 6180 6181 6182 6183 6184 6185 6186 6187 6188 6189 6190 6191 6192 6193 6194 6195 6196 6197 6198 6199 6200 6201 6202 6203 6204 6205 6206 6207 6208 6209 6210 6211 6212 6213 6214 6215 6216 6217 6218 6219 6220 6221 6222 6223 6224 6225 6226 6227 6228 6229 6230 6231 6232 6233 6234 6235 6236 6237 6238 6239 6240 6241 6242 6243 6244 6245 6246 6247 6248 6249 6250 6251 6252 6253 6254 6255 6256 6257 6258 6259 6260 6261 6262 6263 6264 6265 6266 6267 6268 6269 6270 6271 6272 6273 6274 6275 6276 6277 6278 6279 6280 6281 6282 6283 6284 6285 6286 6287 6288 6289 6290 6291 6292 6293 6294 6295 6296 6297 6298 6299 6300 6301 6302 6303 6304 6305 6306 6307 6308 6309 6310 6311 6312 6313 6314 6315 6316 6317 6318 6319 6320 6321 6322 6323 6324 6325 6326 6327 6328 6329 6330 6331 6332 6333 6334 6335 6336 6337 6338 6339 6340 6341 6342 6343 6344 6345 6346 6347 6348 6349 6350 6351 6352 6353 6354 6355 6356 6357 6358 6359 6360 6361 6362 6363 6364 6365 6366 6367 6368 6369 6370 6371 6372 6373 6374 6375 6376 6377 6378 6379 6380 6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422 6423 6424 6425 6426 6427 6428 6429 6430 6431 6432 6433 6434 6435 6436 6437 6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463 6464 6465 6466 6467 6468 6469 6470 6471 6472 6473 6474 6475 6476 6477 6478 6479 6480 6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522 6523 6524 6525 6526 6527 6528 6529 6530 6531 6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544 6545 6546 6547 6548 6549 6550 6551 6552 6553 6554 6555 6556 6557 6558 6559 6560 6561 6562 6563 6564 6565 6566 6567 6568 6569 6570 6571 6572 6573 6574 6575 6576 6577 6578 6579 6580 6581 6582 6583 6584 6585 6586 6587 6588 6589 6590 6591 6592 6593 6594 6595 6596 6597 6598 6599 6600 6601 6602 6603 6604 6605 6606 6607 6608 6609 6610 6611 6612 6613 6614 6615 6616 6617 6618 6619 6620 6621 6622 6623 6624 6625 6626 6627 6628 6629 6630 6631 6632 6633 6634 6635 6636 6637 6638 6639 6640 6641 6642 6643 6644 6645 6646 6647 6648 6649 6650 6651 6652 6653 6654 6655 6656 6657 6658 6659 6660 6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683 6684 6685 6686 6687 6688 6689 6690 6691 6692 6693 6694 6695 6696 6697 6698 6699 6700 6701 6702 6703 6704 6705 6706 6707 6708 6709 6710 6711 6712 6713 6714 6715 6716 6717 6718 6719 6720 6721 6722 6723 6724 6725 6726 6727 6728 6729 6730 6731 6732 6733 6734 6735 6736 6737 6738 6739 6740 6741 6742 6743 6744 6745 6746 6747 6748 6749 6750 6751 6752 6753 6754 6755 6756 6757 6758 6759 6760 6761 6762 6763 6764 6765 6766 6767 6768 6769 6770 6771 6772 6773 6774 6775 6776 6777 6778 6779 6780 6781 6782 6783 6784 6785 6786 6787 6788 6789 6790 6791 6792 6793 6794 6795 6796 6797 6798 6799 6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812 6813 6814 6815 6816 6817 6818 6819 6820 6821 6822 6823 6824 6825 6826 6827 6828 6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840 6841 6842 6843 6844 6845 6846 6847 6848 6849 6850 6851 6852 6853 6854 6855 6856 6857 6858 6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870 6871 6872 6873 6874 6875 6876 6877 6878 6879 6880 6881 6882 6883 6884 6885 6886 6887 6888 6889 6890 6891 6892 6893 6894 6895 6896 6897 6898 6899 6900 6901 6902 6903 6904 6905 6906 6907 6908 6909 6910 6911 6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943 6944 6945 6946 6947 6948 6949 6950 6951 6952 6953 6954 6955 6956 6957 6958 6959 6960 6961 6962 6963 6964 6965 6966 6967 6968 6969 6970 6971 6972 6973 6974 6975 6976 6977 6978 6979 6980 6981 6982 6983 6984 6985 6986 6987 6988 6989 6990 6991 6992 6993 6994 6995 6996 6997 6998 6999 7000 7001 7002 7003 7004 7005 7006 7007 7008 7009 7010 7011 7012 7013 7014 7015 7016 7017 7018 7019 7020 7021 7022 7023 7024 7025 7026 7027 7028 7029 7030 7031 7032 7033 7034 7035 7036 7037 7038 7039 7040 7041 7042 7043 7044 7045 7046 7047 7048 7049 7050 7051 7052 7053 7054 7055 7056 7057 7058 7059 7060 7061 7062 7063 7064 7065 7066 7067 7068 7069 7070 7071 7072 7073 7074 7075 7076 7077 7078 7079 7080 7081 7082 7083 7084 7085 7086 7087 7088 7089 7090 7091 7092 7093 7094 7095 7096 7097 7098 7099 7100 7101 7102 7103 7104 7105 7106 7107 7108 7109 7110 7111 7112 7113 7114 7115 7116 7117 7118 7119 7120 7121 7122 7123 7124 7125 7126 7127 7128 7129 7130 7131 7132 7133 7134 7135 7136 7137 7138 7139 7140 7141 7142 7143 7144 7145 7146 7147 7148 7149 7150 7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161 7162 7163 7164 7165 7166 7167 7168 7169 7170 7171 7172 7173 7174 7175 7176 7177 7178 7179 7180 7181 7182 7183 7184 7185 7186 7187 7188 7189 7190 7191 7192 7193 7194 7195 7196 7197 7198 7199 7200 7201 7202 7203 7204 7205 7206 7207 7208 7209 7210 7211 7212 7213 7214 7215 7216 7217 7218 7219 7220 7221 7222 7223 7224 7225 7226 7227 7228 7229 7230 7231 7232 7233 7234 7235 7236 7237 7238 7239 7240 7241 7242 7243 7244 7245 7246 7247 7248 7249 7250 7251 7252 7253 7254 7255 7256 7257 7258 7259 7260 7261 7262 7263 7264 7265 7266 7267 7268 7269 7270 7271 7272 7273 7274 7275 7276 7277 7278 7279 7280 7281 7282 7283 7284 7285 7286 7287 7288 7289 7290 7291 7292 7293 7294 7295 7296 7297 7298 7299 7300 7301 7302 7303 7304 7305 7306 7307 7308 7309 7310 7311 7312 7313 7314 7315 7316 7317 7318 7319 7320 7321 7322 7323 7324 7325 7326 7327 7328 7329 7330 7331 7332 7333 7334 7335 7336 7337 7338 7339 7340 7341 7342 7343 7344 7345 7346 7347 7348 7349 7350 7351 7352 7353 7354 7355 7356 7357 7358 7359 7360 7361 7362 7363 7364 7365 7366 7367 7368 7369 7370 7371 7372 7373 7374 7375 7376 7377 7378 7379 7380 7381 7382 7383 7384 7385 7386 7387 7388 7389 7390 7391 7392 7393 7394 7395 7396 7397 7398 7399 7400 7401 7402 7403 7404 7405 7406 7407 7408 7409 7410 7411 7412 7413 7414 7415 7416 7417 7418 7419 7420 7421 7422 7423 7424 7425 7426 7427 7428 7429 7430 7431 7432 7433 7434 7435 7436 7437 7438 7439 7440 7441 7442 7443 7444 7445 7446 7447 7448 7449 7450 7451 7452 7453 7454 7455 7456 7457 7458 7459 7460 7461 7462 7463 7464 7465 7466 7467 7468 7469 7470 7471 7472 7473 7474 7475 7476 7477 7478 7479 7480 7481 7482 7483 7484 7485 7486 7487 7488 7489 7490 7491 7492 7493 7494 7495 7496 7497 7498 7499 7500 7501 7502 7503 7504 7505 7506 7507 7508 7509 7510 7511 7512 7513 7514 7515 7516 7517 7518 7519 7520 7521 7522 7523 7524 7525 7526 7527 7528 7529 7530 7531 7532 7533 7534 7535 7536 7537 7538 7539 7540 7541 7542 7543 7544 7545 7546 7547 7548 7549 7550 7551 7552 7553 7554 7555 7556 7557 7558 7559 7560 7561 7562 7563 7564 7565 7566 7567 7568 7569 7570 7571 7572 7573 7574 7575 7576 7577 7578 7579 7580 7581 7582 7583 7584 7585 7586 7587 7588 7589 7590 7591 7592 7593 7594 7595 7596 7597 7598 7599 7600 7601 7602 7603 7604 7605 7606 7607 7608 7609 7610 7611 7612 7613 7614 7615 7616 7617 7618 7619 7620 7621 7622 7623 7624 7625 7626 7627 7628 7629 7630 7631 7632 7633 7634 7635 7636 7637 7638 7639 7640 7641 7642 7643 7644 7645 7646 7647 7648 7649 7650 7651 7652 7653 7654 7655 7656 7657 7658 7659 7660 7661 7662 7663 7664 7665 7666 7667 7668 7669 7670 7671 7672 7673 7674 7675 7676 7677 7678 7679 7680 7681 7682 7683 7684 7685 7686 7687 7688 7689 7690 7691 7692 7693 7694 7695 7696 7697 7698 7699 7700 7701 7702 7703 7704 7705 7706 7707 7708 7709 7710 7711 7712 7713 7714 7715 7716 7717 7718 7719 7720 7721 7722 7723 7724 7725 7726 7727 7728 7729 7730 7731 7732 7733 7734 7735 7736 7737 7738 7739 7740 7741 7742 7743 7744 7745 7746 7747 7748 7749 7750 7751 7752 7753 7754 7755 7756 7757 7758 7759 7760 7761 7762 7763 7764 7765 7766 7767 7768 7769 7770 7771 7772 7773 7774 7775 7776 7777 7778 7779 7780 7781 7782 7783 7784 7785 7786 7787 7788 7789 7790 7791 7792 7793 7794 7795 7796 7797 7798 7799 7800 7801 7802 7803 7804 7805 7806 7807 7808 7809 7810 7811 7812 7813 7814 7815 7816 7817 7818 7819 7820 7821 7822 7823 7824 7825 7826 7827 7828 7829 7830 7831 7832 7833 7834 7835 7836 7837 7838 7839 7840 7841 7842 7843 7844 7845 7846 7847 7848 7849 7850 7851 7852 7853 7854 7855 7856 7857 7858 7859 7860 7861 7862 7863 7864 7865 7866 7867 7868 7869 7870 7871 7872 7873 7874 7875 7876 7877 7878 7879 7880 7881 7882 7883 7884 7885 7886 7887 7888 7889 7890 7891 7892 7893 7894 7895 7896 7897 7898 7899 7900 7901 7902 7903 7904 7905 7906 7907 7908 7909 7910 7911 7912 7913 7914 7915 7916 7917 7918 7919 7920 7921 7922 7923 7924 7925 7926 7927 7928 7929 7930 7931 7932 7933 7934 7935 7936 7937 7938 7939 7940 7941 7942 7943 7944 7945 7946 7947 7948 7949 7950 7951 7952 7953 7954 7955 7956 7957 7958 7959 7960 7961 7962 7963 7964 7965 7966 7967 7968 7969 7970 7971 7972 7973 7974 7975 7976 7977 7978 7979 7980 7981 7982 7983 7984 7985 7986 7987 7988 7989 7990 7991 7992 7993 7994 7995 7996 7997 7998 7999 8000 8001 8002 8003 8004 8005 8006 8007 8008 8009 8010 8011 8012 8013 8014 8015 8016 8017 8018 8019 8020 8021 8022 8023 8024 8025 8026 8027 8028 8029 8030 8031 8032 8033 8034 8035 8036 8037 8038 8039 8040 8041 8042 8043 8044 8045 8046 8047 8048 8049 8050 8051 8052 8053 8054 8055 8056 8057 8058 8059 8060 8061 8062 8063 8064 8065 8066 8067 8068 8069 8070 8071 8072 8073 8074 8075 8076 8077 8078 8079 8080 8081 8082 8083 8084 8085 8086 8087 8088 8089 8090 8091 8092 8093 8094 8095 8096 8097 8098 8099 8100 8101 8102 8103 8104 8105 8106 8107 8108 8109 8110 8111 8112 8113 8114 8115 8116 8117 8118 8119 8120 8121 8122 8123 8124 8125 8126 8127 8128 8129 8130 8131 8132 8133 8134 8135 8136 8137 8138 8139 8140 8141 8142 8143 8144 8145 8146 8147 8148 8149 8150 8151 8152 8153 8154 8155 8156 8157 8158 8159 8160 8161 8162 8163 8164 8165 8166 8167 8168
|
from __future__ import annotations
import asyncio
import contextlib
import dataclasses
import heapq
import inspect
import itertools
import json
import logging
import math
import operator
import os
import pickle
import random
import sys
import uuid
import warnings
import weakref
from collections import defaultdict, deque
from collections.abc import (
Callable,
Collection,
Container,
Hashable,
Iterable,
Iterator,
Mapping,
Sequence,
Set,
)
from contextlib import suppress
from functools import partial
from numbers import Number
from typing import TYPE_CHECKING, Any, ClassVar, Literal, NamedTuple, cast, overload
import psutil
from sortedcontainers import SortedDict, SortedSet
from tlz import (
first,
groupby,
merge,
merge_sorted,
merge_with,
partition,
pluck,
second,
valmap,
)
from tornado.ioloop import IOLoop
import dask
from dask.highlevelgraph import HighLevelGraph
from dask.utils import (
format_bytes,
format_time,
key_split,
parse_bytes,
parse_timedelta,
tmpfile,
)
from dask.widgets import get_template
from distributed import cluster_dump, preloading, profile
from distributed import versions as version_module
from distributed._stories import scheduler_story
from distributed.active_memory_manager import ActiveMemoryManagerExtension, RetireWorker
from distributed.batched import BatchedSend
from distributed.collections import HeapSet
from distributed.comm import (
Comm,
CommClosedError,
get_address_host,
normalize_address,
resolve_address,
unparse_host_port,
)
from distributed.comm.addressing import addresses_from_user_args
from distributed.compatibility import PeriodicCallback
from distributed.core import Status, clean_exception, rpc, send_recv
from distributed.diagnostics.memory_sampler import MemorySamplerExtension
from distributed.diagnostics.plugin import SchedulerPlugin, _get_plugin_name
from distributed.event import EventExtension
from distributed.http import get_handlers
from distributed.lock import LockExtension
from distributed.metrics import monotonic, time
from distributed.multi_lock import MultiLockExtension
from distributed.node import ServerNode
from distributed.proctitle import setproctitle
from distributed.protocol.pickle import dumps, loads
from distributed.protocol.serialize import Serialized, serialize
from distributed.publish import PublishExtension
from distributed.pubsub import PubSubSchedulerExtension
from distributed.queues import QueueExtension
from distributed.recreate_tasks import ReplayTaskScheduler
from distributed.security import Security
from distributed.semaphore import SemaphoreExtension
from distributed.shuffle import ShuffleSchedulerExtension
from distributed.stealing import WorkStealing
from distributed.utils import (
All,
TimeoutError,
empty_context,
get_fileno_limit,
key_split_group,
log_errors,
no_default,
recursive_to_dict,
validate_key,
)
from distributed.utils_comm import (
gather_from_workers,
retry_operation,
scatter_to_workers,
)
from distributed.utils_perf import disable_gc_diagnosis, enable_gc_diagnosis
from distributed.variable import VariableExtension
if TYPE_CHECKING:
# TODO import from typing (requires Python >=3.10)
from typing_extensions import TypeAlias
# TODO move out of TYPE_CHECKING (requires Python >=3.10)
# Not to be confused with distributed.worker_state_machine.TaskStateState
TaskStateState: TypeAlias = Literal[
"released",
"waiting",
"no-worker",
"queued",
"processing",
"memory",
"erred",
"forgotten",
]
# TODO remove quotes (requires Python >=3.9)
# {task key -> finish state}
# Not to be confused with distributed.worker_state_machine.Recs
Recs: TypeAlias = "dict[str, TaskStateState]"
# {client or worker address: [{op: <key>, ...}, ...]}
Msgs: TypeAlias = "dict[str, list[dict[str, Any]]]"
# (recommendations, client messages, worker messages)
RecsMsgs: TypeAlias = "tuple[Recs, Msgs, Msgs]"
else:
TaskStateState = str
Recs = dict
Msgs = dict
RecsMsgs = tuple
ALL_TASK_STATES: set[TaskStateState] = {
"released",
"waiting",
"no-worker",
"queued",
"processing",
"memory",
"erred",
"forgotten",
}
logger = logging.getLogger(__name__)
LOG_PDB = dask.config.get("distributed.admin.pdb-on-err")
DEFAULT_DATA_SIZE = parse_bytes(
dask.config.get("distributed.scheduler.default-data-size")
)
STIMULUS_ID_UNSET = "<stimulus_id unset>"
DEFAULT_EXTENSIONS = {
"locks": LockExtension,
"multi_locks": MultiLockExtension,
"publish": PublishExtension,
"replay-tasks": ReplayTaskScheduler,
"queues": QueueExtension,
"variables": VariableExtension,
"pubsub": PubSubSchedulerExtension,
"semaphores": SemaphoreExtension,
"events": EventExtension,
"amm": ActiveMemoryManagerExtension,
"memory_sampler": MemorySamplerExtension,
"shuffle": ShuffleSchedulerExtension,
"stealing": WorkStealing,
}
class ClientState:
"""A simple object holding information about a client."""
#: A unique identifier for this client. This is generally an opaque
#: string generated by the client itself.
client_key: str
#: Cached hash of :attr:`~ClientState.client_key`
_hash: int
#: A set of tasks this client wants to be kept in memory, so that it can download
#: its result when desired. This is the reverse mapping of
#: :class:`TaskState.who_wants`. Tasks are typically removed from this set when the
#: corresponding object in the client's space (for example a ``Future`` or a Dask
#: collection) gets garbage-collected.
wants_what: set[TaskState]
#: The last time we received a heartbeat from this client, in local scheduler time.
last_seen: float
#: Output of :func:`distributed.versions.get_versions` on the client
versions: dict[str, Any]
__slots__ = tuple(__annotations__)
def __init__(self, client: str, *, versions: dict[str, Any] | None = None):
self.client_key = client
self._hash = hash(client)
self.wants_what = set()
self.last_seen = time()
self.versions = versions or {}
def __hash__(self) -> int:
return self._hash
def __eq__(self, other: object) -> bool:
if not isinstance(other, ClientState):
return False
return self.client_key == other.client_key
def __repr__(self) -> str:
return f"<Client {self.client_key!r}>"
def __str__(self) -> str:
return self.client_key
def _to_dict_no_nest(self, *, exclude: Container[str] = ()) -> dict:
"""Dictionary representation for debugging purposes.
Not type stable and not intended for roundtrips.
See also
--------
Client.dump_cluster_state
distributed.utils.recursive_to_dict
TaskState._to_dict
"""
return recursive_to_dict(
self,
exclude=set(exclude) | {"versions"}, # type: ignore
members=True,
)
class MemoryState:
"""Memory readings on a worker or on the whole cluster.
See :doc:`worker-memory`.
Attributes / properties:
managed
Sum of the output of sizeof() for all dask keys held by the worker in memory,
plus number of bytes spilled to disk
managed_in_memory
Sum of the output of sizeof() for the dask keys held in RAM. Note that this may
be inaccurate, which may cause inaccurate unmanaged memory (see below).
managed_spilled
Number of bytes for the dask keys spilled to the hard drive.
Note that this is the size on disk; size in memory may be different due to
compression and inaccuracies in sizeof(). In other words, given the same keys,
'managed' will change depending if the keys are in memory or spilled.
process
Total RSS memory measured by the OS on the worker process.
This is always exactly equal to managed_in_memory + unmanaged.
unmanaged
process - managed_in_memory. This is the sum of
- Python interpreter and modules
- global variables
- memory temporarily allocated by the dask tasks that are currently running
- memory fragmentation
- memory leaks
- memory not yet garbage collected
- memory not yet free()'d by the Python memory manager to the OS
unmanaged_old
Minimum of the 'unmanaged' measures over the last
``distributed.memory.recent-to-old-time`` seconds
unmanaged_recent
unmanaged - unmanaged_old; in other words process memory that has been recently
allocated but is not accounted for by dask; hopefully it's mostly a temporary
spike.
optimistic
managed_in_memory + unmanaged_old; in other words the memory held long-term by
the process under the hopeful assumption that all unmanaged_recent memory is a
temporary spike
.. note::
There is an intentional misalignment in terminology between this class (which is
meant for internal / programmatic use) and the memory readings on the GUI (which
is aimed at the general public:
================= =====================
MemoryState GUI
================= =====================
managed n/a
managed_in_memory managed
managed_spilled spilled
process process (RSS); memory
unmanaged n/a
unmanaged_old unmanaged (old)
unmanaged_recent unmanaged (recent)
optimistic n/a
================= =====================
"""
process: int
unmanaged_old: int
managed_in_memory: int
managed_spilled: int
__slots__ = tuple(__annotations__)
def __init__(
self,
*,
process: int,
unmanaged_old: int,
managed_in_memory: int,
managed_spilled: int,
):
# Some data arrives with the heartbeat, some other arrives in realtime as the
# tasks progress. Also, sizeof() is not guaranteed to return correct results.
# This can cause glitches where a partial measure is larger than the whole, so
# we need to force all numbers to add up exactly by definition.
self.process = process
self.managed_in_memory = min(self.process, managed_in_memory)
self.managed_spilled = managed_spilled
# Subtractions between unsigned ints guaranteed by construction to be >= 0
self.unmanaged_old = min(unmanaged_old, process - self.managed_in_memory)
@staticmethod
def sum(*infos: MemoryState) -> MemoryState:
process = 0
unmanaged_old = 0
managed_in_memory = 0
managed_spilled = 0
for ms in infos:
process += ms.process
unmanaged_old += ms.unmanaged_old
managed_spilled += ms.managed_spilled
managed_in_memory += ms.managed_in_memory
return MemoryState(
process=process,
unmanaged_old=unmanaged_old,
managed_in_memory=managed_in_memory,
managed_spilled=managed_spilled,
)
@property
def managed(self) -> int:
return self.managed_in_memory + self.managed_spilled
@property
def unmanaged(self) -> int:
# This is never negative thanks to __init__
return self.process - self.managed_in_memory
@property
def unmanaged_recent(self) -> int:
# This is never negative thanks to __init__
return self.process - self.managed_in_memory - self.unmanaged_old
@property
def optimistic(self) -> int:
return self.managed_in_memory + self.unmanaged_old
def __repr__(self) -> str:
return (
f"Process memory (RSS) : {format_bytes(self.process)}\n"
f" - managed by Dask : {format_bytes(self.managed_in_memory)}\n"
f" - unmanaged (old) : {format_bytes(self.unmanaged_old)}\n"
f" - unmanaged (recent): {format_bytes(self.unmanaged_recent)}\n"
f"Spilled to disk : {format_bytes(self.managed_spilled)}\n"
)
def _to_dict(self, *, exclude: Container[str] = ()) -> dict:
"""Dictionary representation for debugging purposes.
Not type stable and not intended for roundtrips.
See also
--------
Client.dump_cluster_state
distributed.utils.recursive_to_dict
"""
return recursive_to_dict(self, exclude=exclude, members=True)
class WorkerState:
"""A simple object holding information about a worker.
Not to be confused with :class:`distributed.worker_state_machine.WorkerState`.
"""
#: This worker's unique key. This can be its connected address
#: (such as ``"tcp://127.0.0.1:8891"``) or an alias (such as ``"alice"``).
address: str
pid: int
name: Hashable
#: The number of CPU threads made available on this worker
nthreads: int
#: Memory available to the worker, in bytes
memory_limit: int
local_directory: str
services: dict[str, int]
#: Output of :meth:`distributed.versions.get_versions` on the worker
versions: dict[str, Any]
#: Address of the associated :class:`~distributed.nanny.Nanny`, if present
nanny: str
#: Read-only worker status, synced one way from the remote Worker object
status: Status
#: Cached hash of :attr:`~WorkerState.address`
_hash: int
#: The total memory size, in bytes, used by the tasks this worker holds in memory
#: (i.e. the tasks in this worker's :attr:`~WorkerState.has_what`).
nbytes: int
#: Worker memory unknown to the worker, in bytes, which has been there for more than
#: 30 seconds. See :class:`MemoryState`.
_memory_unmanaged_old: int
#: History of the last 30 seconds' worth of unmanaged memory. Used to differentiate
#: between "old" and "new" unmanaged memory.
#: Format: ``[(timestamp, bytes), (timestamp, bytes), ...]``
_memory_unmanaged_history: deque[tuple[float, int]]
metrics: dict[str, Any]
#: The last time we received a heartbeat from this worker, in local scheduler time.
last_seen: float
time_delay: float
bandwidth: float
#: A set of all TaskStates on this worker that are actors. This only includes those
#: actors whose state actually lives on this worker, not actors to which this worker
#: has a reference.
actors: set[TaskState]
#: Underlying data of :meth:`WorkerState.has_what`
_has_what: dict[TaskState, None]
#: A set of tasks that have been submitted to this worker. Multiple tasks may be
# submitted to a worker in advance and the worker will run them eventually,
# depending on its execution resources (but see :doc:`work-stealing`).
#:
#: All the tasks here are in the "processing" state.
#: This attribute is kept in sync with :attr:`TaskState.processing_on`.
processing: set[TaskState]
#: Running tasks that invoked :func:`distributed.secede`
long_running: set[TaskState]
#: A dictionary of tasks that are currently being run on this worker.
#: Each task state is associated with the duration in seconds which the task has
#: been running.
executing: dict[TaskState, float]
#: The available resources on this worker, e.g. ``{"GPU": 2}``.
#: These are abstract quantities that constrain certain tasks from running at the
#: same time on this worker.
resources: dict[str, float]
#: The sum of each resource used by all tasks allocated to this worker.
#: The numbers in this dictionary can only be less or equal than those in this
#: worker's :attr:`~WorkerState.resources`.
used_resources: dict[str, float]
#: Arbitrary additional metadata to be added to :meth:`~WorkerState.identity`
extra: dict[str, Any]
# The unique server ID this WorkerState is referencing
server_id: str
# Reference to scheduler task_groups
scheduler_ref: weakref.ref[SchedulerState] | None
task_prefix_count: defaultdict[str, int]
_network_occ: float
_occupancy_cache: float | None
#: Keys that may need to be fetched to this worker, and the number of tasks that need them.
#: All tasks are currently in `memory` on a worker other than this one.
#: Much like `processing`, this does not exactly reflect worker state:
#: keys here may be queued to fetch, in flight, or already in memory
#: on the worker.
needs_what: dict[TaskState, int]
__slots__ = tuple(__annotations__)
def __init__(
self,
*,
address: str,
status: Status,
pid: int,
name: object,
nthreads: int = 0,
memory_limit: int,
local_directory: str,
nanny: str,
server_id: str,
services: dict[str, int] | None = None,
versions: dict[str, Any] | None = None,
extra: dict[str, Any] | None = None,
scheduler: SchedulerState | None = None,
):
self.server_id = server_id
self.address = address
self.pid = pid
self.name = name
self.nthreads = nthreads
self.memory_limit = memory_limit
self.local_directory = local_directory
self.services = services or {}
self.versions = versions or {}
self.nanny = nanny
self.status = status
self._hash = hash(self.server_id)
self.nbytes = 0
self._memory_unmanaged_old = 0
self._memory_unmanaged_history = deque()
self.metrics = {}
self.last_seen = 0
self.time_delay = 0
self.bandwidth = parse_bytes(dask.config.get("distributed.scheduler.bandwidth"))
self.actors = set()
self._has_what = {}
self.processing = set()
self.long_running = set()
self.executing = {}
self.resources = {}
self.used_resources = {}
self.extra = extra or {}
self.scheduler_ref = weakref.ref(scheduler) if scheduler else None
self.task_prefix_count = defaultdict(int)
self.needs_what = {}
self._network_occ = 0
self._occupancy_cache = None
def __hash__(self) -> int:
return self._hash
def __eq__(self, other: object) -> bool:
return isinstance(other, WorkerState) and other.server_id == self.server_id
@property
def has_what(self) -> Set[TaskState]:
"""An insertion-sorted set-like of tasks which currently reside on this worker.
All the tasks here are in the "memory" state.
This is the reverse mapping of :attr:`TaskState.who_has`.
This is a read-only public accessor. The data is implemented as a dict without
values, because rebalance() relies on dicts being insertion-sorted.
"""
return self._has_what.keys()
@property
def host(self) -> str:
return get_address_host(self.address)
@property
def memory(self) -> MemoryState:
"""Polished memory metrics for the worker.
**Design note on managed memory**
There are two measures available for managed memory:
- ``self.nbytes``
- ``self.metrics["managed_bytes"]``
At rest, the two numbers must be identical. However, ``self.nbytes`` is
immediately updated through the batched comms as soon as each task lands in
memory on the worker; ``self.metrics["managed_bytes"]`` instead is updated by
the heartbeat, which can lag several seconds behind.
Below we are mixing likely newer managed memory info from ``self.nbytes`` with
process and spilled memory from the heartbeat. This is deliberate, so that
managed memory total is updated more frequently.
Managed memory directly and immediately contributes to optimistic memory, which
is in turn used in Active Memory Manager heuristics (at the moment of writing;
more uses will likely be added in the future). So it's important to have it
up to date; much more than it is for process memory.
Having up-to-date managed memory info as soon as the scheduler learns about
task completion also substantially simplifies unit tests.
The flip side of this design is that it may cause some noise in the
unmanaged_recent measure. e.g.:
1. Delete 100MB of managed data
2. The updated managed memory reaches the scheduler faster than the
updated process memory
3. There's a blip where the scheduler thinks that there's a sudden 100MB
increase in unmanaged_recent, since process memory hasn't changed but managed
memory has decreased by 100MB
4. When the heartbeat arrives, process memory goes down and so does the
unmanaged_recent.
This is OK - one of the main reasons for the unmanaged_recent / unmanaged_old
split is exactly to concentrate all the noise in unmanaged_recent and exclude it
from optimistic memory, which is used for heuristics.
Something that is less OK, but also less frequent, is that the sudden deletion
of spilled keys will cause a negative blip of managed_in_memory:
1. Delete 100MB of spilled data
2. The updated managed memory *total* reaches the scheduler faster than the
updated spilled portion
3. This causes managed_in_memory to temporarily plummet and be replaced by
unmanaged_recent, while managed_spilled remains unaltered
4. When the heartbeat arrives, managed_in_memory goes back up, unmanaged_recent
goes back down, and managed_spilled goes down by 100MB as it should have to
begin with.
https://github.com/dask/distributed/issues/6002 will let us solve this.
"""
return MemoryState(
process=self.metrics["memory"],
managed_in_memory=max(
0, self.nbytes - self.metrics["spilled_bytes"]["memory"]
),
managed_spilled=self.metrics["spilled_bytes"]["disk"],
unmanaged_old=self._memory_unmanaged_old,
)
def clean(self) -> WorkerState:
"""Return a version of this object that is appropriate for serialization"""
ws = WorkerState(
address=self.address,
status=self.status,
pid=self.pid,
name=self.name,
nthreads=self.nthreads,
memory_limit=self.memory_limit,
local_directory=self.local_directory,
services=self.services,
nanny=self.nanny,
extra=self.extra,
server_id=self.server_id,
)
ws._occupancy_cache = self.occupancy
ws.executing = {
ts.key: duration for ts, duration in self.executing.items() # type: ignore
}
return ws
def __repr__(self) -> str:
name = f", name: {self.name}" if self.name != self.address else ""
return (
f"<WorkerState {self.address!r}{name}, "
f"status: {self.status.name}, "
f"memory: {len(self.has_what)}, "
f"processing: {len(self.processing)}>"
)
def _repr_html_(self) -> str:
return get_template("worker_state.html.j2").render(
address=self.address,
name=self.name,
status=self.status.name,
has_what=self.has_what,
processing=self.processing,
)
def identity(self) -> dict[str, Any]:
return {
"type": "Worker",
"id": self.name,
"host": self.host,
"resources": self.resources,
"local_directory": self.local_directory,
"name": self.name,
"nthreads": self.nthreads,
"memory_limit": self.memory_limit,
"last_seen": self.last_seen,
"services": self.services,
"metrics": self.metrics,
"status": self.status.name,
"nanny": self.nanny,
**self.extra,
}
def _to_dict_no_nest(self, *, exclude: Container[str] = ()) -> dict[str, Any]:
"""Dictionary representation for debugging purposes.
Not type stable and not intended for roundtrips.
See also
--------
Client.dump_cluster_state
distributed.utils.recursive_to_dict
TaskState._to_dict
"""
return recursive_to_dict(
self,
exclude=set(exclude) | {"versions"}, # type: ignore
members=True,
)
@property
def scheduler(self) -> SchedulerState:
assert self.scheduler_ref
s = self.scheduler_ref()
assert s
return s
def add_to_processing(self, ts: TaskState) -> None:
"""Assign a task to this worker for compute."""
if self.scheduler.validate:
assert ts not in self.processing
tp = ts.prefix
self.task_prefix_count[tp.name] += 1
self.scheduler._task_prefix_count_global[tp.name] += 1
self.processing.add(ts)
for dts in ts.dependencies:
if self not in dts.who_has:
self._inc_needs_replica(dts)
def add_to_long_running(self, ts: TaskState) -> None:
if self.scheduler.validate:
assert ts in self.processing
assert ts not in self.long_running
self._remove_from_task_prefix_count(ts)
# Cannot remove from processing since we're using this for things like
# idleness detection. Idle workers are typically targeted for
# downscaling but we should not downscale workers with long running
# tasks
self.long_running.add(ts)
def remove_from_processing(self, ts: TaskState) -> None:
"""Remove a task from a workers processing"""
if self.scheduler.validate:
assert ts in self.processing
if ts in self.long_running:
self.long_running.discard(ts)
else:
self._remove_from_task_prefix_count(ts)
self.processing.remove(ts)
for dts in ts.dependencies:
if dts in self.needs_what:
self._dec_needs_replica(dts)
def _remove_from_task_prefix_count(self, ts: TaskState) -> None:
count = self.task_prefix_count[ts.prefix.name] - 1
if count:
self.task_prefix_count[ts.prefix.name] = count
else:
del self.task_prefix_count[ts.prefix.name]
count = self.scheduler._task_prefix_count_global[ts.prefix.name] - 1
if count:
self.scheduler._task_prefix_count_global[ts.prefix.name] = count
else:
del self.scheduler._task_prefix_count_global[ts.prefix.name]
def remove_replica(self, ts: TaskState) -> None:
"""The worker no longer has a task in memory"""
if self.scheduler.validate:
assert self in ts.who_has
assert ts in self.has_what
assert ts not in self.needs_what
self.nbytes -= ts.get_nbytes()
del self._has_what[ts]
ts.who_has.remove(self)
def _inc_needs_replica(self, ts: TaskState) -> None:
"""Assign a task fetch to this worker and update network occupancies"""
if self.scheduler.validate:
assert self not in ts.who_has
assert ts not in self.has_what
if ts not in self.needs_what:
self.needs_what[ts] = 1
nbytes = ts.get_nbytes()
self._network_occ += nbytes
self.scheduler._network_occ_global += nbytes
else:
self.needs_what[ts] += 1
def _dec_needs_replica(self, ts: TaskState) -> None:
if self.scheduler.validate:
assert ts in self.needs_what
self.needs_what[ts] -= 1
if self.needs_what[ts] == 0:
del self.needs_what[ts]
nbytes = ts.get_nbytes()
self._network_occ -= nbytes
self.scheduler._network_occ_global -= nbytes
def add_replica(self, ts: TaskState) -> None:
"""The worker acquired a replica of task"""
if self.scheduler.validate:
assert self not in ts.who_has
assert ts not in self.has_what
nbytes = ts.get_nbytes()
if ts in self.needs_what:
del self.needs_what[ts]
self._network_occ -= nbytes
self.scheduler._network_occ_global -= nbytes
ts.who_has.add(self)
self.nbytes += nbytes
self._has_what[ts] = None
@property
def occupancy(self) -> float:
return self._occupancy_cache or self.scheduler._calc_occupancy(
self.task_prefix_count, self._network_occ
)
@dataclasses.dataclass
class ErredTask:
"""Lightweight representation of an erred task without any dependency information
or runspec.
See also
--------
TaskState
"""
key: Hashable
timestamp: float
erred_on: set[str]
exception_text: str
traceback_text: str
class Computation:
"""Collection tracking a single compute or persist call
See also
--------
TaskPrefix
TaskGroup
TaskState
"""
start: float
groups: set[TaskGroup]
code: SortedSet
id: uuid.UUID
__slots__ = tuple(__annotations__)
def __init__(self):
self.start = time()
self.groups = set()
self.code = SortedSet()
self.id = uuid.uuid4()
@property
def stop(self) -> float:
if self.groups:
return max(tg.stop for tg in self.groups)
else:
return -1
@property
def states(self) -> dict[TaskStateState, int]:
return merge_with(sum, (tg.states for tg in self.groups))
def __repr__(self) -> str:
return (
f"<Computation {self.id}: "
+ "Tasks: "
+ ", ".join(
"%s: %d" % (k, v) for (k, v) in sorted(self.states.items()) if v
)
+ ">"
)
def _repr_html_(self) -> str:
return get_template("computation.html.j2").render(
id=self.id,
start=self.start,
stop=self.stop,
groups=self.groups,
states=self.states,
code=self.code,
)
class TaskPrefix:
"""Collection tracking all tasks within a group
Keys often have a structure like ``("x-123", 0)``
A group takes the first section, like ``"x"``
See Also
--------
TaskGroup
"""
#: The name of a group of tasks.
#: For a task like ``("x-123", 0)`` this is the text ``"x"``
name: str
#: An exponentially weighted moving average duration of all tasks with this prefix
duration_average: float
#: Numbers of times a task was marked as suspicious with this prefix
suspicious: int
#: Store timings for each prefix-action
all_durations: defaultdict[str, float]
#: This measures the maximum recorded live execution time and can be used to
#: detect outliers
max_exec_time: float
#: Task groups associated to this prefix
groups: list[TaskGroup]
#: Accumulate count of number of tasks in each state
state_counts: defaultdict[str, int]
__slots__ = tuple(__annotations__)
def __init__(self, name: str):
self.name = name
self.groups = []
self.all_durations = defaultdict(float)
self.state_counts = defaultdict(int)
task_durations = dask.config.get("distributed.scheduler.default-task-durations")
if self.name in task_durations:
self.duration_average = parse_timedelta(task_durations[self.name])
else:
self.duration_average = -1
self.max_exec_time = -1
self.suspicious = 0
def add_exec_time(self, duration: float) -> None:
self.max_exec_time = max(duration, self.max_exec_time)
if duration > 2 * self.duration_average:
self.duration_average = -1
def add_duration(self, action: str, start: float, stop: float) -> None:
duration = stop - start
self.all_durations[action] += duration
if action == "compute":
old = self.duration_average
if old < 0:
self.duration_average = duration
else:
self.duration_average = 0.5 * duration + 0.5 * old
@property
def states(self) -> dict[str, int]:
"""The number of tasks in each state,
like ``{"memory": 10, "processing": 3, "released": 4, ...}``
"""
return merge_with(sum, [tg.states for tg in self.groups])
@property
def active(self) -> list[TaskGroup]:
return [
tg
for tg in self.groups
if any(k != "forgotten" and v != 0 for k, v in tg.states.items())
]
@property
def active_states(self) -> dict[str, int]:
return merge_with(sum, [tg.states for tg in self.active])
def __repr__(self) -> str:
return (
"<"
+ self.name
+ ": "
+ ", ".join(
"%s: %d" % (k, v) for (k, v) in sorted(self.states.items()) if v
)
+ ">"
)
@property
def nbytes_total(self) -> int:
return sum(tg.nbytes_total for tg in self.groups)
def __len__(self) -> int:
return sum(map(len, self.groups))
@property
def duration(self) -> float:
return sum(tg.duration for tg in self.groups)
@property
def types(self) -> set[str]:
return {typ for tg in self.groups for typ in tg.types}
class TaskGroup:
"""Collection tracking all tasks within a group
Keys often have a structure like ``("x-123", 0)``
A group takes the first section, like ``"x-123"``
See also
--------
TaskPrefix
"""
#: The name of a group of tasks.
#: For a task like ``("x-123", 0)`` this is the text ``"x-123"``
name: str
#: The number of tasks in each state,
#: like ``{"memory": 10, "processing": 3, "released": 4, ...}``
states: dict[TaskStateState, int]
#: The other TaskGroups on which this one depends
dependencies: set[TaskGroup]
#: The total number of bytes that this task group has produced
nbytes_total: int
#: The total amount of time spent on all tasks in this TaskGroup
duration: float
#: The result types of this TaskGroup
types: set[str]
#: The worker most recently assigned a task from this group, or None when the group
#: is not identified to be root-like by `SchedulerState.decide_worker`.
last_worker: WorkerState | None
#: If `last_worker` is not None, the number of times that worker should be assigned
#: subsequent tasks until a new worker is chosen.
last_worker_tasks_left: int
prefix: TaskPrefix | None
start: float
stop: float
all_durations: defaultdict[str, float]
__slots__ = tuple(__annotations__)
def __init__(self, name: str):
self.name = name
self.prefix = None
self.states = dict.fromkeys(ALL_TASK_STATES, 0)
self.dependencies = set()
self.nbytes_total = 0
self.duration = 0
self.types = set()
self.start = 0.0
self.stop = 0.0
self.all_durations = defaultdict(float)
self.last_worker = None
self.last_worker_tasks_left = 0
def add_duration(self, action: str, start: float, stop: float) -> None:
duration = stop - start
self.all_durations[action] += duration
if action == "compute":
if self.stop < stop:
self.stop = stop
self.start = self.start or start
self.duration += duration
assert self.prefix is not None
self.prefix.add_duration(action, start, stop)
def add(self, other: TaskState) -> None:
self.states[other.state] += 1
other.group = self
def __repr__(self) -> str:
return (
"<"
+ (self.name or "no-group")
+ ": "
+ ", ".join(
"%s: %d" % (k, v) for (k, v) in sorted(self.states.items()) if v
)
+ ">"
)
def __len__(self) -> int:
return sum(self.states.values())
def _to_dict_no_nest(self, *, exclude: Container[str] = ()) -> dict[str, Any]:
"""Dictionary representation for debugging purposes.
Not type stable and not intended for roundtrips.
See also
--------
Client.dump_cluster_state
distributed.utils.recursive_to_dict
TaskState._to_dict
"""
return recursive_to_dict(self, exclude=exclude, members=True)
class TaskState:
"""A simple object holding information about a task.
Not to be confused with :class:`distributed.worker_state_machine.TaskState`, which
holds similar information on the Worker side.
"""
#: The key is the unique identifier of a task, generally formed from the name of the
#: function, followed by a hash of the function and arguments, like
#: ``'inc-ab31c010444977004d656610d2d421ec'``.
key: str
#: The broad class of tasks to which this task belongs like "inc" or "read_csv"
prefix: TaskPrefix
#: A specification of how to run the task. The type and meaning of this value is
#: opaque to the scheduler, as it is only interpreted by the worker to which the
#: task is sent for executing.
#:
#: As a special case, this attribute may also be ``None``, in which case the task is
#: "pure data" (such as, for example, a piece of data loaded in the scheduler using
#: :meth:`Client.scatter`). A "pure data" task cannot be computed again if its
#: value is lost.
run_spec: object
#: The priority provides each task with a relative ranking which is used to break
#: ties when many tasks are being considered for execution.
#:
#: This ranking is generally a 2-item tuple. The first (and dominant) item
#: corresponds to when it was submitted. Generally, earlier tasks take precedence.
#: The second item is determined by the client, and is a way to prioritize tasks
#: within a large graph that may be important, such as if they are on the critical
#: path, or good to run in order to release many dependencies. This is explained
#: further in :doc:`Scheduling Policy <scheduling-policies>`.
priority: tuple[int, ...]
# Attribute underlying the state property
_state: TaskStateState
#: The set of tasks this task depends on for proper execution. Only tasks still
#: alive are listed in this set. If, for whatever reason, this task also depends on
#: a forgotten task, the :attr:`has_lost_dependencies` flag is set.
#:
#: A task can only be executed once all its dependencies have already been
#: successfully executed and have their result stored on at least one worker. This
#: is tracked by progressively draining the :attr:`waiting_on` set.
dependencies: set[TaskState]
#: The set of tasks which depend on this task. Only tasks still alive are listed in
#: this set. This is the reverse mapping of :attr:`dependencies`.
dependents: set[TaskState]
#: Whether any of the dependencies of this task has been forgotten. For memory
#: consumption reasons, forgotten tasks are not kept in memory even though they may
#: have dependent tasks. When a task is forgotten, therefore, each of its
#: dependents has their :attr:`has_lost_dependencies` attribute set to ``True``.
#:
#: If :attr:`has_lost_dependencies` is true, this task cannot go into the
#: "processing" state anymore.
has_lost_dependencies: bool
#: The set of tasks this task is waiting on *before* it can be executed. This is
#: always a subset of :attr:`dependencies`. Each time one of the dependencies has
#: finished processing, it is removed from the :attr:`waiting_on` set.
#:
#: Once :attr:`waiting_on` becomes empty, this task can move from the "waiting"
#: state to the "processing" state (unless one of the dependencies errored out, in
#: which case this task is instead marked "erred").
waiting_on: set[TaskState]
#: The set of tasks which need this task to remain alive. This is always a subset
#: of :attr:`dependents`. Each time one of the dependents has finished processing,
#: it is removed from the :attr:`waiters` set.
#:
#: Once both :attr:`waiters` and :attr:`who_wants` become empty, this task can be
#: released (if it has a non-empty :attr:`run_spec`) or forgotten (otherwise) by the
#: scheduler, and by any workers in :attr:`who_has`.
#:
#: .. note::
#: Counter-intuitively, :attr:`waiting_on` and :attr:`waiters` are not reverse
#: mappings of each other.
waiters: set[TaskState]
#: The set of clients who want the result of this task to remain alive.
#: This is the reverse mapping of :attr:`ClientState.wants_what`.
#:
#: When a client submits a graph to the scheduler it also specifies which output
#: tasks it desires, such that their results are not released from memory.
#:
#: Once a task has finished executing (i.e. moves into the "memory" or "erred"
#: state), the clients in :attr:`who_wants` are notified.
#:
#: Once both :attr:`waiters` and :attr:`who_wants` become empty, this task can be
#: released (if it has a non-empty :attr:`run_spec`) or forgotten (otherwise) by the
#: scheduler, and by any workers in :attr:`who_has`.
who_wants: set[ClientState]
#: The set of workers who have this task's result in memory. It is non-empty iff the
#: task is in the "memory" state. There can be more than one worker in this set if,
#: for example, :meth:`Client.scatter` or :meth:`Client.replicate` was used.
#:
#: This is the reverse mapping of :attr:`WorkerState.has_what`.
who_has: set[WorkerState]
#: If this task is in the "processing" state, which worker is currently processing
#: it. This attribute is kept in sync with :attr:`WorkerState.processing`.
processing_on: WorkerState | None
#: The number of times this task can automatically be retried in case of failure.
#: If a task fails executing (the worker returns with an error), its :attr:`retries`
#: attribute is checked. If it is equal to 0, the task is marked "erred". If it is
#: greater than 0, the :attr:`retries` attribute is decremented and execution is
#: attempted again.
retries: int
#: The number of bytes, as determined by ``sizeof``, of the result of a finished
#: task. This number is used for diagnostics and to help prioritize work.
#: Set to -1 for unfinished tasks.
nbytes: int
#: The type of the object as a string. Only present for tasks that have been
#: computed.
type: str
#: If this task failed executing, the exception object is stored here.
exception: Serialized | None
#: If this task failed executing, the traceback object is stored here.
traceback: Serialized | None
#: string representation of exception
exception_text: str
#: string representation of traceback
traceback_text: str
#: If this task or one of its dependencies failed executing, the failed task is
#: stored here (possibly itself).
exception_blame: TaskState | None
#: Worker addresses on which errors appeared, causing this task to be in an error
#: state.
erred_on: set[str]
#: The number of times this task has been involved in a worker death.
#:
#: Some tasks may cause workers to die (such as calling ``os._exit(0)``). When a
#: worker dies, all of the tasks on that worker are reassigned to others. This
#: combination of behaviors can cause a bad task to catastrophically destroy all
#: workers on the cluster, one after another. Whenever a worker dies, we mark each
#: task currently processing on that worker (as recorded by
#: :attr:`WorkerState.processing`) as suspicious. If a task is involved in three
#: deaths (or some other fixed constant) then we mark the task as ``erred``.
suspicious: int
#: A set of hostnames where this task can be run (or ``None`` if empty). Usually
#: this is empty unless the task has been specifically restricted to only run on
#: certain hosts. A hostname may correspond to one or several connected workers.
host_restrictions: set[str]
#: A set of complete worker addresses where this can be run (or ``None`` if empty).
#: Usually this is empty unless the task has been specifically restricted to only
#: run on certain workers.
#: Note this is tracking worker addresses, not worker states, since the specific
#: workers may not be connected at this time.
worker_restrictions: set[str]
#: Resources required by this task, such as ``{'gpu': 1}`` or ``{'memory': 1e9}``
#: These are user-defined names and are matched against the : contents of each
#: :attr:`WorkerState.resources` dictionary.
resource_restrictions: dict[str, float]
#: False
#: Each of :attr:`host_restrictions`, :attr:`worker_restrictions` and
#: :attr:`resource_restrictions` is a hard constraint: if no worker is available
#: satisfying those restrictions, the task cannot go into the "processing" state
#: and will instead go into the "no-worker" state.
#: True
#: The above restrictions are mere preferences: if no worker is available
#: satisfying those restrictions, the task can still go into the
#: "processing" state and be sent for execution to another connected worker.
loose_restrictions: bool
#: Whether this task is an Actor
actor: bool
#: The group of tasks to which this one belongs
group: TaskGroup
#: Same as of group.name
group_key: str
#: Metadata related to task
metadata: dict[str, Any]
#: Task annotations
annotations: dict[str, Any]
#: Cached hash of :attr:`~TaskState.client_key`
_hash: int
# Support for weakrefs to a class with __slots__
__weakref__: Any = None
__slots__ = tuple(__annotations__)
# Instances not part of slots since class variable
_instances: ClassVar[weakref.WeakSet[TaskState]] = weakref.WeakSet()
def __init__(self, key: str, run_spec: object, state: TaskStateState):
self.key = key
self._hash = hash(key)
self.run_spec = run_spec
self._state = state
self.exception = None
self.exception_blame = None
self.traceback = None
self.exception_text = ""
self.traceback_text = ""
self.suspicious = 0
self.retries = 0
self.nbytes = -1
self.priority = None # type: ignore
self.who_wants = set()
self.dependencies = set()
self.dependents = set()
self.waiting_on = set()
self.waiters = set()
self.who_has = set()
self.processing_on = None
self.has_lost_dependencies = False
self.host_restrictions = None # type: ignore
self.worker_restrictions = None # type: ignore
self.resource_restrictions = {}
self.loose_restrictions = False
self.actor = False
self.prefix = None # type: ignore
self.type = None # type: ignore
self.group_key = key_split_group(key)
self.group = None # type: ignore
self.metadata = {}
self.annotations = {}
self.erred_on = set()
TaskState._instances.add(self)
def __hash__(self) -> int:
return self._hash
def __eq__(self, other: object) -> bool:
return isinstance(other, TaskState) and self.key == other.key
@property
def state(self) -> TaskStateState:
"""This task's current state. Valid states are ``released``, ``waiting``,
``no-worker``, ``processing``, ``memory``, ``erred`` and ``forgotten``. If it
is ``forgotten``, the task isn't stored in the ``tasks`` dictionary anymore and
will probably disappear soon from memory.
"""
return self._state
@state.setter
def state(self, value: TaskStateState) -> None:
self.group.states[self._state] -= 1
self.group.states[value] += 1
self._state = value
self.prefix.state_counts[value] += 1
def add_dependency(self, other: TaskState) -> None:
"""Add another task as a dependency of this task"""
self.dependencies.add(other)
self.group.dependencies.add(other.group)
other.dependents.add(self)
def get_nbytes(self) -> int:
return self.nbytes if self.nbytes >= 0 else DEFAULT_DATA_SIZE
def set_nbytes(self, nbytes: int) -> None:
diff = nbytes
old_nbytes = self.nbytes
if old_nbytes >= 0:
diff -= old_nbytes
self.group.nbytes_total += diff
for ws in self.who_has:
ws.nbytes += diff
self.nbytes = nbytes
def __repr__(self) -> str:
return f"<TaskState {self.key!r} {self._state}>"
def _repr_html_(self) -> str:
return get_template("task_state.html.j2").render(
state=self.state,
nbytes=self.nbytes,
key=self.key,
)
def validate(self) -> None:
try:
for cs in self.who_wants:
assert isinstance(cs, ClientState), (repr(cs), self.who_wants)
for ws in self.who_has:
assert isinstance(ws, WorkerState), (repr(ws), self.who_has)
for ts in self.dependencies:
assert isinstance(ts, TaskState), (repr(ts), self.dependencies)
for ts in self.dependents:
assert isinstance(ts, TaskState), (repr(ts), self.dependents)
validate_task_state(self)
except Exception as e:
logger.exception(e)
if LOG_PDB:
import pdb
pdb.set_trace()
def get_nbytes_deps(self) -> int:
return sum(ts.get_nbytes() for ts in self.dependencies)
def _to_dict_no_nest(self, *, exclude: Container[str] = ()) -> dict[str, Any]:
"""Dictionary representation for debugging purposes.
Not type stable and not intended for roundtrips.
See also
--------
Client.dump_cluster_state
distributed.utils.recursive_to_dict
Notes
-----
This class uses ``_to_dict_no_nest`` instead of ``_to_dict``.
When a task references another task, or when a WorkerState.tasks contains tasks,
this method is not executed for the inner task, even if the inner task was never
seen before; you get a repr instead. All tasks should neatly appear under
Scheduler.tasks. This also prevents a RecursionError during particularly heavy
loads, which have been observed to happen whenever there's an acyclic dependency
chain of ~200+ tasks.
"""
return recursive_to_dict(self, exclude=exclude, members=True)
class Transition(NamedTuple):
"""An entry in :attr:`SchedulerState.transition_log`"""
key: str
start: TaskStateState
finish: TaskStateState
recommendations: Recs
stimulus_id: str
timestamp: float
class SchedulerState:
"""Underlying task state of dynamic scheduler
Tracks the current state of workers, data, and computations.
Handles transitions between different task states. Notifies the
Scheduler of changes by messaging passing through Queues, which the
Scheduler listens to responds accordingly.
All events are handled quickly, in linear time with respect to their
input (which is often of constant size) and generally within a
millisecond.
Users typically do not interact with ``Transitions`` directly. Instead
users interact with the ``Client``, which in turn engages the
``Scheduler`` affecting different transitions here under-the-hood. In
the background ``Worker``s also engage with the ``Scheduler``
affecting these state transitions as well.
"""
bandwidth: int
#: Clients currently connected to the scheduler
clients: dict[str, ClientState]
extensions: dict[str, Any] # TODO write a scheduler extension Protocol
plugins: dict[str, SchedulerPlugin]
host_info: dict[str, dict[str, Any]]
#: If True, enable expensive internal consistency check.
#: Typically disabled in production.
validate: bool
#######################
# Workers-related state
#######################
#: Workers currently connected to the scheduler
#: (actually a SortedDict, but the sortedcontainers package isn't annotated)
workers: dict[str, WorkerState]
#: Worker {name: address}
aliases: dict[Hashable, str]
#: Workers that are currently in running state
running: set[WorkerState]
#: Workers that are currently in running state and not fully utilized
#: Definition based on occupancy
#: (actually a SortedDict, but the sortedcontainers package isn't annotated)
idle: dict[str, WorkerState]
#: Similar to `idle`
#: Definition based on assigned tasks
idle_task_count: set[WorkerState]
#: Workers that are fully utilized. May include non-running workers.
saturated: set[WorkerState]
total_nthreads: int
#: Cluster-wide resources. {resource name: {worker address: amount}}
resources: dict[str, dict[str, float]]
#####################
# Tasks-related state
#####################
#: Total number of tasks ever processed
n_tasks: int
#: All tasks currently known to the scheduler
tasks: dict[str, TaskState]
#: Tasks in the "queued" state, ordered by priority
queued: HeapSet[TaskState]
#: Tasks in the "no-worker" state
unrunnable: set[TaskState]
#: Subset of tasks that exist in memory on more than one worker
replicated_tasks: set[TaskState]
unknown_durations: dict[str, set[TaskState]]
task_groups: dict[str, TaskGroup]
task_prefixes: dict[str, TaskPrefix]
task_metadata: dict[str, Any]
#########
# History
#########
#: History of computations.
#: The length can be tweaked through
#: distributed.diagnostics.computations.max-history
computations: deque[Computation]
#: History of erred tasks.
#: The length can be tweaked through
#: distributed.diagnostics.erred-tasks.max-history
erred_tasks: deque[ErredTask]
#: History of task state transitions.
#: The length can be tweaked through
#: distributed.scheduler.transition-log-length
transition_log: deque[Transition]
#: Total number of transitions since the cluster was started
transition_counter: int
#: Total number of transitions as of the previous call to check_idle()
_idle_transition_counter: int
#: Raise an error if the :attr:`transition_counter` ever reaches this value.
#: This is meant for debugging only, to catch infinite recursion loops.
#: In production, it should always be set to False.
transition_counter_max: int | Literal[False]
_task_prefix_count_global: defaultdict[str, int]
_network_occ_global: float
######################
# Cached configuration
######################
#: distributed.scheduler.unknown-task-duration
UNKNOWN_TASK_DURATION: float
#: distributed.worker.memory.recent-to-old-time
MEMORY_RECENT_TO_OLD_TIME: float
#: distributed.worker.memory.rebalance.measure
MEMORY_REBALANCE_MEASURE: str
#: distributed.worker.memory.rebalance.sender-min
MEMORY_REBALANCE_SENDER_MIN: float
#: distributed.worker.memory.rebalance.recipient-max
MEMORY_REBALANCE_RECIPIENT_MAX: float
#: distributed.worker.memory.rebalance.sender-recipient-gap / 2
MEMORY_REBALANCE_HALF_GAP: float
#: distributed.scheduler.worker-saturation
WORKER_SATURATION: float
__slots__ = tuple(__annotations__)
def __init__(
self,
aliases: dict[Hashable, str],
clients: dict[str, ClientState],
workers: SortedDict[str, WorkerState],
host_info: dict[str, dict[str, Any]],
resources: dict[str, dict[str, float]],
tasks: dict[str, TaskState],
unrunnable: set[TaskState],
queued: HeapSet[TaskState],
validate: bool,
plugins: Iterable[SchedulerPlugin] = (),
transition_counter_max: int | Literal[False] = False,
**kwargs: Any, # Passed verbatim to Server.__init__()
):
logger.info("State start")
self.aliases = aliases
self.bandwidth = parse_bytes(dask.config.get("distributed.scheduler.bandwidth"))
self.clients = clients
self.clients["fire-and-forget"] = ClientState("fire-and-forget")
self.extensions = {}
self.host_info = host_info
self.idle = SortedDict()
self.idle_task_count = set()
self.n_tasks = 0
self.resources = resources
self.saturated = set()
self.tasks = tasks
self.replicated_tasks = {
ts for ts in self.tasks.values() if len(ts.who_has) > 1
}
self.computations = deque(
maxlen=dask.config.get("distributed.diagnostics.computations.max-history")
)
self.erred_tasks = deque(
maxlen=dask.config.get("distributed.diagnostics.erred-tasks.max-history")
)
self.task_groups = {}
self.task_prefixes = {}
self.task_metadata = {}
self.total_nthreads = 0
self.unknown_durations = {}
self.queued = queued
self.unrunnable = unrunnable
self.validate = validate
self.workers = workers
self._task_prefix_count_global = defaultdict(int)
self._network_occ_global = 0.0
self.running = {
ws for ws in self.workers.values() if ws.status == Status.running
}
self.plugins = {} if not plugins else {_get_plugin_name(p): p for p in plugins}
self.transition_log = deque(
maxlen=dask.config.get("distributed.scheduler.transition-log-length")
)
self.transition_counter = 0
self._idle_transition_counter = 0
self.transition_counter_max = transition_counter_max
# Variables from dask.config, cached by __init__ for performance
self.UNKNOWN_TASK_DURATION = parse_timedelta(
dask.config.get("distributed.scheduler.unknown-task-duration")
)
self.MEMORY_RECENT_TO_OLD_TIME = parse_timedelta(
dask.config.get("distributed.worker.memory.recent-to-old-time")
)
self.MEMORY_REBALANCE_MEASURE = dask.config.get(
"distributed.worker.memory.rebalance.measure"
)
self.MEMORY_REBALANCE_SENDER_MIN = dask.config.get(
"distributed.worker.memory.rebalance.sender-min"
)
self.MEMORY_REBALANCE_RECIPIENT_MAX = dask.config.get(
"distributed.worker.memory.rebalance.recipient-max"
)
self.MEMORY_REBALANCE_HALF_GAP = (
dask.config.get("distributed.worker.memory.rebalance.sender-recipient-gap")
/ 2.0
)
self.WORKER_SATURATION = dask.config.get(
"distributed.scheduler.worker-saturation"
)
if self.WORKER_SATURATION == "inf":
# Special case necessary because there's no way to parse a float infinity
# from a DASK_* environment variable
self.WORKER_SATURATION = math.inf
if (
not isinstance(self.WORKER_SATURATION, (int, float))
or self.WORKER_SATURATION <= 0
):
raise ValueError( # pragma: nocover
"`distributed.scheduler.worker-saturation` must be a float > 0; got "
+ repr(self.WORKER_SATURATION)
)
@property
def memory(self) -> MemoryState:
return MemoryState.sum(*(w.memory for w in self.workers.values()))
@property
def __pdict__(self) -> dict[str, Any]:
return {
"bandwidth": self.bandwidth,
"resources": self.resources,
"saturated": self.saturated,
"unrunnable": self.unrunnable,
"queued": self.queued,
"n_tasks": self.n_tasks,
"unknown_durations": self.unknown_durations,
"validate": self.validate,
"tasks": self.tasks,
"task_groups": self.task_groups,
"task_prefixes": self.task_prefixes,
"total_nthreads": self.total_nthreads,
"total_occupancy": self.total_occupancy,
"erred_tasks": self.erred_tasks,
"extensions": self.extensions,
"clients": self.clients,
"workers": self.workers,
"idle": self.idle,
"host_info": self.host_info,
}
def new_task(
self,
key: str,
spec: object,
state: TaskStateState,
computation: Computation | None = None,
) -> TaskState:
"""Create a new task, and associated states"""
ts = TaskState(key, spec, state)
prefix_key = key_split(key)
tp = self.task_prefixes.get(prefix_key)
if tp is None:
self.task_prefixes[prefix_key] = tp = TaskPrefix(prefix_key)
ts.prefix = tp
group_key = ts.group_key
tg = self.task_groups.get(group_key)
if tg is None:
self.task_groups[group_key] = tg = TaskGroup(group_key)
if computation:
computation.groups.add(tg)
tg.prefix = tp
tp.groups.append(tg)
tg.add(ts)
self.tasks[key] = ts
return ts
def _clear_task_state(self) -> None:
logger.debug("Clear task state")
for collection in (
self.unrunnable,
self.erred_tasks,
self.computations,
self.task_prefixes,
self.task_groups,
self.task_metadata,
self.unknown_durations,
self.replicated_tasks,
):
collection.clear() # type: ignore
@property
def total_occupancy(self) -> float:
return self._calc_occupancy(
self._task_prefix_count_global,
self._network_occ_global,
)
def _calc_occupancy(
self,
task_prefix_count: dict[str, int],
network_occ: float,
) -> float:
res = 0.0
for prefix_name, count in task_prefix_count.items():
# TODO: Deal with unknown tasks better
prefix = self.task_prefixes[prefix_name]
assert prefix is not None
duration = prefix.duration_average
if duration < 0:
if prefix.max_exec_time > 0:
duration = 2 * prefix.max_exec_time
else:
duration = self.UNKNOWN_TASK_DURATION
res += duration * count
occ = res + network_occ / self.bandwidth
assert occ >= 0, occ
return occ
#####################
# State Transitions #
#####################
def _transition(
self, key: str, finish: TaskStateState, stimulus_id: str, **kwargs: Any
) -> RecsMsgs:
"""Transition a key from its current state to the finish state
Examples
--------
>>> self._transition('x', 'waiting')
{'x': 'processing'}, {}, {}
Returns
-------
Tuple of:
- Dictionary of recommendations for future transitions {key: new state}
- Messages to clients {client address: [msg, msg, ...]}
- Messages to workers {worker address: [msg, msg, ...]}
See Also
--------
Scheduler.transitions : transitive version of this function
"""
try:
ts = self.tasks.get(key)
if ts is None:
return {}, {}, {}
start = ts._state
if start == finish:
return {}, {}, {}
# Notes:
# - in case of transition through released, this counter is incremented by 2
# - this increase happens before the actual transitions, so that it can
# catch potential infinite recursions
self.transition_counter += 1
if self.transition_counter_max:
assert self.transition_counter < self.transition_counter_max
recommendations: dict = {}
worker_msgs: dict = {}
client_msgs: dict = {}
if self.plugins:
dependents = set(ts.dependents)
dependencies = set(ts.dependencies)
func = self._TRANSITIONS_TABLE.get((start, finish))
if func is not None:
recommendations, client_msgs, worker_msgs = func(
self, key, stimulus_id, **kwargs
)
elif "released" not in (start, finish):
assert not kwargs, (kwargs, start, finish)
a_recs, a_cmsgs, a_wmsgs = self._transition(
key, "released", stimulus_id
)
v = a_recs.get(key, finish)
func = self._TRANSITIONS_TABLE["released", v]
b_recs, b_cmsgs, b_wmsgs = func(self, key, stimulus_id)
recommendations.update(a_recs)
for c, new_msgs in a_cmsgs.items():
client_msgs.setdefault(c, []).extend(new_msgs)
for w, new_msgs in a_wmsgs.items():
worker_msgs.setdefault(w, []).extend(new_msgs)
recommendations.update(b_recs)
for c, new_msgs in b_cmsgs.items():
client_msgs.setdefault(c, []).extend(new_msgs)
for w, new_msgs in b_wmsgs.items():
worker_msgs.setdefault(w, []).extend(new_msgs)
start = "released"
else:
raise RuntimeError(
f"Impossible transition from {start} to {finish} for {key!r}: "
f"{stimulus_id=}, {kwargs=}, story={self.story(ts)}"
)
if not stimulus_id:
stimulus_id = STIMULUS_ID_UNSET
actual_finish = ts._state
self.transition_log.append(
Transition(
key, start, actual_finish, recommendations, stimulus_id, time()
)
)
if self.validate:
if stimulus_id == STIMULUS_ID_UNSET:
raise RuntimeError(
"stimulus_id not set during Scheduler transition"
)
logger.debug(
"Transitioned %r %s->%s (actual: %s). Consequence: %s",
key,
start,
finish,
actual_finish,
dict(recommendations),
)
if self.plugins:
# Temporarily put back forgotten key for plugin to retrieve it
if ts._state == "forgotten":
ts.dependents = dependents
ts.dependencies = dependencies
self.tasks[ts.key] = ts
for plugin in list(self.plugins.values()):
try:
plugin.transition(key, start, actual_finish, **kwargs)
except Exception:
logger.info("Plugin failed with exception", exc_info=True)
if ts.state == "forgotten":
del self.tasks[ts.key]
tg = ts.group
if ts.state == "forgotten" and tg.name in self.task_groups:
# Remove TaskGroup if all tasks are in the forgotten state
if all(v == 0 or k == "forgotten" for k, v in tg.states.items()):
ts.prefix.groups.remove(tg)
del self.task_groups[tg.name]
return recommendations, client_msgs, worker_msgs
except Exception:
logger.exception("Error transitioning %r from %r to %r", key, start, finish)
if LOG_PDB:
import pdb
pdb.set_trace()
raise
def _transitions(
self,
recommendations: Recs,
client_msgs: Msgs,
worker_msgs: Msgs,
stimulus_id: str,
) -> None:
"""Process transitions until none are left
This includes feedback from previous transitions and continues until we
reach a steady state
"""
keys: set[str] = set()
recommendations = recommendations.copy()
while recommendations:
key, finish = recommendations.popitem()
keys.add(key)
new_recs, new_cmsgs, new_wmsgs = self._transition(key, finish, stimulus_id)
recommendations.update(new_recs)
for c, new_msgs in new_cmsgs.items():
client_msgs.setdefault(c, []).extend(new_msgs)
for w, new_msgs in new_wmsgs.items():
worker_msgs.setdefault(w, []).extend(new_msgs)
if self.validate:
# FIXME downcast antipattern
scheduler = cast(Scheduler, self)
for key in keys:
scheduler.validate_key(key)
def transition_released_waiting(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
if self.validate:
assert ts.run_spec
assert not ts.waiting_on
assert not ts.who_has
assert not ts.processing_on
for dts in ts.dependencies:
assert dts.state not in {"forgotten", "erred"}
if ts.has_lost_dependencies:
return {key: "forgotten"}, {}, {}
ts.state = "waiting"
recommendations: Recs = {}
for dts in ts.dependencies:
if not dts.who_has:
ts.waiting_on.add(dts)
if dts.state == "released":
recommendations[dts.key] = "waiting"
else:
dts.waiters.add(ts)
ts.waiters = {dts for dts in ts.dependents if dts.state == "waiting"}
if not ts.waiting_on:
# NOTE: waiting->processing will send tasks to queued or no-worker as
# necessary
recommendations[key] = "processing"
return recommendations, {}, {}
def transition_no_worker_processing(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
worker_msgs: Msgs = {}
if self.validate:
assert not ts.actor, f"Actors can't be in `no-worker`: {ts}"
assert ts in self.unrunnable
if ws := self.decide_worker_non_rootish(ts):
self.unrunnable.discard(ts)
worker_msgs = self._add_to_processing(ts, ws)
# If no worker, task just stays in `no-worker`
return {}, {}, worker_msgs
def decide_worker_rootish_queuing_disabled(
self, ts: TaskState
) -> WorkerState | None:
"""Pick a worker for a runnable root-ish task, without queuing.
This attempts to schedule sibling tasks on the same worker, reducing future data
transfer. It does not consider the location of dependencies, since they'll end
up on every worker anyway.
It assumes it's being called on a batch of tasks in priority order, and
maintains state in `SchedulerState.last_root_worker` and
`SchedulerState.last_root_worker_tasks_left` to achieve this.
This will send every runnable task to a worker, often causing root task
overproduction.
Returns
-------
ws: WorkerState | None
The worker to assign the task to. If there are no workers in the cluster,
returns None, in which case the task should be transitioned to
``no-worker``.
"""
if self.validate:
# See root-ish-ness note below in `decide_worker_rootish_queuing_enabled`
assert math.isinf(self.WORKER_SATURATION)
pool = self.idle.values() if self.idle else self.running
if not pool:
return None
tg = ts.group
lws = tg.last_worker
if (
lws
and tg.last_worker_tasks_left
and lws.status == Status.running
and self.workers.get(lws.address) is lws
):
ws = lws
else:
# Last-used worker is full, unknown, retiring, or paused;
# pick a new worker for the next few tasks
ws = min(pool, key=partial(self.worker_objective, ts))
tg.last_worker_tasks_left = math.floor(
(len(tg) / self.total_nthreads) * ws.nthreads
)
# Record `last_worker`, or clear it on the final task
tg.last_worker = (
ws if tg.states["released"] + tg.states["waiting"] > 1 else None
)
tg.last_worker_tasks_left -= 1
if self.validate and ws is not None:
assert self.workers.get(ws.address) is ws
assert ws in self.running, (ws, self.running)
return ws
def decide_worker_rootish_queuing_enabled(self) -> WorkerState | None:
"""Pick a worker for a runnable root-ish task, if not all are busy.
Picks the least-busy worker out of the ``idle`` workers (idle workers have fewer
tasks running than threads, as set by ``distributed.scheduler.worker-saturation``).
It does not consider the location of dependencies, since they'll end up on every
worker anyway.
If all workers are full, returns None, meaning the task should transition to
``queued``. The scheduler will wait to send it to a worker until a thread opens
up. This ensures that downstream tasks always run before new root tasks are
started.
This does not try to schedule sibling tasks on the same worker; in fact, it
usually does the opposite. Even though this increases subsequent data transfer,
it typically reduces overall memory use by eliminating root task overproduction.
Returns
-------
ws: WorkerState | None
The worker to assign the task to. If there are no idle workers, returns
None, in which case the task should be transitioned to ``queued``.
"""
if self.validate:
# We don't `assert self.is_rootish(ts)` here, because that check is
# dependent on cluster size. It's possible a task looked root-ish when it
# was queued, but the cluster has since scaled up and it no longer does when
# coming out of the queue. If `is_rootish` changes to a static definition,
# then add that assertion here (and actually pass in the task).
assert not math.isinf(self.WORKER_SATURATION)
if not self.idle_task_count:
# All workers busy? Task gets/stays queued.
return None
# Just pick the least busy worker.
# NOTE: this will lead to worst-case scheduling with regards to co-assignment.
ws = min(
self.idle_task_count,
key=lambda ws: len(ws.processing) / ws.nthreads,
)
if self.validate:
assert not _worker_full(ws, self.WORKER_SATURATION), (
ws,
_task_slots_available(ws, self.WORKER_SATURATION),
)
assert ws in self.running, (ws, self.running)
if self.validate and ws is not None:
assert self.workers.get(ws.address) is ws
assert ws in self.running, (ws, self.running)
return ws
def decide_worker_non_rootish(self, ts: TaskState) -> WorkerState | None:
"""Pick a worker for a runnable non-root task, considering dependencies and
restrictions.
Out of eligible workers holding dependencies of ``ts``, selects the worker
where, considering worker backlong and data-transfer costs, the task is
estimated to start running the soonest.
Returns
-------
ws: WorkerState | None
The worker to assign the task to. If no workers satisfy the restrictions of
``ts`` or there are no running workers, returns None, in which case the task
should be transitioned to ``no-worker``.
"""
if not self.running:
return None
valid_workers = self.valid_workers(ts)
if valid_workers is None and len(self.running) < len(self.workers):
if not self.running:
return None
# If there were no restrictions, `valid_workers()` didn't subset by
# `running`.
valid_workers = self.running
if ts.dependencies or valid_workers is not None:
ws = decide_worker(
ts,
self.running,
valid_workers,
partial(self.worker_objective, ts),
)
else:
# TODO if `is_rootish` would always return True for tasks without
# dependencies, we could remove all this logic. The rootish assignment logic
# would behave more or less the same as this, maybe without guaranteed
# round-robin though? This path is only reachable when `ts` doesn't have
# dependencies, but its group is also smaller than the cluster.
# Fastpath when there are no related tasks or restrictions
worker_pool = self.idle or self.workers
# FIXME idle and workers are SortedDict's declared as dicts
# because sortedcontainers is not annotated
wp_vals = cast("Sequence[WorkerState]", worker_pool.values())
n_workers = len(wp_vals)
if n_workers < 20: # smart but linear in small case
ws = min(wp_vals, key=operator.attrgetter("occupancy"))
assert ws
if ws.occupancy == 0:
# special case to use round-robin; linear search
# for next worker with zero occupancy (or just
# land back where we started).
start = self.n_tasks % n_workers
for i in range(n_workers):
wp_i = wp_vals[(i + start) % n_workers]
if wp_i.occupancy == 0:
ws = wp_i
break
else: # dumb but fast in large case
ws = wp_vals[self.n_tasks % n_workers]
if self.validate and ws is not None:
assert self.workers.get(ws.address) is ws
assert ws in self.running, (ws, self.running)
return ws
def transition_waiting_processing(self, key: str, stimulus_id: str) -> RecsMsgs:
"""Possibly schedule a ready task. This is the primary dispatch for ready tasks.
If there's no appropriate worker for the task (but the task is otherwise
runnable), it will be recommended to ``no-worker`` or ``queued``.
"""
ts = self.tasks[key]
if self.is_rootish(ts):
# NOTE: having two root-ish methods is temporary. When the feature flag is
# removed, there should only be one, which combines co-assignment and
# queuing. Eventually, special-casing root tasks might be removed entirely,
# with better heuristics.
if math.isinf(self.WORKER_SATURATION):
if not (ws := self.decide_worker_rootish_queuing_disabled(ts)):
return {ts.key: "no-worker"}, {}, {}
else:
if not (ws := self.decide_worker_rootish_queuing_enabled()):
return {ts.key: "queued"}, {}, {}
else:
if not (ws := self.decide_worker_non_rootish(ts)):
return {ts.key: "no-worker"}, {}, {}
worker_msgs = self._add_to_processing(ts, ws)
return {}, {}, worker_msgs
def transition_waiting_memory(
self,
key: str,
stimulus_id: str,
*,
nbytes: int | None = None,
type: bytes | None = None,
typename: str | None = None,
worker: str,
**kwargs: Any,
) -> RecsMsgs:
"""This transition exclusively happens in a race condition where the scheduler
believes that the only copy of a dependency task has just been lost, so it
transitions all dependents back to waiting, but actually a replica has already
been acquired by a worker computing the dependency - the scheduler just doesn't
know yet - and the execution finishes before the cancellation message from the
scheduler has a chance to reach the worker. Shortly, the cancellation request
will reach the worker, thus deleting the data from memory.
"""
ts = self.tasks[key]
if self.validate:
assert not ts.processing_on
assert ts.waiting_on
assert ts.state == "waiting"
return {}, {}, {}
def transition_processing_memory(
self,
key: str,
stimulus_id: str,
*,
nbytes: int | None = None,
type: bytes | None = None,
typename: str | None = None,
worker: str,
startstops: list[dict] | None = None,
**kwargs: Any,
) -> RecsMsgs:
ts = self.tasks[key]
assert worker
assert isinstance(worker, str)
if self.validate:
assert ts.processing_on
wss = ts.processing_on
assert wss
assert ts in wss.processing
del wss
assert not ts.waiting_on
assert not ts.who_has, (ts, ts.who_has)
assert not ts.exception_blame
assert ts.state == "processing"
ws = self.workers.get(worker)
if ws is None:
return {key: "released"}, {}, {}
if ws != ts.processing_on: # pragma: nocover
assert ts.processing_on
raise RuntimeError(
f"Task {ts.key!r} transitioned from processing to memory on worker "
f"{ws}, while it was expected from {ts.processing_on}. This should "
f"be impossible. {stimulus_id=}, story={self.story(ts)}"
)
#############################
# Update Timing Information #
#############################
if startstops:
for startstop in startstops:
ts.group.add_duration(
stop=startstop["stop"],
start=startstop["start"],
action=startstop["action"],
)
s = self.unknown_durations.pop(ts.prefix.name, set())
steal = self.extensions.get("stealing")
if steal:
for tts in s:
if tts.processing_on:
steal.recalculate_cost(tts)
############################
# Update State Information #
############################
if nbytes is not None:
ts.set_nbytes(nbytes)
self._exit_processing_common(ts)
recommendations: Recs = {}
client_msgs: Msgs = {}
self._add_to_memory(
ts, ws, recommendations, client_msgs, type=type, typename=typename
)
if self.validate:
assert not ts.processing_on
assert not ts.waiting_on
return recommendations, client_msgs, {}
def transition_memory_released(
self, key: str, stimulus_id: str, *, safe: bool = False
) -> RecsMsgs:
ts = self.tasks[key]
if self.validate:
assert not ts.waiting_on
assert not ts.processing_on
if safe:
assert not ts.waiters
if ts.actor:
for ws in ts.who_has:
ws.actors.discard(ts)
if ts.who_wants:
ts.exception_blame = ts
ts.exception = Serialized(
*serialize(ValueError("Worker holding Actor was lost"))
)
return {ts.key: "erred"}, {}, {} # don't try to recreate
recommendations: Recs = {}
client_msgs: Msgs = {}
worker_msgs: Msgs = {}
# XXX factor this out?
worker_msg = {
"op": "free-keys",
"keys": [key],
"stimulus_id": stimulus_id,
}
for ws in ts.who_has:
worker_msgs[ws.address] = [worker_msg]
self.remove_all_replicas(ts)
ts.state = "released"
report_msg = {"op": "lost-data", "key": key}
for cs in ts.who_wants:
client_msgs[cs.client_key] = [report_msg]
if not ts.run_spec: # pure data
recommendations[key] = "forgotten"
elif ts.has_lost_dependencies:
recommendations[key] = "forgotten"
elif ts.who_wants or ts.waiters:
recommendations[key] = "waiting"
for dts in ts.waiters:
if dts.state in ("no-worker", "processing"):
recommendations[dts.key] = "waiting"
elif dts.state == "waiting":
dts.waiting_on.add(ts)
if self.validate:
assert not ts.waiting_on
return recommendations, client_msgs, worker_msgs
def transition_released_erred(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
recommendations: Recs = {}
client_msgs: Msgs = {}
if self.validate:
with log_errors(pdb=LOG_PDB):
assert ts.exception_blame
assert not ts.who_has
assert not ts.waiting_on
assert not ts.waiters
failing_ts = ts.exception_blame
assert failing_ts
for dts in ts.dependents:
dts.exception_blame = failing_ts
if not dts.who_has:
recommendations[dts.key] = "erred"
report_msg = {
"op": "task-erred",
"key": key,
"exception": failing_ts.exception,
"traceback": failing_ts.traceback,
}
for cs in ts.who_wants:
client_msgs[cs.client_key] = [report_msg]
ts.state = "erred"
# TODO: waiting data?
return recommendations, client_msgs, {}
def transition_erred_released(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
recommendations: Recs = {}
client_msgs: Msgs = {}
worker_msgs: Msgs = {}
if self.validate:
with log_errors(pdb=LOG_PDB):
assert ts.exception_blame
assert not ts.who_has
assert not ts.waiting_on
assert not ts.waiters
ts.exception = None
ts.exception_blame = None
ts.traceback = None
for dts in ts.dependents:
if dts.state == "erred":
recommendations[dts.key] = "waiting"
w_msg = {
"op": "free-keys",
"keys": [key],
"stimulus_id": stimulus_id,
}
for ws_addr in ts.erred_on:
worker_msgs[ws_addr] = [w_msg]
ts.erred_on.clear()
report_msg = {"op": "task-retried", "key": key}
for cs in ts.who_wants:
client_msgs[cs.client_key] = [report_msg]
ts.state = "released"
return recommendations, client_msgs, worker_msgs
def transition_waiting_released(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
recommendations: Recs = {}
if self.validate:
assert not ts.who_has
assert not ts.processing_on
for dts in ts.dependencies:
if ts in dts.waiters:
dts.waiters.discard(ts)
if not dts.waiters and not dts.who_wants:
recommendations[dts.key] = "released"
ts.waiting_on.clear()
ts.state = "released"
if ts.has_lost_dependencies:
recommendations[key] = "forgotten"
elif not ts.exception_blame and (ts.who_wants or ts.waiters):
recommendations[key] = "waiting"
else:
ts.waiters.clear()
return recommendations, {}, {}
def transition_processing_released(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
recommendations: Recs = {}
worker_msgs: Msgs = {}
if self.validate:
assert ts.processing_on
assert not ts.who_has
assert not ts.waiting_on
assert ts.state == "processing"
ws = self._exit_processing_common(ts)
if ws:
worker_msgs[ws.address] = [
{
"op": "free-keys",
"keys": [key],
"stimulus_id": stimulus_id,
}
]
self._propagate_released(ts, recommendations)
return recommendations, {}, worker_msgs
def transition_processing_erred(
self,
key: str,
stimulus_id: str,
*,
worker: str,
cause: str | None = None,
exception: Serialized | None = None,
traceback: Serialized | None = None,
exception_text: str | None = None,
traceback_text: str | None = None,
**kwargs: Any,
) -> RecsMsgs:
"""Processed a recommended transition processing -> erred.
Parameters
----------
key
Key of the task to transition
stimulus_id
ID of the stimulus causing the transition
worker
Address of the worker where the task erred.
Not necessarily ``ts.processing_on``.
cause
Address of the task that caused this task to be transitioned to erred
exception
Exception caused by the task
traceback
Traceback caused by the task
exception_text
String representation of the exception
traceback_text
String representation of the traceback
Returns
-------
Recommendations, client messages and worker messages to process
"""
ts = self.tasks[key]
recommendations: Recs = {}
client_msgs: Msgs = {}
if self.validate:
assert cause or ts.exception_blame
assert ts.processing_on
assert not ts.who_has
assert not ts.waiting_on
if ts.actor:
ws = ts.processing_on
assert ws
ws.actors.remove(ts)
self._exit_processing_common(ts)
ts.erred_on.add(worker)
if exception is not None:
ts.exception = exception
ts.exception_text = exception_text # type: ignore
if traceback is not None:
ts.traceback = traceback
ts.traceback_text = traceback_text # type: ignore
if cause is not None:
failing_ts = self.tasks[cause]
ts.exception_blame = failing_ts
else:
failing_ts = ts.exception_blame # type: ignore
self.erred_tasks.appendleft(
ErredTask(
ts.key,
time(),
ts.erred_on.copy(),
exception_text or "",
traceback_text or "",
)
)
for dts in ts.dependents:
dts.exception_blame = failing_ts
recommendations[dts.key] = "erred"
for dts in ts.dependencies:
dts.waiters.discard(ts)
if not dts.waiters and not dts.who_wants:
recommendations[dts.key] = "released"
ts.waiters.clear() # do anything with this?
ts.state = "erred"
report_msg = {
"op": "task-erred",
"key": key,
"exception": failing_ts.exception,
"traceback": failing_ts.traceback,
}
for cs in ts.who_wants:
client_msgs[cs.client_key] = [report_msg]
cs = self.clients["fire-and-forget"]
if ts in cs.wants_what:
self._client_releases_keys(
cs=cs,
keys=[key],
recommendations=recommendations,
)
if self.validate:
assert not ts.processing_on
return recommendations, client_msgs, {}
def transition_no_worker_released(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
if self.validate:
assert self.tasks[key].state == "no-worker"
assert not ts.who_has
assert not ts.waiting_on
self.unrunnable.remove(ts)
ts.state = "released"
for dts in ts.dependencies:
dts.waiters.discard(ts)
ts.waiters.clear()
return {}, {}, {}
def transition_waiting_queued(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
if self.validate:
assert not self.idle_task_count, (ts, self.idle_task_count)
self._validate_ready(ts)
ts.state = "queued"
self.queued.add(ts)
return {}, {}, {}
def transition_waiting_no_worker(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
if self.validate:
self._validate_ready(ts)
ts.state = "no-worker"
self.unrunnable.add(ts)
return {}, {}, {}
def transition_queued_released(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
if self.validate:
assert ts in self.queued
assert not ts.processing_on
self.queued.remove(ts)
recommendations: Recs = {}
self._propagate_released(ts, recommendations)
return recommendations, {}, {}
def transition_queued_processing(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
recommendations: Recs = {}
worker_msgs: Msgs = {}
if self.validate:
assert not ts.actor, f"Actors can't be queued: {ts}"
assert ts in self.queued
if ws := self.decide_worker_rootish_queuing_enabled():
self.queued.discard(ts)
worker_msgs = self._add_to_processing(ts, ws)
# If no worker, task just stays `queued`
return recommendations, {}, worker_msgs
def _remove_key(self, key: str) -> None:
ts = self.tasks.pop(key)
assert ts.state == "forgotten"
self.unrunnable.discard(ts)
for cs in ts.who_wants:
cs.wants_what.remove(ts)
ts.who_wants.clear()
ts.processing_on = None
ts.exception_blame = ts.exception = ts.traceback = None
self.task_metadata.pop(key, None)
def transition_memory_forgotten(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
if self.validate:
assert ts.state == "memory"
assert not ts.processing_on
assert not ts.waiting_on
if not ts.run_spec:
# It's ok to forget a pure data task
pass
elif ts.has_lost_dependencies:
# It's ok to forget a task with forgotten dependencies
pass
elif not ts.who_wants and not ts.waiters and not ts.dependents:
# It's ok to forget a task that nobody needs
pass
else:
raise AssertionError("Unreachable", ts) # pragma: nocover
if ts.actor:
for ws in ts.who_has:
ws.actors.discard(ts)
recommendations: Recs = {}
worker_msgs: Msgs = {}
self._propagate_forgotten(ts, recommendations, worker_msgs, stimulus_id)
client_msgs = _task_to_client_msgs(ts)
self._remove_key(key)
return recommendations, client_msgs, worker_msgs
def transition_released_forgotten(self, key: str, stimulus_id: str) -> RecsMsgs:
ts = self.tasks[key]
if self.validate:
assert ts.state in ("released", "erred")
assert not ts.who_has
assert not ts.processing_on
assert ts not in self.queued
assert not ts.waiting_on, (ts, ts.waiting_on)
if not ts.run_spec:
# It's ok to forget a pure data task
pass
elif ts.has_lost_dependencies:
# It's ok to forget a task with forgotten dependencies
pass
elif not ts.who_wants and not ts.waiters and not ts.dependents:
# It's ok to forget a task that nobody needs
pass
else:
raise AssertionError("Unreachable", str(ts)) # pragma: nocover
recommendations: Recs = {}
worker_msgs: Msgs = {}
self._propagate_forgotten(ts, recommendations, worker_msgs, stimulus_id)
client_msgs = _task_to_client_msgs(ts)
self._remove_key(key)
return recommendations, client_msgs, worker_msgs
# {
# (start, finish):
# transition_<start>_<finish>(
# self, key: str, stimulus_id: str, **kwargs
# ) -> (recommendations, client_msgs, worker_msgs)
# }
_TRANSITIONS_TABLE: ClassVar[
Mapping[
tuple[TaskStateState, TaskStateState],
Callable[..., RecsMsgs],
]
] = {
("released", "waiting"): transition_released_waiting,
("waiting", "released"): transition_waiting_released,
("waiting", "processing"): transition_waiting_processing,
("waiting", "no-worker"): transition_waiting_no_worker,
("waiting", "queued"): transition_waiting_queued,
("waiting", "memory"): transition_waiting_memory,
("queued", "released"): transition_queued_released,
("queued", "processing"): transition_queued_processing,
("processing", "released"): transition_processing_released,
("processing", "memory"): transition_processing_memory,
("processing", "erred"): transition_processing_erred,
("no-worker", "released"): transition_no_worker_released,
("no-worker", "processing"): transition_no_worker_processing,
("released", "forgotten"): transition_released_forgotten,
("memory", "forgotten"): transition_memory_forgotten,
("erred", "released"): transition_erred_released,
("memory", "released"): transition_memory_released,
("released", "erred"): transition_released_erred,
}
def story(self, *keys_or_tasks_or_stimuli: str | TaskState) -> list[Transition]:
"""Get all transitions that touch one of the input keys or stimulus_id's"""
keys_or_stimuli = {
key.key if isinstance(key, TaskState) else key
for key in keys_or_tasks_or_stimuli
}
return scheduler_story(keys_or_stimuli, self.transition_log)
##############################
# Assigning Tasks to Workers #
##############################
def is_rootish(self, ts: TaskState) -> bool:
"""
Whether ``ts`` is a root or root-like task.
Root-ish tasks are part of a group that's much larger than the cluster,
and have few or no dependencies.
"""
if ts.resource_restrictions or ts.worker_restrictions or ts.host_restrictions:
return False
tg = ts.group
# TODO short-circuit to True if `not ts.dependencies`?
return (
len(tg) > self.total_nthreads * 2
and len(tg.dependencies) < 5
and sum(map(len, tg.dependencies)) < 5
)
def check_idle_saturated(self, ws: WorkerState, occ: float = -1.0) -> None:
"""Update the status of the idle and saturated state
The scheduler keeps track of workers that are ..
- Saturated: have enough work to stay busy
- Idle: do not have enough work to stay busy
They are considered saturated if they both have enough tasks to occupy
all of their threads, and if the expected runtime of those tasks is
large enough.
If ``distributed.scheduler.worker-saturation`` is not ``inf``
(scheduler-side queuing is enabled), they are considered idle
if they have fewer tasks processing than the ``worker-saturation``
threshold dictates.
Otherwise, they are considered idle if they have fewer tasks processing
than threads, or if their tasks' total expected runtime is less than half
the expected runtime of the same number of average tasks.
This is useful for load balancing and adaptivity.
"""
if self.total_nthreads == 0 or ws.status == Status.closed:
return
if occ < 0:
occ = ws.occupancy
p = len(ws.processing)
idle = self.idle
saturated = self.saturated
saturated.discard(ws)
if self.is_unoccupied(ws, occ, p):
if ws.status == Status.running:
idle[ws.address] = ws
else:
idle.pop(ws.address, None)
nc = ws.nthreads
if p > nc:
pending = occ * (p - nc) / (p * nc)
if 0.4 < pending > 1.9 * (self.total_occupancy / self.total_nthreads):
saturated.add(ws)
if not _worker_full(ws, self.WORKER_SATURATION):
if ws.status == Status.running:
self.idle_task_count.add(ws)
else:
self.idle_task_count.discard(ws)
def is_unoccupied(
self, ws: WorkerState, occupancy: float, nprocessing: int
) -> bool:
nthreads = ws.nthreads
return (
nprocessing < nthreads
or occupancy < nthreads * (self.total_occupancy / self.total_nthreads) / 2
)
def get_comm_cost(self, ts: TaskState, ws: WorkerState) -> float:
"""
Get the estimated communication cost (in s.) to compute the task
on the given worker.
"""
if 10 * len(ts.dependencies) < len(ws.has_what):
# In the common case where the number of dependencies is
# much less than the number of tasks that we have,
# construct the set of deps that require communication in
# O(len(dependencies)) rather than O(len(has_what)) time.
# Factor of 10 is a guess at the overhead of explicit
# iteration as opposed to just calling set.difference
deps = {dep for dep in ts.dependencies if dep not in ws.has_what}
else:
deps = ts.dependencies.difference(ws.has_what)
nbytes = sum(dts.nbytes for dts in deps)
return nbytes / self.bandwidth
def get_task_duration(self, ts: TaskState) -> float:
"""Get the estimated computation cost of the given task (not including
any communication cost).
If no data has been observed, value of
`distributed.scheduler.default-task-durations` are used. If none is set
for this task, `distributed.scheduler.unknown-task-duration` is used
instead.
"""
duration: float = ts.prefix.duration_average
if duration >= 0:
return duration
s = self.unknown_durations.get(ts.prefix.name)
if s is None:
self.unknown_durations[ts.prefix.name] = s = set()
s.add(ts)
return self.UNKNOWN_TASK_DURATION
def valid_workers(self, ts: TaskState) -> set[WorkerState] | None:
"""Return set of currently valid workers for key
If all workers are valid then this returns ``None``, in which case
any *running* worker can be used.
Otherwise, the subset of running workers valid for this task
is returned.
This checks tracks the following state:
* worker_restrictions
* host_restrictions
* resource_restrictions
"""
s: set[str] | None = None
if ts.worker_restrictions:
s = {addr for addr in ts.worker_restrictions if addr in self.workers}
if ts.host_restrictions:
# Resolve the alias here rather than early, for the worker
# may not be connected when host_restrictions is populated
hr = [self.coerce_hostname(h) for h in ts.host_restrictions]
# XXX need HostState?
sl = []
for h in hr:
dh = self.host_info.get(h)
if dh is not None:
sl.append(dh["addresses"])
ss = set.union(*sl) if sl else set()
if s is None:
s = ss
else:
s |= ss
if ts.resource_restrictions:
dw = {}
for resource, required in ts.resource_restrictions.items():
dr = self.resources.get(resource)
if dr is None:
self.resources[resource] = dr = {}
sw = set()
for addr, supplied in dr.items():
if supplied >= required:
sw.add(addr)
dw[resource] = sw
ww = set.intersection(*dw.values())
if s is None:
s = ww
else:
s &= ww
if s is None:
return None # All workers are valid
if not s:
return set() # No workers are valid
# Some workers are valid
s_ws = {self.workers[addr] for addr in s}
if len(self.running) < len(self.workers):
s_ws &= self.running
return s_ws
def acquire_resources(self, ts: TaskState, ws: WorkerState) -> None:
for r, required in ts.resource_restrictions.items():
ws.used_resources[r] += required
def release_resources(self, ts: TaskState, ws: WorkerState) -> None:
for r, required in ts.resource_restrictions.items():
ws.used_resources[r] -= required
def coerce_hostname(self, host: Hashable) -> str:
"""
Coerce the hostname of a worker.
"""
alias = self.aliases.get(host)
if alias is not None:
ws = self.workers[alias]
return ws.host
else:
assert isinstance(host, str)
return host
def worker_objective(self, ts: TaskState, ws: WorkerState) -> tuple:
"""Objective function to determine which worker should get the task
Minimize expected start time. If a tie then break with data storage.
"""
comm_bytes = sum(
dts.get_nbytes() for dts in ts.dependencies if ws not in dts.who_has
)
stack_time = ws.occupancy / ws.nthreads
start_time = stack_time + comm_bytes / self.bandwidth
if ts.actor:
return (len(ws.actors), start_time, ws.nbytes)
else:
return (start_time, ws.nbytes)
def add_replica(self, ts: TaskState, ws: WorkerState) -> None:
"""Note that a worker holds a replica of a task with state='memory'"""
ws.add_replica(ts)
if len(ts.who_has) == 2:
self.replicated_tasks.add(ts)
def remove_replica(self, ts: TaskState, ws: WorkerState) -> None:
"""Note that a worker no longer holds a replica of a task"""
ws.remove_replica(ts)
if len(ts.who_has) == 1:
self.replicated_tasks.remove(ts)
def remove_all_replicas(self, ts: TaskState) -> None:
"""Remove all replicas of a task from all workers"""
nbytes = ts.get_nbytes()
for ws in ts.who_has:
ws.nbytes -= nbytes
del ws._has_what[ts]
if len(ts.who_has) > 1:
self.replicated_tasks.remove(ts)
ts.who_has.clear()
def bulk_schedule_unrunnable_after_adding_worker(self, ws: WorkerState) -> Recs:
"""Send ``no-worker`` tasks to ``processing`` that this worker can handle.
Returns priority-ordered recommendations.
"""
runnable: list[TaskState] = []
for ts in self.unrunnable:
valid = self.valid_workers(ts)
if valid is None or ws in valid:
runnable.append(ts)
# Recommendations are processed LIFO, hence the reversed order
runnable.sort(key=operator.attrgetter("priority"), reverse=True)
return {ts.key: "processing" for ts in runnable}
def _validate_ready(self, ts: TaskState) -> None:
"""Validation for ready states (processing, queued, no-worker)"""
assert not ts.waiting_on
assert not ts.who_has
assert not ts.exception_blame
assert not ts.processing_on
assert not ts.has_lost_dependencies
assert ts not in self.unrunnable
assert ts not in self.queued
assert all(dts.who_has for dts in ts.dependencies)
def _add_to_processing(self, ts: TaskState, ws: WorkerState) -> Msgs:
"""Set a task as processing on a worker and return the worker messages to send"""
if self.validate:
self._validate_ready(ts)
assert ws in self.running, self.running
assert (o := self.workers.get(ws.address)) is ws, (ws, o)
ws.add_to_processing(ts)
ts.processing_on = ws
ts.state = "processing"
self.acquire_resources(ts, ws)
self.check_idle_saturated(ws)
self.n_tasks += 1
if ts.actor:
ws.actors.add(ts)
return {ws.address: [self._task_to_msg(ts)]}
def _exit_processing_common(self, ts: TaskState) -> WorkerState | None:
"""Remove *ts* from the set of processing tasks.
Returns
-------
Worker state of the worker that processed *ts* if the worker is current,
None if the worker is stale.
See also
--------
Scheduler._set_duration_estimate
"""
ws = ts.processing_on
assert ws
ts.processing_on = None
ws.remove_from_processing(ts)
if self.workers.get(ws.address) is not ws: # may have been removed
return None
self.check_idle_saturated(ws)
self.release_resources(ts, ws)
return ws
def _add_to_memory(
self,
ts: TaskState,
ws: WorkerState,
recommendations: Recs,
client_msgs: Msgs,
type: bytes | None = None,
typename: str | None = None,
) -> None:
"""Add ts to the set of in-memory tasks"""
if self.validate:
assert ts not in ws.has_what
self.add_replica(ts, ws)
deps = list(ts.dependents)
if len(deps) > 1:
deps.sort(key=operator.attrgetter("priority"), reverse=True)
for dts in deps:
s = dts.waiting_on
if ts in s:
s.discard(ts)
if not s: # new task ready to run
recommendations[dts.key] = "processing"
for dts in ts.dependencies:
s = dts.waiters
s.discard(ts)
if not s and not dts.who_wants:
recommendations[dts.key] = "released"
if not ts.waiters and not ts.who_wants:
recommendations[ts.key] = "released"
else:
report_msg: dict[str, Any] = {"op": "key-in-memory", "key": ts.key}
if type is not None:
report_msg["type"] = type
for cs in ts.who_wants:
client_msgs[cs.client_key] = [report_msg]
ts.state = "memory"
ts.type = typename # type: ignore
ts.group.types.add(typename) # type: ignore
cs = self.clients["fire-and-forget"]
if ts in cs.wants_what:
self._client_releases_keys(
cs=cs,
keys=[ts.key],
recommendations=recommendations,
)
def _propagate_released(self, ts: TaskState, recommendations: Recs) -> None:
ts.state = "released"
key = ts.key
if ts.has_lost_dependencies:
recommendations[key] = "forgotten"
elif ts.waiters or ts.who_wants:
recommendations[key] = "waiting"
if recommendations.get(key) != "waiting":
for dts in ts.dependencies:
if dts.state != "released":
dts.waiters.discard(ts)
if not dts.waiters and not dts.who_wants:
recommendations[dts.key] = "released"
ts.waiters.clear()
if self.validate:
assert not ts.processing_on
assert ts not in self.queued
def _propagate_forgotten(
self,
ts: TaskState,
recommendations: Recs,
worker_msgs: Msgs,
stimulus_id: str,
) -> None:
ts.state = "forgotten"
for dts in ts.dependents:
dts.has_lost_dependencies = True
dts.dependencies.remove(ts)
dts.waiting_on.discard(ts)
if dts.state not in ("memory", "erred"):
# Cannot compute task anymore
recommendations[dts.key] = "forgotten"
ts.dependents.clear()
ts.waiters.clear()
for dts in ts.dependencies:
dts.dependents.remove(ts)
dts.waiters.discard(ts)
if not dts.dependents and not dts.who_wants:
# Task not needed anymore
assert dts is not ts
recommendations[dts.key] = "forgotten"
ts.dependencies.clear()
ts.waiting_on.clear()
for ws in ts.who_has:
if ws.address in self.workers: # in case worker has died
worker_msgs[ws.address] = [
{
"op": "free-keys",
"keys": [ts.key],
"stimulus_id": stimulus_id,
}
]
self.remove_all_replicas(ts)
def _client_releases_keys(
self,
keys: Collection[str],
cs: ClientState,
recommendations: Recs,
) -> None:
"""Remove keys from client desired list"""
logger.debug("Client %s releases keys: %s", cs.client_key, keys)
for key in keys:
ts = self.tasks.get(key)
if ts is not None and ts in cs.wants_what:
cs.wants_what.remove(ts)
ts.who_wants.remove(cs)
if not ts.who_wants:
if not ts.dependents:
# No live dependents, can forget
recommendations[ts.key] = "forgotten"
elif ts.state != "erred" and not ts.waiters:
recommendations[ts.key] = "released"
def _task_to_msg(self, ts: TaskState, duration: float = -1) -> dict[str, Any]:
"""Convert a single computational task to a message"""
# FIXME: The duration attribute is not used on worker. We could save ourselves the
# time to compute and submit this
if duration < 0:
duration = self.get_task_duration(ts)
msg: dict[str, Any] = {
"op": "compute-task",
"key": ts.key,
"priority": ts.priority,
"duration": duration,
"stimulus_id": f"compute-task-{time()}",
"who_has": {
dts.key: [ws.address for ws in dts.who_has] for dts in ts.dependencies
},
"nbytes": {dts.key: dts.nbytes for dts in ts.dependencies},
"run_spec": None,
"function": None,
"args": None,
"kwargs": None,
"resource_restrictions": ts.resource_restrictions,
"actor": ts.actor,
"annotations": ts.annotations,
}
if self.validate:
assert all(msg["who_has"].values())
if isinstance(ts.run_spec, dict):
msg.update(ts.run_spec)
else:
msg["run_spec"] = ts.run_spec
return msg
class Scheduler(SchedulerState, ServerNode):
"""Dynamic distributed task scheduler
The scheduler tracks the current state of workers, data, and computations.
The scheduler listens for events and responds by controlling workers
appropriately. It continuously tries to use the workers to execute an ever
growing dask graph.
All events are handled quickly, in linear time with respect to their input
(which is often of constant size) and generally within a millisecond. To
accomplish this the scheduler tracks a lot of state. Every operation
maintains the consistency of this state.
The scheduler communicates with the outside world through Comm objects.
It maintains a consistent and valid view of the world even when listening
to several clients at once.
A Scheduler is typically started either with the ``dask scheduler``
executable::
$ dask scheduler
Scheduler started at 127.0.0.1:8786
Or within a LocalCluster a Client starts up without connection
information::
>>> c = Client() # doctest: +SKIP
>>> c.cluster.scheduler # doctest: +SKIP
Scheduler(...)
Users typically do not interact with the scheduler directly but rather with
the client object ``Client``.
The ``contact_address`` parameter allows to advertise a specific address to
the workers for communication with the scheduler, which is different than
the address the scheduler binds to. This is useful when the scheduler
listens on a private address, which therefore cannot be used by the workers
to contact it.
**State**
The scheduler contains the following state variables. Each variable is
listed along with what it stores and a brief description.
* **tasks:** ``{task key: TaskState}``
Tasks currently known to the scheduler
* **unrunnable:** ``{TaskState}``
Tasks in the "no-worker" state
* **workers:** ``{worker key: WorkerState}``
Workers currently connected to the scheduler
* **idle:** ``{WorkerState}``:
Set of workers that are not fully utilized
* **saturated:** ``{WorkerState}``:
Set of workers that are not over-utilized
* **host_info:** ``{hostname: dict}``:
Information about each worker host
* **clients:** ``{client key: ClientState}``
Clients currently connected to the scheduler
* **services:** ``{str: port}``:
Other services running on this scheduler, like Bokeh
* **loop:** ``IOLoop``:
The running Tornado IOLoop
* **client_comms:** ``{client key: Comm}``
For each client, a Comm object used to receive task requests and
report task status updates.
* **stream_comms:** ``{worker key: Comm}``
For each worker, a Comm object from which we both accept stimuli and
report results
* **task_duration:** ``{key-prefix: time}``
Time we expect certain functions to take, e.g. ``{'sum': 0.25}``
"""
default_port = 8786
_instances: ClassVar[weakref.WeakSet[Scheduler]] = weakref.WeakSet()
def __init__(
self,
loop=None,
delete_interval="500ms",
synchronize_worker_interval="60s",
services=None,
service_kwargs=None,
allowed_failures=None,
extensions=None,
validate=None,
scheduler_file=None,
security=None,
worker_ttl=None,
idle_timeout=None,
interface=None,
host=None,
port=0,
protocol=None,
dashboard_address=None,
dashboard=None,
http_prefix="/",
preload=None,
preload_argv=(),
plugins=(),
contact_address=None,
transition_counter_max=False,
jupyter=False,
**kwargs,
):
if loop is not None:
warnings.warn(
"the loop kwarg to Scheduler is deprecated",
DeprecationWarning,
stacklevel=2,
)
self.loop = self.io_loop = IOLoop.current()
self._setup_logging(logger)
# Attributes
if contact_address is None:
contact_address = dask.config.get("distributed.scheduler.contact-address")
self.contact_address = contact_address
if allowed_failures is None:
allowed_failures = dask.config.get("distributed.scheduler.allowed-failures")
self.allowed_failures = allowed_failures
if validate is None:
validate = dask.config.get("distributed.scheduler.validate")
self.proc = psutil.Process()
self.delete_interval = parse_timedelta(delete_interval, default="ms")
self.synchronize_worker_interval = parse_timedelta(
synchronize_worker_interval, default="ms"
)
self.service_specs = services or {}
self.service_kwargs = service_kwargs or {}
self.services = {}
self.scheduler_file = scheduler_file
worker_ttl = worker_ttl or dask.config.get("distributed.scheduler.worker-ttl")
self.worker_ttl = parse_timedelta(worker_ttl) if worker_ttl else None
idle_timeout = idle_timeout or dask.config.get(
"distributed.scheduler.idle-timeout"
)
if idle_timeout:
self.idle_timeout = parse_timedelta(idle_timeout)
else:
self.idle_timeout = None
self.idle_since = time()
self.time_started = self.idle_since # compatibility for dask-gateway
self._lock = asyncio.Lock()
self.bandwidth_workers = defaultdict(float)
self.bandwidth_types = defaultdict(float)
if not preload:
preload = dask.config.get("distributed.scheduler.preload")
if not preload_argv:
preload_argv = dask.config.get("distributed.scheduler.preload-argv")
self.preloads = preloading.process_preloads(self, preload, preload_argv)
if isinstance(security, dict):
security = Security(**security)
self.security = security or Security()
assert isinstance(self.security, Security)
self.connection_args = self.security.get_connection_args("scheduler")
self.connection_args["handshake_overrides"] = { # common denominator
"pickle-protocol": 4
}
self._start_address = addresses_from_user_args(
host=host,
port=port,
interface=interface,
protocol=protocol,
security=security,
default_port=self.default_port,
)
http_server_modules = dask.config.get("distributed.scheduler.http.routes")
show_dashboard = dashboard or (dashboard is None and dashboard_address)
# install vanilla route if show_dashboard but bokeh is not installed
if show_dashboard:
try:
import distributed.dashboard.scheduler
except ImportError:
show_dashboard = False
http_server_modules.append("distributed.http.scheduler.missing_bokeh")
routes = get_handlers(
server=self, modules=http_server_modules, prefix=http_prefix
)
self.start_http_server(routes, dashboard_address, default_port=8787)
if show_dashboard:
distributed.dashboard.scheduler.connect(
self.http_application, self.http_server, self, prefix=http_prefix
)
self.jupyter = jupyter
if self.jupyter:
try:
from jupyter_server.serverapp import ServerApp
except ImportError:
raise ImportError(
"In order to use the Dask jupyter option you "
"need to have jupyterlab installed"
)
from traitlets.config import Config
j = ServerApp.instance(
config=Config(
{
"ServerApp": {
"base_url": "jupyter",
# SECURITY: We usually expect the dashboard to be a read-only view into
# the scheduler activity. However, by adding an open Jupyter application
# we are allowing arbitrary remote code execution on the scheduler via the
# dashboard server. This option should only be used when the dashboard is
# protected via other means, or when you don't care about cluster security.
"token": "",
"allow_remote_access": True,
}
}
)
)
j.initialize(
new_httpserver=False,
)
self._jupyter_server_application = j
self.http_application.add_application(j.web_app)
# Communication state
self.client_comms = {}
self.stream_comms = {}
# Task state
tasks = {}
self.generation = 0
self._last_client = None
self._last_time = 0
unrunnable = set()
queued: HeapSet[TaskState] = HeapSet(key=operator.attrgetter("priority"))
self.datasets = {}
# Prefix-keyed containers
# Client state
clients = {}
# Worker state
workers = SortedDict()
host_info = {}
resources = {}
aliases = {}
self._worker_collections = [
workers,
host_info,
resources,
aliases,
]
self.events = defaultdict(
partial(
deque, maxlen=dask.config.get("distributed.scheduler.events-log-length")
)
)
self.event_counts = defaultdict(int)
self.event_subscriber = defaultdict(set)
self.worker_plugins = {}
self.nanny_plugins = {}
worker_handlers = {
"task-finished": self.handle_task_finished,
"task-erred": self.handle_task_erred,
"release-worker-data": self.release_worker_data,
"add-keys": self.add_keys,
"long-running": self.handle_long_running,
"reschedule": self._reschedule,
"keep-alive": lambda *args, **kwargs: None,
"log-event": self.log_worker_event,
"worker-status-change": self.handle_worker_status_change,
"request-refresh-who-has": self.handle_request_refresh_who_has,
}
client_handlers = {
"update-graph": self.update_graph,
"update-graph-hlg": self.update_graph_hlg,
"client-desires-keys": self.client_desires_keys,
"update-data": self.update_data,
"report-key": self.report_on_key,
"client-releases-keys": self.client_releases_keys,
"heartbeat-client": self.client_heartbeat,
"close-client": self.remove_client,
"subscribe-topic": self.subscribe_topic,
"unsubscribe-topic": self.unsubscribe_topic,
}
self.handlers = {
"register-client": self.add_client,
"scatter": self.scatter,
"register-worker": self.add_worker,
"register_nanny": self.add_nanny,
"unregister": self.remove_worker,
"gather": self.gather,
"cancel": self.stimulus_cancel,
"retry": self.stimulus_retry,
"feed": self.feed,
"terminate": self.close,
"broadcast": self.broadcast,
"proxy": self.proxy,
"ncores": self.get_ncores,
"ncores_running": self.get_ncores_running,
"has_what": self.get_has_what,
"who_has": self.get_who_has,
"processing": self.get_processing,
"call_stack": self.get_call_stack,
"profile": self.get_profile,
"performance_report": self.performance_report,
"get_logs": self.get_logs,
"logs": self.get_logs,
"worker_logs": self.get_worker_logs,
"log_event": self.log_event,
"events": self.get_events,
"nbytes": self.get_nbytes,
"versions": self.versions,
"add_keys": self.add_keys,
"rebalance": self.rebalance,
"replicate": self.replicate,
"run_function": self.run_function,
"restart": self.restart,
"update_data": self.update_data,
"set_resources": self.add_resources,
"retire_workers": self.retire_workers,
"get_metadata": self.get_metadata,
"set_metadata": self.set_metadata,
"set_restrictions": self.set_restrictions,
"heartbeat_worker": self.heartbeat_worker,
"get_task_status": self.get_task_status,
"get_task_stream": self.get_task_stream,
"get_task_prefix_states": self.get_task_prefix_states,
"register_scheduler_plugin": self.register_scheduler_plugin,
"register_worker_plugin": self.register_worker_plugin,
"unregister_worker_plugin": self.unregister_worker_plugin,
"register_nanny_plugin": self.register_nanny_plugin,
"unregister_nanny_plugin": self.unregister_nanny_plugin,
"adaptive_target": self.adaptive_target,
"workers_to_close": self.workers_to_close,
"subscribe_worker_status": self.subscribe_worker_status,
"start_task_metadata": self.start_task_metadata,
"stop_task_metadata": self.stop_task_metadata,
"get_cluster_state": self.get_cluster_state,
"dump_cluster_state_to_url": self.dump_cluster_state_to_url,
"benchmark_hardware": self.benchmark_hardware,
"get_story": self.get_story,
}
connection_limit = get_fileno_limit() / 2
SchedulerState.__init__(
self,
aliases=aliases,
clients=clients,
workers=workers,
host_info=host_info,
resources=resources,
tasks=tasks,
unrunnable=unrunnable,
queued=queued,
validate=validate,
plugins=plugins,
transition_counter_max=transition_counter_max,
)
ServerNode.__init__(
self,
handlers=self.handlers,
stream_handlers=merge(worker_handlers, client_handlers),
connection_limit=connection_limit,
deserialize=False,
connection_args=self.connection_args,
**kwargs,
)
if self.worker_ttl:
pc = PeriodicCallback(self.check_worker_ttl, self.worker_ttl * 1000)
self.periodic_callbacks["worker-ttl"] = pc
if self.idle_timeout:
pc = PeriodicCallback(self.check_idle, self.idle_timeout * 1000 / 4)
self.periodic_callbacks["idle-timeout"] = pc
if extensions is None:
extensions = DEFAULT_EXTENSIONS.copy()
if not dask.config.get("distributed.scheduler.work-stealing"):
if "stealing" in extensions:
del extensions["stealing"]
for name, extension in extensions.items():
self.extensions[name] = extension(self)
setproctitle("dask scheduler [not started]")
Scheduler._instances.add(self)
self.rpc.allow_offload = False
##################
# Administration #
##################
def __repr__(self):
return (
f"<Scheduler {self.address_safe!r}, "
f"workers: {len(self.workers)}, "
f"cores: {self.total_nthreads}, "
f"tasks: {len(self.tasks)}>"
)
def _repr_html_(self):
return get_template("scheduler.html.j2").render(
address=self.address,
workers=self.workers,
threads=self.total_nthreads,
tasks=self.tasks,
)
def identity(self):
"""Basic information about ourselves and our cluster"""
d = {
"type": type(self).__name__,
"id": str(self.id),
"address": self.address,
"services": {key: v.port for (key, v) in self.services.items()},
"started": self.time_started,
"workers": {
worker.address: worker.identity() for worker in self.workers.values()
},
}
return d
def _to_dict(self, *, exclude: Container[str] = ()) -> dict:
"""Dictionary representation for debugging purposes.
Not type stable and not intended for roundtrips.
See also
--------
Server.identity
Client.dump_cluster_state
distributed.utils.recursive_to_dict
"""
info = super()._to_dict(exclude=exclude)
extra = {
"transition_log": self.transition_log,
"transition_counter": self.transition_counter,
"tasks": self.tasks,
"task_groups": self.task_groups,
# Overwrite dict of WorkerState.identity from info
"workers": self.workers,
"clients": self.clients,
"memory": self.memory,
"events": self.events,
"extensions": self.extensions,
}
extra = {k: v for k, v in extra.items() if k not in exclude}
info.update(recursive_to_dict(extra, exclude=exclude))
return info
async def get_cluster_state(
self,
exclude: "Collection[str]",
) -> dict:
"Produce the state dict used in a cluster state dump"
# Kick off state-dumping on workers before we block the event loop in `self._to_dict`.
workers_future = asyncio.gather(
self.broadcast(
msg={"op": "dump_state", "exclude": exclude},
on_error="return",
),
self.broadcast(
msg={"op": "versions"},
on_error="ignore",
),
)
try:
scheduler_state = self._to_dict(exclude=exclude)
worker_states, worker_versions = await workers_future
finally:
# Ensure the tasks aren't left running if anything fails.
# Someday (py3.11), use a trio-style TaskGroup for this.
workers_future.cancel()
# Convert any RPC errors to strings
worker_states = {
k: repr(v) if isinstance(v, Exception) else v
for k, v in worker_states.items()
}
return {
"scheduler": scheduler_state,
"workers": worker_states,
"versions": {"scheduler": self.versions(), "workers": worker_versions},
}
async def dump_cluster_state_to_url(
self,
url: str,
exclude: "Collection[str]",
format: Literal["msgpack", "yaml"],
**storage_options: dict[str, Any],
) -> None:
"Write a cluster state dump to an fsspec-compatible URL."
await cluster_dump.write_state(
partial(self.get_cluster_state, exclude), url, format, **storage_options
)
def get_worker_service_addr(
self, worker: str, service_name: str, protocol: bool = False
) -> tuple[str, int] | str | None:
"""
Get the (host, port) address of the named service on the *worker*.
Returns None if the service doesn't exist.
Parameters
----------
worker : address
service_name : str
Common services include 'bokeh' and 'nanny'
protocol : boolean
Whether or not to include a full address with protocol (True)
or just a (host, port) pair
"""
ws = self.workers[worker]
port = ws.services.get(service_name)
if port is None:
return None
elif protocol:
return "%(protocol)s://%(host)s:%(port)d" % {
"protocol": ws.address.split("://")[0],
"host": ws.host,
"port": port,
}
else:
return ws.host, port
async def start_unsafe(self):
"""Clear out old state and restart all running coroutines"""
await super().start_unsafe()
enable_gc_diagnosis()
self._clear_task_state()
for addr in self._start_address:
await self.listen(
addr,
allow_offload=False,
handshake_overrides={"pickle-protocol": 4, "compression": None},
**self.security.get_listen_args("scheduler"),
)
self.ip = get_address_host(self.listen_address)
listen_ip = self.ip
if listen_ip == "0.0.0.0":
listen_ip = ""
if self.address.startswith("inproc://"):
listen_ip = "localhost"
# Services listen on all addresses
self.start_services(listen_ip)
for listener in self.listeners:
logger.info(" Scheduler at: %25s", listener.contact_address)
for k, v in self.services.items():
logger.info("%11s at: %25s", k, "%s:%d" % (listen_ip, v.port))
if self.scheduler_file:
with open(self.scheduler_file, "w") as f:
json.dump(self.identity(), f, indent=2)
fn = self.scheduler_file # remove file when we close the process
def del_scheduler_file():
if os.path.exists(fn):
os.remove(fn)
weakref.finalize(self, del_scheduler_file)
for preload in self.preloads:
try:
await preload.start()
except Exception:
logger.exception("Failed to start preload")
if self.jupyter:
# Allow insecure communications from local users
if self.address.startswith("tls://"):
await self.listen("tcp://localhost:0")
os.environ["DASK_SCHEDULER_ADDRESS"] = self.listeners[-1].contact_address
await asyncio.gather(
*[plugin.start(self) for plugin in list(self.plugins.values())]
)
self.start_periodic_callbacks()
setproctitle(f"dask scheduler [{self.address}]")
return self
async def close(self, fast=None, close_workers=None):
"""Send cleanup signal to all coroutines then wait until finished
See Also
--------
Scheduler.cleanup
"""
if fast is not None or close_workers is not None:
warnings.warn(
"The 'fast' and 'close_workers' parameters in Scheduler.close have no effect and will be removed in a future version of distributed.",
FutureWarning,
)
if self.status in (Status.closing, Status.closed):
await self.finished()
return
async def log_errors(func):
try:
await func()
except Exception:
logger.exception("Plugin call failed during scheduler.close")
await asyncio.gather(
*[log_errors(plugin.before_close) for plugin in list(self.plugins.values())]
)
self.status = Status.closing
logger.info("Scheduler closing...")
setproctitle("dask scheduler [closing]")
for preload in self.preloads:
try:
await preload.teardown()
except Exception:
logger.exception("Failed to tear down preload")
await asyncio.gather(
*[log_errors(plugin.close) for plugin in list(self.plugins.values())]
)
for pc in self.periodic_callbacks.values():
pc.stop()
self.periodic_callbacks.clear()
self.stop_services()
for ext in self.extensions.values():
with suppress(AttributeError):
ext.teardown()
logger.info("Scheduler closing all comms")
futures = []
for _, comm in list(self.stream_comms.items()):
# FIXME use `self.remove_worker()` instead after https://github.com/dask/distributed/issues/6390
if not comm.closed():
# This closes the Worker and ensures that if a Nanny is around,
# it is closed as well
comm.send({"op": "close", "reason": "scheduler-close"})
comm.send({"op": "close-stream"})
# ^ TODO remove? `Worker.close` will close the stream anyway.
with suppress(AttributeError):
futures.append(comm.close())
await asyncio.gather(*futures)
if self.jupyter:
await self._jupyter_server_application._cleanup()
for comm in self.client_comms.values():
comm.abort()
await self.rpc.close()
self.status = Status.closed
self.stop()
await super().close()
setproctitle("dask scheduler [closed]")
disable_gc_diagnosis()
###########
# Stimuli #
###########
def heartbeat_worker(
self,
comm=None,
*,
address,
resolve_address: bool = True,
now: float | None = None,
resources: dict[str, float] | None = None,
host_info: dict | None = None,
metrics: dict,
executing: dict[str, float] | None = None,
extensions: dict | None = None,
) -> dict[str, Any]:
address = self.coerce_address(address, resolve_address)
address = normalize_address(address)
ws = self.workers.get(address)
if ws is None:
logger.warning(f"Received heartbeat from unregistered worker {address!r}.")
return {"status": "missing"}
host = get_address_host(address)
local_now = time()
host_info = host_info or {}
dh: dict = self.host_info.setdefault(host, {})
dh["last-seen"] = local_now
frac = 1 / len(self.workers)
self.bandwidth = (
self.bandwidth * (1 - frac) + metrics["bandwidth"]["total"] * frac
)
for other, (bw, count) in metrics["bandwidth"]["workers"].items():
if (address, other) not in self.bandwidth_workers:
self.bandwidth_workers[address, other] = bw / count
else:
alpha = (1 - frac) ** count
self.bandwidth_workers[address, other] = self.bandwidth_workers[
address, other
] * alpha + bw * (1 - alpha)
for typ, (bw, count) in metrics["bandwidth"]["types"].items():
if typ not in self.bandwidth_types:
self.bandwidth_types[typ] = bw / count
else:
alpha = (1 - frac) ** count
self.bandwidth_types[typ] = self.bandwidth_types[typ] * alpha + bw * (
1 - alpha
)
ws.last_seen = local_now
if executing is not None:
# NOTE: the executing dict is unused
ws.executing = {}
for key, duration in executing.items():
if key in self.tasks:
ts = self.tasks[key]
ws.executing[ts] = duration
ts.prefix.add_exec_time(duration)
ws.metrics = metrics
# Calculate RSS - dask keys, separating "old" and "new" usage
# See MemoryState for details
max_memory_unmanaged_old_hist_age = local_now - self.MEMORY_RECENT_TO_OLD_TIME
memory_unmanaged_old = ws._memory_unmanaged_old
while ws._memory_unmanaged_history:
timestamp, size = ws._memory_unmanaged_history[0]
if timestamp >= max_memory_unmanaged_old_hist_age:
break
ws._memory_unmanaged_history.popleft()
if size == memory_unmanaged_old:
memory_unmanaged_old = 0 # recalculate min()
# ws._nbytes is updated at a different time and sizeof() may not be accurate,
# so size may be (temporarily) negative; floor it to zero.
size = max(
0, metrics["memory"] - ws.nbytes + metrics["spilled_bytes"]["memory"]
)
ws._memory_unmanaged_history.append((local_now, size))
if not memory_unmanaged_old:
# The worker has just been started or the previous minimum has been expunged
# because too old.
# Note: this algorithm is capped to 200 * MEMORY_RECENT_TO_OLD_TIME elements
# cluster-wide by heartbeat_interval(), regardless of the number of workers
ws._memory_unmanaged_old = min(map(second, ws._memory_unmanaged_history))
elif size < memory_unmanaged_old:
ws._memory_unmanaged_old = size
if host_info:
dh = self.host_info.setdefault(host, {})
dh.update(host_info)
if now:
ws.time_delay = local_now - now
if resources:
self.add_resources(worker=address, resources=resources)
if extensions:
for name, data in extensions.items():
self.extensions[name].heartbeat(ws, data)
return {
"status": "OK",
"time": local_now,
"heartbeat-interval": heartbeat_interval(len(self.workers)),
}
@log_errors
async def add_worker(
self,
comm: Comm,
*,
address: str,
status: str,
server_id: str,
keys=(),
nthreads=None,
name=None,
resolve_address=True,
nbytes=None,
types=None,
now=None,
resources=None,
host_info=None,
memory_limit=None,
metrics=None,
pid=0,
services=None,
local_directory=None,
versions: dict[str, Any] | None = None,
nanny=None,
extra=None,
stimulus_id: str,
) -> None:
"""Add a new worker to the cluster"""
address = self.coerce_address(address, resolve_address)
address = normalize_address(address)
host = get_address_host(address)
if address in self.workers:
raise ValueError("Worker already exists %s" % address)
if nbytes:
err = (
f"Worker {address!r} connected with {len(nbytes)} key(s) in memory! Worker reconnection is not supported. "
f"Keys: {list(nbytes)}"
)
logger.error(err)
await comm.write({"status": "error", "message": err, "time": time()})
return
if name in self.aliases:
logger.warning("Worker tried to connect with a duplicate name: %s", name)
msg = {
"status": "error",
"message": "name taken, %s" % name,
"time": time(),
}
await comm.write(msg)
return
self.log_event(address, {"action": "add-worker"})
self.log_event("all", {"action": "add-worker", "worker": address})
self.workers[address] = ws = WorkerState(
address=address,
status=Status.lookup[status], # type: ignore
pid=pid,
nthreads=nthreads,
memory_limit=memory_limit or 0,
name=name,
local_directory=local_directory,
services=services,
versions=versions,
nanny=nanny,
extra=extra,
server_id=server_id,
scheduler=self,
)
if ws.status == Status.running:
self.running.add(ws)
dh = self.host_info.get(host)
if dh is None:
self.host_info[host] = dh = {}
dh_addresses = dh.get("addresses")
if dh_addresses is None:
dh["addresses"] = dh_addresses = set()
dh["nthreads"] = 0
dh_addresses.add(address)
dh["nthreads"] += nthreads
self.total_nthreads += nthreads
self.aliases[name] = address
self.heartbeat_worker(
address=address,
resolve_address=resolve_address,
now=now,
resources=resources,
host_info=host_info,
metrics=metrics,
)
# Do not need to adjust self.total_occupancy as self.occupancy[ws] cannot
# exist before this.
self.check_idle_saturated(ws)
# for key in keys: # TODO
# self.mark_key_in_memory(key, [address])
self.stream_comms[address] = BatchedSend(interval="5ms", loop=self.loop)
for plugin in list(self.plugins.values()):
try:
result = plugin.add_worker(scheduler=self, worker=address)
if result is not None and inspect.isawaitable(result):
await result
except Exception as e:
logger.exception(e)
if ws.status == Status.running:
self.transitions(
self.bulk_schedule_unrunnable_after_adding_worker(ws), stimulus_id
)
self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id)
logger.info("Register worker %s", ws)
msg = {
"status": "OK",
"time": time(),
"heartbeat-interval": heartbeat_interval(len(self.workers)),
"worker-plugins": self.worker_plugins,
}
version_warning = version_module.error_message(
version_module.get_versions(),
{w: ws.versions for w, ws in self.workers.items()},
versions,
source_name=str(ws.server_id),
)
msg.update(version_warning)
await comm.write(msg)
# This will keep running until the worker is removed
await self.handle_worker(comm, address)
async def add_nanny(self) -> dict[str, Any]:
msg = {
"status": "OK",
"nanny-plugins": self.nanny_plugins,
}
return msg
def update_graph_hlg(
self,
client=None,
hlg=None,
keys=None,
dependencies=None,
restrictions=None,
priority=None,
loose_restrictions=None,
resources=None,
submitting_task=None,
retries=None,
user_priority=0,
actors=None,
fifo_timeout=0,
code=None,
):
unpacked_graph = HighLevelGraph.__dask_distributed_unpack__(hlg)
dsk = unpacked_graph["dsk"]
dependencies = unpacked_graph["deps"]
annotations = unpacked_graph["annotations"]
# Remove any self-dependencies (happens on test_publish_bag() and others)
for k, v in dependencies.items():
deps = set(v)
if k in deps:
deps.remove(k)
dependencies[k] = deps
if priority is None:
# Removing all non-local keys before calling order()
dsk_keys = set(dsk) # intersection() of sets is much faster than dict_keys
stripped_deps = {
k: v.intersection(dsk_keys)
for k, v in dependencies.items()
if k in dsk_keys
}
priority = dask.order.order(dsk, dependencies=stripped_deps)
return self.update_graph(
client,
dsk,
keys,
dependencies,
restrictions,
priority,
loose_restrictions,
resources,
submitting_task,
retries,
user_priority,
actors,
fifo_timeout,
annotations,
code=code,
stimulus_id=f"update-graph-{time()}",
)
def update_graph(
self,
client=None,
tasks=None,
keys=None,
dependencies=None,
restrictions=None,
priority=None,
loose_restrictions=None,
resources=None,
submitting_task=None,
retries=None,
user_priority=0,
actors=None,
fifo_timeout=0,
annotations=None,
code=None,
stimulus_id=None,
):
"""
Add new computations to the internal dask graph
This happens whenever the Client calls submit, map, get, or compute.
"""
stimulus_id = stimulus_id or f"update-graph-{time()}"
start = time()
fifo_timeout = parse_timedelta(fifo_timeout)
keys = set(keys)
if len(tasks) > 1:
self.log_event(
["all", client], {"action": "update_graph", "count": len(tasks)}
)
# Remove aliases
for k in list(tasks):
if tasks[k] is k:
del tasks[k]
dependencies = dependencies or {}
if self.total_occupancy > 1e-9 and self.computations:
# Still working on something. Assign new tasks to same computation
computation = self.computations[-1]
else:
computation = Computation()
self.computations.append(computation)
if code and code not in computation.code: # add new code blocks
computation.code.add(code)
n = 0
while len(tasks) != n: # walk through new tasks, cancel any bad deps
n = len(tasks)
for k, deps in list(dependencies.items()):
if any(
dep not in self.tasks and dep not in tasks for dep in deps
): # bad key
logger.info("User asked for computation on lost data, %s", k)
del tasks[k]
del dependencies[k]
if k in keys:
keys.remove(k)
self.report({"op": "cancelled-key", "key": k}, client=client)
self.client_releases_keys(
keys=[k], client=client, stimulus_id=stimulus_id
)
# Avoid computation that is already finished
already_in_memory = set() # tasks that are already done
for k, v in dependencies.items():
if v and k in self.tasks:
ts = self.tasks[k]
if ts.state in ("memory", "erred"):
already_in_memory.add(k)
if already_in_memory:
dependents = dask.core.reverse_dict(dependencies)
stack = list(already_in_memory)
done = set(already_in_memory)
while stack: # remove unnecessary dependencies
key = stack.pop()
try:
deps = dependencies[key]
except KeyError:
deps = self.tasks[key].dependencies
for dep in deps:
if dep in dependents:
child_deps = dependents[dep]
elif dep in self.tasks:
child_deps = self.tasks[dep].dependencies
else:
child_deps = set()
if all(d in done for d in child_deps):
if dep in self.tasks and dep not in done:
done.add(dep)
stack.append(dep)
for d in done:
tasks.pop(d, None)
dependencies.pop(d, None)
# Get or create task states
stack = list(keys)
touched_keys = set()
touched_tasks = []
while stack:
k = stack.pop()
if k in touched_keys:
continue
# XXX Have a method get_task_state(self, k) ?
ts = self.tasks.get(k)
if ts is None:
ts = self.new_task(k, tasks.get(k), "released", computation=computation)
elif not ts.run_spec:
ts.run_spec = tasks.get(k)
touched_keys.add(k)
touched_tasks.append(ts)
stack.extend(dependencies.get(k, ()))
self.client_desires_keys(keys=keys, client=client)
# Add dependencies
for key, deps in dependencies.items():
ts = self.tasks.get(key)
if ts is None or ts.dependencies:
continue
for dep in deps:
dts = self.tasks[dep]
ts.add_dependency(dts)
# Compute priorities
if isinstance(user_priority, Number):
user_priority = {k: user_priority for k in tasks}
annotations = annotations or {}
restrictions = restrictions or {}
loose_restrictions = loose_restrictions or []
resources = resources or {}
retries = retries or {}
# Override existing taxonomy with per task annotations
if annotations:
if "priority" in annotations:
user_priority.update(annotations["priority"])
if "workers" in annotations:
restrictions.update(annotations["workers"])
if "allow_other_workers" in annotations:
loose_restrictions.extend(
k for k, v in annotations["allow_other_workers"].items() if v
)
if "retries" in annotations:
retries.update(annotations["retries"])
if "resources" in annotations:
resources.update(annotations["resources"])
for a, kv in annotations.items():
for k, v in kv.items():
# Tasks might have been culled, in which case
# we have nothing to annotate.
ts = self.tasks.get(k)
if ts is not None:
ts.annotations[a] = v
# Add actors
if actors is True:
actors = list(keys)
for actor in actors or []:
ts = self.tasks[actor]
ts.actor = True
priority = priority or dask.order.order(
tasks
) # TODO: define order wrt old graph
if submitting_task: # sub-tasks get better priority than parent tasks
ts = self.tasks.get(submitting_task)
if ts is not None:
generation = ts.priority[0] - 0.01
else: # super-task already cleaned up
generation = self.generation
elif self._last_time + fifo_timeout < start:
self.generation += 1 # older graph generations take precedence
generation = self.generation
self._last_time = start
else:
generation = self.generation
for key in set(priority) & touched_keys:
ts = self.tasks[key]
if ts.priority is None:
ts.priority = (-(user_priority.get(key, 0)), generation, priority[key])
# Ensure all runnables have a priority
runnables = [ts for ts in touched_tasks if ts.run_spec]
for ts in runnables:
if ts.priority is None and ts.run_spec:
ts.priority = (self.generation, 0)
if restrictions:
# *restrictions* is a dict keying task ids to lists of
# restriction specifications (either worker names or addresses)
for k, v in restrictions.items():
if v is None:
continue
ts = self.tasks.get(k)
if ts is None:
continue
ts.host_restrictions = set()
ts.worker_restrictions = set()
# Make sure `v` is a collection and not a single worker name / address
if not isinstance(v, (list, tuple, set)):
v = [v]
for w in v:
try:
w = self.coerce_address(w)
except ValueError:
# Not a valid address, but perhaps it's a hostname
ts.host_restrictions.add(w)
else:
ts.worker_restrictions.add(w)
if loose_restrictions:
for k in loose_restrictions:
ts = self.tasks[k]
ts.loose_restrictions = True
if resources:
for k, v in resources.items():
if v is None:
continue
assert isinstance(v, dict)
ts = self.tasks.get(k)
if ts is None:
continue
ts.resource_restrictions = v
if retries:
for k, v in retries.items():
assert isinstance(v, int)
ts = self.tasks.get(k)
if ts is None:
continue
ts.retries = v
# Compute recommendations
recommendations: Recs = {}
for ts in sorted(runnables, key=operator.attrgetter("priority"), reverse=True):
if ts.state == "released" and ts.run_spec:
recommendations[ts.key] = "waiting"
for ts in touched_tasks:
for dts in ts.dependencies:
if dts.exception_blame:
ts.exception_blame = dts.exception_blame
recommendations[ts.key] = "erred"
break
for plugin in list(self.plugins.values()):
try:
plugin.update_graph(
self,
client=client,
tasks=tasks,
keys=keys,
restrictions=restrictions or {},
dependencies=dependencies,
priority=priority,
loose_restrictions=loose_restrictions,
resources=resources,
annotations=annotations,
)
except Exception as e:
logger.exception(e)
self.transitions(recommendations, stimulus_id)
for ts in touched_tasks:
if ts.state in ("memory", "erred"):
self.report_on_key(ts=ts, client=client)
end = time()
self.digest_metric("update-graph-duration", end - start)
# TODO: balance workers
def stimulus_queue_slots_maybe_opened(self, *, stimulus_id: str) -> None:
"""Respond to an event which may have opened spots on worker threadpools
Selects the appropriate number of tasks from the front of the queue according to
the total number of task slots available on workers (potentially 0), and
transitions them to ``processing``.
Notes
-----
Other transitions related to this stimulus should be fully processed beforehand,
so any tasks that became runnable are already in ``processing``. Otherwise,
overproduction can occur if queued tasks get scheduled before downstream tasks.
Must be called after `check_idle_saturated`; i.e. `idle_task_count` must be up
to date.
"""
if not self.queued:
return
slots_available = sum(
_task_slots_available(ws, self.WORKER_SATURATION)
for ws in self.idle_task_count
)
if slots_available == 0:
return
recommendations: Recs = {}
for qts in self.queued.peekn(slots_available):
if self.validate:
assert qts.state == "queued", qts.state
assert not qts.processing_on, (qts, qts.processing_on)
assert not qts.waiting_on, (qts, qts.processing_on)
assert qts.who_wants or qts.waiters, qts
recommendations[qts.key] = "processing"
self.transitions(recommendations, stimulus_id)
def stimulus_task_finished(self, key=None, worker=None, stimulus_id=None, **kwargs):
"""Mark that a task has finished execution on a particular worker"""
logger.debug("Stimulus task finished %s, %s", key, worker)
recommendations: Recs = {}
client_msgs: Msgs = {}
worker_msgs: Msgs = {}
ws: WorkerState = self.workers[worker]
ts: TaskState = self.tasks.get(key)
if ts is None or ts.state in ("released", "queued"):
logger.debug(
"Received already computed task, worker: %s, state: %s"
", key: %s, who_has: %s",
worker,
ts.state if ts else "forgotten",
key,
ts.who_has if ts else {},
)
worker_msgs[worker] = [
{
"op": "free-keys",
"keys": [key],
"stimulus_id": stimulus_id,
}
]
elif ts.state == "memory":
self.add_keys(worker=worker, keys=[key])
else:
ts.metadata.update(kwargs["metadata"])
r: tuple = self._transition(
key, "memory", stimulus_id, worker=worker, **kwargs
)
recommendations, client_msgs, worker_msgs = r
if ts.state == "memory":
assert ws in ts.who_has
return recommendations, client_msgs, worker_msgs
def stimulus_task_erred(
self,
key=None,
worker=None,
exception=None,
stimulus_id=None,
traceback=None,
**kwargs,
):
"""Mark that a task has erred on a particular worker"""
logger.debug("Stimulus task erred %s, %s", key, worker)
ts: TaskState = self.tasks.get(key)
if ts is None or ts.state != "processing":
return {}, {}, {}
if ts.retries > 0:
ts.retries -= 1
return self._transition(key, "waiting", stimulus_id)
else:
return self._transition(
key,
"erred",
stimulus_id,
cause=key,
exception=exception,
traceback=traceback,
worker=worker,
**kwargs,
)
def stimulus_retry(self, keys, client=None):
logger.info("Client %s requests to retry %d keys", client, len(keys))
if client:
self.log_event(client, {"action": "retry", "count": len(keys)})
stack = list(keys)
seen = set()
roots = []
ts: TaskState
dts: TaskState
while stack:
key = stack.pop()
seen.add(key)
ts = self.tasks[key]
erred_deps = [dts.key for dts in ts.dependencies if dts.state == "erred"]
if erred_deps:
stack.extend(erred_deps)
else:
roots.append(key)
recommendations: Recs = {key: "waiting" for key in roots}
self.transitions(recommendations, f"stimulus-retry-{time()}")
if self.validate:
for key in seen:
assert not self.tasks[key].exception_blame
return tuple(seen)
def close_worker(self, worker: str) -> None:
"""Ask a worker to shut itself down. Do not wait for it to take effect.
Note that there is no guarantee that the worker will actually accept the
command.
Note that :meth:`remove_worker` sends the same command internally if close=True.
See also
--------
retire_workers
remove_worker
"""
if worker not in self.workers:
return
logger.info("Closing worker %s", worker)
self.log_event(worker, {"action": "close-worker"})
self.worker_send(worker, {"op": "close", "reason": "scheduler-close-worker"})
@log_errors
async def remove_worker(
self, address: str, *, stimulus_id: str, safe: bool = False, close: bool = True
) -> Literal["OK", "already-removed"]:
"""Remove worker from cluster.
We do this when a worker reports that it plans to leave or when it appears to be
unresponsive. This may send its tasks back to a released state.
See also
--------
retire_workers
close_worker
"""
if self.status == Status.closed:
return "already-removed"
address = self.coerce_address(address)
if address not in self.workers:
return "already-removed"
host = get_address_host(address)
ws: WorkerState = self.workers[address]
event_msg = {
"action": "remove-worker",
"processing-tasks": {ts.key for ts in ws.processing},
}
self.log_event(address, event_msg.copy())
event_msg["worker"] = address
self.log_event("all", event_msg)
logger.info("Remove worker %s", ws)
if close:
with suppress(AttributeError, CommClosedError):
self.stream_comms[address].send(
{"op": "close", "reason": "scheduler-remove-worker"}
)
self.remove_resources(address)
dh: dict = self.host_info[host]
dh_addresses: set = dh["addresses"]
dh_addresses.remove(address)
dh["nthreads"] -= ws.nthreads
self.total_nthreads -= ws.nthreads
if not dh_addresses:
del self.host_info[host]
self.rpc.remove(address)
del self.stream_comms[address]
del self.aliases[ws.name]
self.idle.pop(ws.address, None)
self.idle_task_count.discard(ws)
self.saturated.discard(ws)
del self.workers[address]
ws.status = Status.closed
self.running.discard(ws)
recommendations: Recs = {}
ts: TaskState
for ts in list(ws.processing):
k = ts.key
recommendations[k] = "released"
if not safe:
ts.suspicious += 1
ts.prefix.suspicious += 1
if ts.suspicious > self.allowed_failures:
del recommendations[k]
e = pickle.dumps(
KilledWorker(
task=k,
last_worker=ws.clean(),
allowed_failures=self.allowed_failures,
),
)
r = self.transition(
k,
"erred",
exception=e,
cause=k,
stimulus_id=stimulus_id,
worker=address,
)
recommendations.update(r)
logger.info(
"Task %s marked as failed because %d workers died"
" while trying to run it",
ts.key,
self.allowed_failures,
)
for ts in list(ws.has_what):
self.remove_replica(ts, ws)
if not ts.who_has:
if ts.run_spec:
recommendations[ts.key] = "released"
else: # pure data
recommendations[ts.key] = "forgotten"
self.transitions(recommendations, stimulus_id=stimulus_id)
for plugin in list(self.plugins.values()):
try:
result = plugin.remove_worker(scheduler=self, worker=address)
if inspect.isawaitable(result):
await result
except Exception as e:
logger.exception(e)
if not self.workers:
logger.info("Lost all workers")
for w in self.workers:
self.bandwidth_workers.pop((address, w), None)
self.bandwidth_workers.pop((w, address), None)
async def remove_worker_from_events():
# If the worker isn't registered anymore after the delay, remove from events
if address not in self.workers and address in self.events:
del self.events[address]
cleanup_delay = parse_timedelta(
dask.config.get("distributed.scheduler.events-cleanup-delay")
)
self._ongoing_background_tasks.call_later(
cleanup_delay, remove_worker_from_events
)
logger.debug("Removed worker %s", ws)
return "OK"
async def stimulus_cancel(self, keys, client, force=False):
"""Stop execution on a list of keys"""
logger.info("Client %s requests to cancel %d keys", client, len(keys))
if client:
self.log_event(
client, {"action": "cancel", "count": len(keys), "force": force}
)
await asyncio.gather(
*[self._cancel_key(key, client, force=force) for key in keys]
)
async def _cancel_key(self, key, client, force=False):
"""Cancel a particular key and all dependents"""
# TODO: this should be converted to use the transition mechanism
ts: TaskState | None = self.tasks.get(key)
dts: TaskState
try:
cs: ClientState = self.clients[client]
except KeyError:
return
# no key yet, lets try again in a moment
start = time()
while ts is None or not ts.who_wants:
await asyncio.sleep(0.1)
ts = self.tasks.get(key)
if time() - start >= 1:
return
if force or ts.who_wants == {cs}: # no one else wants this key
await asyncio.gather(
*[
self._cancel_key(dts.key, client, force=force)
for dts in ts.dependents
]
)
logger.info("Scheduler cancels key %s. Force=%s", key, force)
self.report({"op": "cancelled-key", "key": key})
clients = list(ts.who_wants) if force else [cs]
for cs in clients:
self.client_releases_keys(
keys=[key], client=cs.client_key, stimulus_id=f"cancel-key-{time()}"
)
def client_desires_keys(self, keys=None, client=None):
cs: ClientState = self.clients.get(client)
if cs is None:
# For publish, queues etc.
self.clients[client] = cs = ClientState(client)
for k in keys:
ts = self.tasks.get(k)
if ts is None:
# For publish, queues etc.
ts = self.new_task(k, None, "released")
ts.who_wants.add(cs)
cs.wants_what.add(ts)
if ts.state in ("memory", "erred"):
self.report_on_key(ts=ts, client=client)
def client_releases_keys(self, keys=None, client=None, stimulus_id=None):
"""Remove keys from client desired list"""
stimulus_id = stimulus_id or f"client-releases-keys-{time()}"
if not isinstance(keys, list):
keys = list(keys)
cs = self.clients[client]
recommendations: Recs = {}
self._client_releases_keys(keys=keys, cs=cs, recommendations=recommendations)
self.transitions(recommendations, stimulus_id)
self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id)
def client_heartbeat(self, client=None):
"""Handle heartbeats from Client"""
cs: ClientState = self.clients[client]
cs.last_seen = time()
###################
# Task Validation #
###################
def validate_released(self, key):
ts: TaskState = self.tasks[key]
assert ts.state == "released"
assert not ts.waiters
assert not ts.waiting_on
assert not ts.who_has
assert not ts.processing_on
assert not any([ts in dts.waiters for dts in ts.dependencies])
assert ts not in self.unrunnable
assert ts not in self.queued
def validate_waiting(self, key):
ts: TaskState = self.tasks[key]
assert ts.waiting_on
assert not ts.who_has
assert not ts.processing_on
assert ts not in self.unrunnable
assert ts not in self.queued
for dts in ts.dependencies:
# We are waiting on a dependency iff it's not stored
assert bool(dts.who_has) != (dts in ts.waiting_on)
assert ts in dts.waiters # XXX even if dts._who_has?
def validate_queued(self, key):
ts: TaskState = self.tasks[key]
dts: TaskState
assert ts in self.queued
assert not ts.waiting_on
assert not ts.who_has
assert not ts.processing_on
assert not (
ts.worker_restrictions or ts.host_restrictions or ts.resource_restrictions
)
for dts in ts.dependencies:
assert dts.who_has
assert ts in dts.waiters
def validate_processing(self, key):
ts: TaskState = self.tasks[key]
dts: TaskState
assert not ts.waiting_on
ws = ts.processing_on
assert ws
assert ts in ws.processing
assert not ts.who_has
assert ts not in self.queued
for dts in ts.dependencies:
assert dts.who_has
assert ts in dts.waiters
def validate_memory(self, key):
ts: TaskState = self.tasks[key]
dts: TaskState
assert ts.who_has
assert bool(ts in self.replicated_tasks) == (len(ts.who_has) > 1)
assert not ts.processing_on
assert not ts.waiting_on
assert ts not in self.unrunnable
assert ts not in self.queued
for dts in ts.dependents:
assert (dts in ts.waiters) == (
dts.state in ("waiting", "queued", "processing", "no-worker")
)
assert ts not in dts.waiting_on
def validate_no_worker(self, key):
ts: TaskState = self.tasks[key]
assert ts in self.unrunnable
assert not ts.waiting_on
assert ts in self.unrunnable
assert not ts.processing_on
assert not ts.who_has
assert ts not in self.queued
for dts in ts.dependencies:
assert dts.who_has
def validate_erred(self, key):
ts: TaskState = self.tasks[key]
assert ts.exception_blame
assert not ts.who_has
assert ts not in self.queued
def validate_key(self, key, ts: TaskState | None = None):
try:
if ts is None:
ts = self.tasks.get(key)
if ts is None:
logger.debug("Key lost: %s", key)
else:
ts.validate()
try:
func = getattr(self, "validate_" + ts.state.replace("-", "_"))
except AttributeError:
logger.error(
"self.validate_%s not found", ts.state.replace("-", "_")
)
else:
func(key)
except Exception as e:
logger.exception(e)
if LOG_PDB:
import pdb
pdb.set_trace()
raise
def validate_state(self, allow_overlap: bool = False) -> None:
validate_state(self.tasks, self.workers, self.clients)
if not (set(self.workers) == set(self.stream_comms)):
raise ValueError("Workers not the same in all collections")
assert self.running.issuperset(self.idle.values()), (
self.running,
list(self.idle.values()),
)
task_prefix_counts: defaultdict[str, int] = defaultdict(int)
for w, ws in self.workers.items():
assert isinstance(w, str), (type(w), w)
assert isinstance(ws, WorkerState), (type(ws), ws)
assert ws.address == w
if ws.status != Status.running:
assert ws.address not in self.idle
assert ws.long_running.issubset(ws.processing)
if not ws.processing:
assert not ws.occupancy
if ws.status == Status.running:
assert ws.address in self.idle
assert not ws.needs_what.keys() & ws.has_what
actual_needs_what: defaultdict[TaskState, int] = defaultdict(int)
for ts in ws.processing:
for tss in ts.dependencies:
if tss not in ws.has_what:
actual_needs_what[tss] += 1
assert actual_needs_what == ws.needs_what
assert (ws.status == Status.running) == (ws in self.running)
for name, count in ws.task_prefix_count.items():
task_prefix_counts[name] += count
assert task_prefix_counts.keys() == self._task_prefix_count_global.keys()
for name, global_count in self._task_prefix_count_global.items():
assert (
task_prefix_counts[name] == global_count
), f"{name}: {task_prefix_counts[name]} (wss), {global_count} (global)"
for ws in self.running:
assert ws.status == Status.running
assert ws.address in self.workers
for k, ts in self.tasks.items():
assert isinstance(ts, TaskState), (type(ts), ts)
assert ts.key == k
assert bool(ts in self.replicated_tasks) == (len(ts.who_has) > 1)
self.validate_key(k, ts)
for ts in self.replicated_tasks:
assert ts.state == "memory"
assert ts.key in self.tasks
for c, cs in self.clients.items():
# client=None is often used in tests...
assert c is None or type(c) == str, (type(c), c)
assert type(cs) == ClientState, (type(cs), cs)
assert cs.client_key == c
a = {w: ws.nbytes for w, ws in self.workers.items()}
b = {
w: sum(ts.get_nbytes() for ts in ws.has_what)
for w, ws in self.workers.items()
}
assert a == b, (a, b)
if self.transition_counter_max:
assert self.transition_counter < self.transition_counter_max
###################
# Manage Messages #
###################
def report(self, msg: dict, ts: TaskState | None = None, client: str | None = None):
"""
Publish updates to all listening Queues and Comms
If the message contains a key then we only send the message to those
comms that care about the key.
"""
if ts is None:
msg_key = msg.get("key")
if msg_key is not None:
tasks: dict = self.tasks
ts = tasks.get(msg_key)
client_comms: dict = self.client_comms
if ts is None:
# Notify all clients
client_keys = list(client_comms)
elif client is None:
# Notify clients interested in key
client_keys = [cs.client_key for cs in ts.who_wants]
else:
# Notify clients interested in key (including `client`)
client_keys = [
cs.client_key for cs in ts.who_wants if cs.client_key != client
]
client_keys.append(client)
k: str
for k in client_keys:
c = client_comms.get(k)
if c is None:
continue
try:
c.send(msg)
# logger.debug("Scheduler sends message to client %s", msg)
except CommClosedError:
if self.status == Status.running:
logger.critical(
"Closed comm %r while trying to write %s", c, msg, exc_info=True
)
async def add_client(
self, comm: Comm, client: str, versions: dict[str, Any]
) -> None:
"""Add client to network
We listen to all future messages from this Comm.
"""
assert client is not None
comm.name = "Scheduler->Client"
logger.info("Receive client connection: %s", client)
self.log_event(["all", client], {"action": "add-client", "client": client})
self.clients[client] = ClientState(client, versions=versions)
for plugin in list(self.plugins.values()):
try:
plugin.add_client(scheduler=self, client=client)
except Exception as e:
logger.exception(e)
try:
bcomm = BatchedSend(interval="2ms", loop=self.loop)
bcomm.start(comm)
self.client_comms[client] = bcomm
msg = {"op": "stream-start"}
version_warning = version_module.error_message(
version_module.get_versions(),
{w: ws.versions for w, ws in self.workers.items()},
versions,
)
msg.update(version_warning)
bcomm.send(msg)
try:
await self.handle_stream(comm=comm, extra={"client": client})
finally:
self.remove_client(client=client, stimulus_id=f"remove-client-{time()}")
logger.debug("Finished handling client %s", client)
finally:
if not comm.closed():
self.client_comms[client].send({"op": "stream-closed"})
try:
if not sys.is_finalizing():
await self.client_comms[client].close()
del self.client_comms[client]
if self.status == Status.running:
logger.info("Close client connection: %s", client)
except TypeError: # comm becomes None during GC
pass
def remove_client(self, client: str, stimulus_id: str | None = None) -> None:
"""Remove client from network"""
stimulus_id = stimulus_id or f"remove-client-{time()}"
if self.status == Status.running:
logger.info("Remove client %s", client)
self.log_event(["all", client], {"action": "remove-client", "client": client})
try:
cs: ClientState = self.clients[client]
except KeyError:
# XXX is this a legitimate condition?
pass
else:
self.client_releases_keys(
keys=[ts.key for ts in cs.wants_what],
client=cs.client_key,
stimulus_id=stimulus_id,
)
del self.clients[client]
for plugin in list(self.plugins.values()):
try:
plugin.remove_client(scheduler=self, client=client)
except Exception as e:
logger.exception(e)
async def remove_client_from_events():
# If the client isn't registered anymore after the delay, remove from events
if client not in self.clients and client in self.events:
del self.events[client]
cleanup_delay = parse_timedelta(
dask.config.get("distributed.scheduler.events-cleanup-delay")
)
if not self._ongoing_background_tasks.closed:
self._ongoing_background_tasks.call_later(
cleanup_delay, remove_client_from_events
)
def send_task_to_worker(self, worker, ts: TaskState, duration: float = -1):
"""Send a single computational task to a worker"""
try:
msg: dict = self._task_to_msg(ts, duration)
self.worker_send(worker, msg)
except Exception as e:
logger.exception(e)
if LOG_PDB:
import pdb
pdb.set_trace()
raise
def handle_uncaught_error(self, **msg):
logger.exception(clean_exception(**msg)[1])
def handle_task_finished(
self, key: str, worker: str, stimulus_id: str, **msg
) -> None:
if worker not in self.workers:
return
validate_key(key)
r: tuple = self.stimulus_task_finished(
key=key, worker=worker, stimulus_id=stimulus_id, **msg
)
recommendations, client_msgs, worker_msgs = r
self._transitions(recommendations, client_msgs, worker_msgs, stimulus_id)
self.send_all(client_msgs, worker_msgs)
self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id)
def handle_task_erred(self, key: str, stimulus_id: str, **msg) -> None:
r: tuple = self.stimulus_task_erred(key=key, stimulus_id=stimulus_id, **msg)
recommendations, client_msgs, worker_msgs = r
self._transitions(recommendations, client_msgs, worker_msgs, stimulus_id)
self.send_all(client_msgs, worker_msgs)
self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id)
def release_worker_data(self, key: str, worker: str, stimulus_id: str) -> None:
ts = self.tasks.get(key)
ws = self.workers.get(worker)
if not ts or not ws or ws not in ts.who_has:
return
self.remove_replica(ts, ws)
if not ts.who_has:
self.transitions({key: "released"}, stimulus_id)
def handle_long_running(
self, key: str, worker: str, compute_duration: float | None, stimulus_id: str
) -> None:
"""A task has seceded from the thread pool
We stop the task from being stolen in the future, and change task
duration accounting as if the task has stopped.
"""
if key not in self.tasks:
logger.debug("Skipping long_running since key %s was already released", key)
return
ts = self.tasks[key]
steal = self.extensions.get("stealing")
if steal is not None:
steal.remove_key_from_stealable(ts)
ws = ts.processing_on
if ws is None:
logger.debug("Received long-running signal from duplicate task. Ignoring.")
return
if compute_duration is not None:
old_duration = ts.prefix.duration_average
if old_duration < 0:
ts.prefix.duration_average = compute_duration
else:
ts.prefix.duration_average = (old_duration + compute_duration) / 2
ws.add_to_long_running(ts)
self.check_idle_saturated(ws)
self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id)
def handle_worker_status_change(
self, status: str | Status, worker: str | WorkerState, stimulus_id: str
) -> None:
ws = self.workers.get(worker) if isinstance(worker, str) else worker
if not ws:
return
prev_status = ws.status
ws.status = Status[status] if isinstance(status, str) else status
if ws.status == prev_status:
return
self.log_event(
ws.address,
{
"action": "worker-status-change",
"prev-status": prev_status.name,
"status": ws.status.name,
},
)
logger.debug(f"Worker status {prev_status.name} -> {status} - {ws}")
if ws.status == Status.running:
self.running.add(ws)
self.check_idle_saturated(ws)
self.transitions(
self.bulk_schedule_unrunnable_after_adding_worker(ws), stimulus_id
)
self.stimulus_queue_slots_maybe_opened(stimulus_id=stimulus_id)
else:
self.running.discard(ws)
self.idle.pop(ws.address, None)
self.idle_task_count.discard(ws)
async def handle_request_refresh_who_has(
self, keys: Iterable[str], worker: str, stimulus_id: str
) -> None:
"""Asynchronous request (through bulk comms) from a Worker to refresh the
who_has for some keys. Not to be confused with scheduler.who_has, which is a
synchronous RPC request from a Client.
"""
who_has = {}
free_keys = []
for key in keys:
if key in self.tasks:
who_has[key] = [ws.address for ws in self.tasks[key].who_has]
else:
free_keys.append(key)
if who_has:
self.stream_comms[worker].send(
{
"op": "refresh-who-has",
"who_has": who_has,
"stimulus_id": stimulus_id,
}
)
if free_keys:
self.stream_comms[worker].send(
{
"op": "free-keys",
"keys": free_keys,
"stimulus_id": stimulus_id,
}
)
async def handle_worker(self, comm: Comm, worker: str) -> None:
"""
Listen to responses from a single worker
This is the main loop for scheduler-worker interaction
See Also
--------
Scheduler.handle_client: Equivalent coroutine for clients
"""
comm.name = "Scheduler connection to worker"
worker_comm = self.stream_comms[worker]
worker_comm.start(comm)
logger.info("Starting worker compute stream, %s", worker)
try:
await self.handle_stream(comm=comm, extra={"worker": worker})
finally:
if worker in self.stream_comms:
worker_comm.abort()
await self.remove_worker(
worker, stimulus_id=f"handle-worker-cleanup-{time()}"
)
def add_plugin(
self,
plugin: SchedulerPlugin,
*,
idempotent: bool = False,
name: str | None = None,
**kwargs,
):
"""Add external plugin to scheduler.
See https://distributed.readthedocs.io/en/latest/plugins.html
Parameters
----------
plugin : SchedulerPlugin
SchedulerPlugin instance to add
idempotent : bool
If true, the plugin is assumed to already exist and no
action is taken.
name : str
A name for the plugin, if None, the name attribute is
checked on the Plugin instance and generated if not
discovered.
"""
if name is None:
name = _get_plugin_name(plugin)
if name in self.plugins:
if idempotent:
return
warnings.warn(
f"Scheduler already contains a plugin with name {name}; overwriting.",
category=UserWarning,
)
self.plugins[name] = plugin
def remove_plugin(
self,
name: str | None = None,
plugin: SchedulerPlugin | None = None,
) -> None:
"""Remove external plugin from scheduler
Parameters
----------
name : str
Name of the plugin to remove
"""
assert name is not None
try:
del self.plugins[name]
except KeyError:
raise ValueError(
f"Could not find plugin {name!r} among the current scheduler plugins"
)
async def register_scheduler_plugin(self, plugin, name=None, idempotent=None):
"""Register a plugin on the scheduler."""
if not dask.config.get("distributed.scheduler.pickle"):
raise ValueError(
"Cannot register a scheduler plugin as the scheduler "
"has been explicitly disallowed from deserializing "
"arbitrary bytestrings using pickle via the "
"'distributed.scheduler.pickle' configuration setting."
)
if not isinstance(plugin, SchedulerPlugin):
plugin = loads(plugin)
if name is None:
name = _get_plugin_name(plugin)
if name in self.plugins and idempotent:
return
if hasattr(plugin, "start"):
result = plugin.start(self)
if inspect.isawaitable(result):
await result
self.add_plugin(plugin, name=name, idempotent=idempotent)
def worker_send(self, worker: str, msg: dict[str, Any]) -> None:
"""Send message to worker
This also handles connection failures by adding a callback to remove
the worker on the next cycle.
"""
stream_comms: dict = self.stream_comms
try:
stream_comms[worker].send(msg)
except (CommClosedError, AttributeError):
self._ongoing_background_tasks.call_soon(
self.remove_worker,
address=worker,
stimulus_id=f"worker-send-comm-fail-{time()}",
)
def client_send(self, client, msg):
"""Send message to client"""
client_comms: dict = self.client_comms
c = client_comms.get(client)
if c is None:
return
try:
c.send(msg)
except CommClosedError:
if self.status == Status.running:
logger.critical(
"Closed comm %r while trying to write %s", c, msg, exc_info=True
)
def send_all(self, client_msgs: Msgs, worker_msgs: Msgs) -> None:
"""Send messages to client and workers"""
for client, msgs in client_msgs.items():
c = self.client_comms.get(client)
if c is None:
continue
try:
c.send(*msgs)
except CommClosedError:
if self.status == Status.running:
logger.critical(
"Closed comm %r while trying to write %s",
c,
msgs,
exc_info=True,
)
for worker, msgs in worker_msgs.items():
try:
w = self.stream_comms[worker]
w.send(*msgs)
except KeyError:
# worker already gone
pass
except (CommClosedError, AttributeError):
self._ongoing_background_tasks.call_soon(
self.remove_worker,
address=worker,
stimulus_id=f"send-all-comm-fail-{time()}",
)
############################
# Less common interactions #
############################
async def scatter(
self,
comm=None,
data=None,
workers=None,
client=None,
broadcast=False,
timeout=2,
):
"""Send data out to workers
See also
--------
Scheduler.broadcast:
"""
start = time()
while True:
if workers is None:
wss = self.running
else:
workers = [self.coerce_address(w) for w in workers]
wss = {self.workers[w] for w in workers}
wss = {ws for ws in wss if ws.status == Status.running}
if wss:
break
if time() > start + timeout:
raise TimeoutError("No valid workers found")
await asyncio.sleep(0.1)
nthreads = {ws.address: ws.nthreads for ws in wss}
assert isinstance(data, dict)
keys, who_has, nbytes = await scatter_to_workers(
nthreads, data, rpc=self.rpc, report=False
)
self.update_data(who_has=who_has, nbytes=nbytes, client=client)
if broadcast:
n = len(nthreads) if broadcast is True else broadcast
await self.replicate(keys=keys, workers=workers, n=n)
self.log_event(
[client, "all"], {"action": "scatter", "client": client, "count": len(data)}
)
return keys
async def gather(self, keys, serializers=None):
"""Collect data from workers to the scheduler"""
stimulus_id = f"gather-{time()}"
keys = list(keys)
who_has = {}
for key in keys:
ts: TaskState = self.tasks.get(key)
if ts is not None:
who_has[key] = [ws.address for ws in ts.who_has]
else:
who_has[key] = []
data, missing_keys, missing_workers = await gather_from_workers(
who_has, rpc=self.rpc, close=False, serializers=serializers
)
if not missing_keys:
result = {"status": "OK", "data": data}
else:
missing_states = [
(self.tasks[key].state if key in self.tasks else None)
for key in missing_keys
]
logger.exception(
"Couldn't gather keys %s state: %s workers: %s",
missing_keys,
missing_states,
missing_workers,
)
result = {"status": "error", "keys": missing_keys}
with log_errors():
# Remove suspicious workers from the scheduler and shut them down.
await asyncio.gather(
*(
self.remove_worker(
address=worker, close=True, stimulus_id=stimulus_id
)
for worker in missing_workers
)
)
for key, workers in missing_keys.items():
logger.exception(
"Shut down workers that don't have promised key: %s, %s",
str(workers),
str(key),
)
self.log_event("all", {"action": "gather", "count": len(keys)})
return result
@log_errors
async def restart(self, client=None, timeout=30, wait_for_workers=True):
"""
Restart all workers. Reset local state. Optionally wait for workers to return.
Workers without nannies are shut down, hoping an external deployment system
will restart them. Therefore, if not using nannies and your deployment system
does not automatically restart workers, ``restart`` will just shut down all
workers, then time out!
After ``restart``, all connected workers are new, regardless of whether ``TimeoutError``
was raised. Any workers that failed to shut down in time are removed, and
may or may not shut down on their own in the future.
Parameters
----------
timeout:
How long to wait for workers to shut down and come back, if ``wait_for_workers``
is True, otherwise just how long to wait for workers to shut down.
Raises ``asyncio.TimeoutError`` if this is exceeded.
wait_for_workers:
Whether to wait for all workers to reconnect, or just for them to shut down
(default True). Use ``restart(wait_for_workers=False)`` combined with
:meth:`Client.wait_for_workers` for granular control over how many workers to
wait for.
See also
--------
Client.restart
Client.restart_workers
"""
stimulus_id = f"restart-{time()}"
logger.info("Restarting workers and releasing all keys.")
for cs in self.clients.values():
self.client_releases_keys(
keys=[ts.key for ts in cs.wants_what],
client=cs.client_key,
stimulus_id=stimulus_id,
)
self._clear_task_state()
assert not self.tasks
self.report({"op": "restart"})
for plugin in list(self.plugins.values()):
try:
plugin.restart(self)
except Exception as e:
logger.exception(e)
n_workers = len(self.workers)
nanny_workers = {
addr: ws.nanny for addr, ws in self.workers.items() if ws.nanny
}
# Close non-Nanny workers. We have no way to restart them, so we just let them go,
# and assume a deployment system is going to restart them for us.
await asyncio.gather(
*(
self.remove_worker(address=addr, stimulus_id=stimulus_id)
for addr in self.workers
if addr not in nanny_workers
)
)
logger.debug("Send kill signal to nannies: %s", nanny_workers)
async with contextlib.AsyncExitStack() as stack:
nannies = await asyncio.gather(
*(
stack.enter_async_context(
rpc(nanny_address, connection_args=self.connection_args)
)
for nanny_address in nanny_workers.values()
)
)
start = monotonic()
resps = await asyncio.gather(
*(
asyncio.wait_for(
# FIXME does not raise if the process fails to shut down,
# see https://github.com/dask/distributed/pull/6427/files#r894917424
# NOTE: Nanny will automatically restart worker process when it's killed
nanny.kill(
reason="scheduler-restart",
timeout=timeout,
),
timeout,
)
for nanny in nannies
),
return_exceptions=True,
)
# NOTE: the `WorkerState` entries for these workers will be removed
# naturally when they disconnect from the scheduler.
# Remove any workers that failed to shut down, so we can guarantee
# that after `restart`, there are no old workers around.
bad_nannies = [
addr for addr, resp in zip(nanny_workers, resps) if resp is not None
]
if bad_nannies:
await asyncio.gather(
*(
self.remove_worker(addr, stimulus_id=stimulus_id)
for addr in bad_nannies
)
)
raise TimeoutError(
f"{len(bad_nannies)}/{len(nannies)} nanny worker(s) did not shut down within {timeout}s"
)
self.log_event([client, "all"], {"action": "restart", "client": client})
if wait_for_workers:
while len(self.workers) < n_workers:
# NOTE: if new (unrelated) workers join while we're waiting, we may return before
# our shut-down workers have come back up. That's fine; workers are interchangeable.
if monotonic() < start + timeout:
await asyncio.sleep(0.2)
else:
msg = (
f"Waited for {n_workers} worker(s) to reconnect after restarting, "
f"but after {timeout}s, only {len(self.workers)} have returned. "
"Consider a longer timeout, or `wait_for_workers=False`."
)
if (n_nanny := len(nanny_workers)) < n_workers:
msg += (
f" The {n_workers - n_nanny} worker(s) not using Nannies were just shut "
"down instead of restarted (restart is only possible with Nannies). If "
"your deployment system does not automatically re-launch terminated "
"processes, then those workers will never come back, and `Client.restart` "
"will always time out. Do not use `Client.restart` in that case."
)
raise TimeoutError(msg) from None
logger.info("Restarting finished.")
async def broadcast(
self,
comm=None,
*,
msg: dict,
workers: "list[str] | None" = None,
hosts: "list[str] | None" = None,
nanny: bool = False,
serializers=None,
on_error: "Literal['raise', 'return', 'return_pickle', 'ignore']" = "raise",
) -> dict: # dict[str, Any]
"""Broadcast message to workers, return all results"""
if workers is None:
if hosts is None:
workers = list(self.workers)
else:
workers = []
if hosts is not None:
for host in hosts:
dh: dict = self.host_info.get(host) # type: ignore
if dh is not None:
workers.extend(dh["addresses"])
# TODO replace with worker_list
if nanny:
addresses = [self.workers[w].nanny for w in workers]
else:
addresses = workers
ERROR = object()
async def send_message(addr):
try:
comm = await self.rpc.connect(addr)
comm.name = "Scheduler Broadcast"
try:
resp = await send_recv(
comm, close=True, serializers=serializers, **msg
)
finally:
self.rpc.reuse(addr, comm)
return resp
except Exception as e:
logger.error(f"broadcast to {addr} failed: {e.__class__.__name__}: {e}")
if on_error == "raise":
raise
elif on_error == "return":
return e
elif on_error == "return_pickle":
return dumps(e)
elif on_error == "ignore":
return ERROR
else:
raise ValueError(
"on_error must be 'raise', 'return', 'return_pickle', "
f"or 'ignore'; got {on_error!r}"
)
results = await All(
[send_message(address) for address in addresses if address is not None]
)
return {k: v for k, v in zip(workers, results) if v is not ERROR}
async def proxy(self, comm=None, msg=None, worker=None, serializers=None):
"""Proxy a communication through the scheduler to some other worker"""
d = await self.broadcast(
comm=comm, msg=msg, workers=[worker], serializers=serializers
)
return d[worker]
async def gather_on_worker(
self, worker_address: str, who_has: "dict[str, list[str]]"
) -> set:
"""Peer-to-peer copy of keys from multiple workers to a single worker
Parameters
----------
worker_address: str
Recipient worker address to copy keys to
who_has: dict[Hashable, list[str]]
{key: [sender address, sender address, ...], key: ...}
Returns
-------
returns:
set of keys that failed to be copied
"""
try:
result = await retry_operation(
self.rpc(addr=worker_address).gather, who_has=who_has
)
except OSError as e:
# This can happen e.g. if the worker is going through controlled shutdown;
# it doesn't necessarily mean that it went unexpectedly missing
logger.warning(
f"Communication with worker {worker_address} failed during "
f"replication: {e.__class__.__name__}: {e}"
)
return set(who_has)
ws = self.workers.get(worker_address)
if not ws:
logger.warning(f"Worker {worker_address} lost during replication")
return set(who_has)
elif result["status"] == "OK":
keys_failed = set()
keys_ok: Set = who_has.keys()
elif result["status"] == "partial-fail":
keys_failed = set(result["keys"])
keys_ok = who_has.keys() - keys_failed
logger.warning(
f"Worker {worker_address} failed to acquire keys: {result['keys']}"
)
else: # pragma: nocover
raise ValueError(f"Unexpected message from {worker_address}: {result}")
for key in keys_ok:
ts: TaskState = self.tasks.get(key) # type: ignore
if ts is None or ts.state != "memory":
logger.warning(f"Key lost during replication: {key}")
continue
if ws not in ts.who_has:
self.add_replica(ts, ws)
return keys_failed
async def delete_worker_data(
self, worker_address: str, keys: "Collection[str]", stimulus_id: str
) -> None:
"""Delete data from a worker and update the corresponding worker/task states
Parameters
----------
worker_address: str
Worker address to delete keys from
keys: list[str]
List of keys to delete on the specified worker
"""
try:
await retry_operation(
self.rpc(addr=worker_address).free_keys,
keys=list(keys),
stimulus_id=f"delete-data-{time()}",
)
except OSError as e:
# This can happen e.g. if the worker is going through controlled shutdown;
# it doesn't necessarily mean that it went unexpectedly missing
logger.warning(
f"Communication with worker {worker_address} failed during "
f"replication: {e.__class__.__name__}: {e}"
)
return
ws = self.workers.get(worker_address)
if not ws:
return
for key in keys:
ts: TaskState = self.tasks.get(key) # type: ignore
if ts is not None and ws in ts.who_has:
assert ts.state == "memory"
self.remove_replica(ts, ws)
if not ts.who_has:
# Last copy deleted
self.transitions({key: "released"}, stimulus_id)
self.log_event(ws.address, {"action": "remove-worker-data", "keys": keys})
@log_errors
async def rebalance(
self,
comm=None,
keys: Iterable[str] | None = None,
workers: Iterable[str] | None = None,
stimulus_id: str | None = None,
) -> dict:
"""Rebalance keys so that each worker ends up with roughly the same process
memory (managed+unmanaged).
.. warning::
This operation is generally not well tested against normal operation of the
scheduler. It is not recommended to use it while waiting on computations.
**Algorithm**
#. Find the mean occupancy of the cluster, defined as data managed by dask +
unmanaged process memory that has been there for at least 30 seconds
(``distributed.worker.memory.recent-to-old-time``).
This lets us ignore temporary spikes caused by task heap usage.
Alternatively, you may change how memory is measured both for the individual
workers as well as to calculate the mean through
``distributed.worker.memory.rebalance.measure``. Namely, this can be useful
to disregard inaccurate OS memory measurements.
#. Discard workers whose occupancy is within 5% of the mean cluster occupancy
(``distributed.worker.memory.rebalance.sender-recipient-gap`` / 2).
This helps avoid data from bouncing around the cluster repeatedly.
#. Workers above the mean are senders; those below are recipients.
#. Discard senders whose absolute occupancy is below 30%
(``distributed.worker.memory.rebalance.sender-min``). In other words, no data
is moved regardless of imbalancing as long as all workers are below 30%.
#. Discard recipients whose absolute occupancy is above 60%
(``distributed.worker.memory.rebalance.recipient-max``).
Note that this threshold by default is the same as
``distributed.worker.memory.target`` to prevent workers from accepting data
and immediately spilling it out to disk.
#. Iteratively pick the sender and recipient that are farthest from the mean and
move the *least recently inserted* key between the two, until either all
senders or all recipients fall within 5% of the mean.
A recipient will be skipped if it already has a copy of the data. In other
words, this method does not degrade replication.
A key will be skipped if there are no recipients available with enough memory
to accept the key and that don't already hold a copy.
The least recently insertd (LRI) policy is a greedy choice with the advantage of
being O(1), trivial to implement (it relies on python dict insertion-sorting)
and hopefully good enough in most cases. Discarded alternative policies were:
- Largest first. O(n*log(n)) save for non-trivial additional data structures and
risks causing the largest chunks of data to repeatedly move around the
cluster like pinballs.
- Least recently used (LRU). This information is currently available on the
workers only and not trivial to replicate on the scheduler; transmitting it
over the network would be very expensive. Also, note that dask will go out of
its way to minimise the amount of time intermediate keys are held in memory,
so in such a case LRI is a close approximation of LRU.
Parameters
----------
keys: optional
allowlist of dask keys that should be considered for moving. All other keys
will be ignored. Note that this offers no guarantee that a key will actually
be moved (e.g. because it is unnecessary or because there are no viable
recipient workers for it).
workers: optional
allowlist of workers addresses to be considered as senders or recipients.
All other workers will be ignored. The mean cluster occupancy will be
calculated only using the allowed workers.
"""
stimulus_id = stimulus_id or f"rebalance-{time()}"
wss: Collection[WorkerState]
if workers is not None:
wss = [self.workers[w] for w in workers]
else:
wss = self.workers.values()
if not wss:
return {"status": "OK"}
if keys is not None:
if not isinstance(keys, Set):
keys = set(keys) # unless already a set-like
if not keys:
return {"status": "OK"}
missing_data = [
k for k in keys if k not in self.tasks or not self.tasks[k].who_has
]
if missing_data:
return {"status": "partial-fail", "keys": missing_data}
msgs = self._rebalance_find_msgs(keys, wss)
if not msgs:
return {"status": "OK"}
async with self._lock:
result = await self._rebalance_move_data(msgs, stimulus_id)
if result["status"] == "partial-fail" and keys is None:
# Only return failed keys if the client explicitly asked for them
result = {"status": "OK"}
return result
def _rebalance_find_msgs(
self,
keys: Set[Hashable] | None,
workers: Iterable[WorkerState],
) -> list[tuple[WorkerState, WorkerState, TaskState]]:
"""Identify workers that need to lose keys and those that can receive them,
together with how many bytes each needs to lose/receive. Then, pair a sender
worker with a recipient worker for each key, until the cluster is rebalanced.
This method only defines the work to be performed; it does not start any network
transfers itself.
The big-O complexity is O(wt + ke*log(we)), where
- wt is the total number of workers on the cluster (or the number of allowed
workers, if explicitly stated by the user)
- we is the number of workers that are eligible to be senders or recipients
- kt is the total number of keys on the cluster (or on the allowed workers)
- ke is the number of keys that need to be moved in order to achieve a balanced
cluster
There is a degenerate edge case O(wt + kt*log(we)) when kt is much greater than
the number of allowed keys, or when most keys are replicated or cannot be
moved for some other reason.
Returns list of tuples to feed into _rebalance_move_data:
- sender worker
- recipient worker
- task to be transferred
"""
# Heaps of workers, managed by the heapq module, that need to send/receive data,
# with how many bytes each needs to send/receive.
#
# Each element of the heap is a tuple constructed as follows:
# - snd_bytes_max/rec_bytes_max: maximum number of bytes to send or receive.
# This number is negative, so that the workers farthest from the cluster mean
# are at the top of the smallest-first heaps.
# - snd_bytes_min/rec_bytes_min: minimum number of bytes after sending/receiving
# which the worker should not be considered anymore. This is also negative.
# - arbitrary unique number, there just to to make sure that WorkerState objects
# are never used for sorting in the unlikely event that two processes have
# exactly the same number of bytes allocated.
# - WorkerState
# - iterator of all tasks in memory on the worker (senders only), insertion
# sorted (least recently inserted first).
# Note that this iterator will typically *not* be exhausted. It will only be
# exhausted if, after moving away from the worker all keys that can be moved,
# is insufficient to drop snd_bytes_min above 0.
senders: list[tuple[int, int, int, WorkerState, Iterator[TaskState]]] = []
recipients: list[tuple[int, int, int, WorkerState]] = []
# Output: [(sender, recipient, task), ...]
msgs: list[tuple[WorkerState, WorkerState, TaskState]] = []
# By default, this is the optimistic memory, meaning total process memory minus
# unmanaged memory that appeared over the last 30 seconds
# (distributed.worker.memory.recent-to-old-time).
# This lets us ignore temporary spikes caused by task heap usage.
memory_by_worker = [
(ws, getattr(ws.memory, self.MEMORY_REBALANCE_MEASURE)) for ws in workers
]
mean_memory = sum(m for _, m in memory_by_worker) // len(memory_by_worker)
for ws, ws_memory in memory_by_worker:
if ws.memory_limit:
half_gap = int(self.MEMORY_REBALANCE_HALF_GAP * ws.memory_limit)
sender_min = self.MEMORY_REBALANCE_SENDER_MIN * ws.memory_limit
recipient_max = self.MEMORY_REBALANCE_RECIPIENT_MAX * ws.memory_limit
else:
half_gap = 0
sender_min = 0.0
recipient_max = math.inf
if (
ws._has_what
and ws_memory >= mean_memory + half_gap
and ws_memory >= sender_min
):
# This may send the worker below sender_min (by design)
snd_bytes_max = mean_memory - ws_memory # negative
snd_bytes_min = snd_bytes_max + half_gap # negative
# See definition of senders above
senders.append(
(snd_bytes_max, snd_bytes_min, id(ws), ws, iter(ws._has_what))
)
elif ws_memory < mean_memory - half_gap and ws_memory < recipient_max:
# This may send the worker above recipient_max (by design)
rec_bytes_max = ws_memory - mean_memory # negative
rec_bytes_min = rec_bytes_max + half_gap # negative
# See definition of recipients above
recipients.append((rec_bytes_max, rec_bytes_min, id(ws), ws))
# Fast exit in case no transfers are necessary or possible
if not senders or not recipients:
self.log_event(
"all",
{
"action": "rebalance",
"senders": len(senders),
"recipients": len(recipients),
"moved_keys": 0,
},
)
return []
heapq.heapify(senders)
heapq.heapify(recipients)
while senders and recipients:
snd_bytes_max, snd_bytes_min, _, snd_ws, ts_iter = senders[0]
# Iterate through tasks in memory, least recently inserted first
for ts in ts_iter:
if keys is not None and ts.key not in keys:
continue
nbytes = ts.nbytes
if nbytes + snd_bytes_max > 0:
# Moving this task would cause the sender to go below mean and
# potentially risk becoming a recipient, which would cause tasks to
# bounce around. Move on to the next task of the same sender.
continue
# Find the recipient, farthest from the mean, which
# 1. has enough available RAM for this task, and
# 2. doesn't hold a copy of this task already
# There may not be any that satisfies these conditions; in this case
# this task won't be moved.
skipped_recipients = []
use_recipient = False
while recipients and not use_recipient:
rec_bytes_max, rec_bytes_min, _, rec_ws = recipients[0]
if nbytes + rec_bytes_max > 0:
# recipients are sorted by rec_bytes_max.
# The next ones will be worse; no reason to continue iterating
break
use_recipient = ts not in rec_ws._has_what
if not use_recipient:
skipped_recipients.append(heapq.heappop(recipients))
for recipient in skipped_recipients:
heapq.heappush(recipients, recipient)
if not use_recipient:
# This task has no recipients available. Leave it on the sender and
# move on to the next task of the same sender.
continue
# Schedule task for transfer from sender to recipient
msgs.append((snd_ws, rec_ws, ts))
# *_bytes_max/min are all negative for heap sorting
snd_bytes_max += nbytes
snd_bytes_min += nbytes
rec_bytes_max += nbytes
rec_bytes_min += nbytes
# Stop iterating on the tasks of this sender for now and, if it still
# has bytes to lose, push it back into the senders heap; it may or may
# not come back on top again.
if snd_bytes_min < 0:
# See definition of senders above
heapq.heapreplace(
senders,
(snd_bytes_max, snd_bytes_min, id(snd_ws), snd_ws, ts_iter),
)
else:
heapq.heappop(senders)
# If recipient still has bytes to gain, push it back into the recipients
# heap; it may or may not come back on top again.
if rec_bytes_min < 0:
# See definition of recipients above
heapq.heapreplace(
recipients,
(rec_bytes_max, rec_bytes_min, id(rec_ws), rec_ws),
)
else:
heapq.heappop(recipients)
# Move to next sender with the most data to lose.
# It may or may not be the same sender again.
break
else: # for ts in ts_iter
# Exhausted tasks on this sender
heapq.heappop(senders)
return msgs
async def _rebalance_move_data(
self, msgs: "list[tuple[WorkerState, WorkerState, TaskState]]", stimulus_id: str
) -> dict:
"""Perform the actual transfer of data across the network in rebalance().
Takes in input the output of _rebalance_find_msgs(), that is a list of tuples:
- sender worker
- recipient worker
- task to be transferred
FIXME this method is not robust when the cluster is not idle.
"""
to_recipients: defaultdict[str, defaultdict[str, list[str]]] = defaultdict(
lambda: defaultdict(list)
)
for snd_ws, rec_ws, ts in msgs:
to_recipients[rec_ws.address][ts.key].append(snd_ws.address)
failed_keys_by_recipient = dict(
zip(
to_recipients,
await asyncio.gather(
*(
# Note: this never raises exceptions
self.gather_on_worker(w, who_has)
for w, who_has in to_recipients.items()
)
),
)
)
to_senders = defaultdict(list)
for snd_ws, rec_ws, ts in msgs:
if ts.key not in failed_keys_by_recipient[rec_ws.address]:
to_senders[snd_ws.address].append(ts.key)
# Note: this never raises exceptions
await asyncio.gather(
*(self.delete_worker_data(r, v, stimulus_id) for r, v in to_senders.items())
)
for r, v in to_recipients.items():
self.log_event(r, {"action": "rebalance", "who_has": v})
self.log_event(
"all",
{
"action": "rebalance",
"senders": valmap(len, to_senders),
"recipients": valmap(len, to_recipients),
"moved_keys": len(msgs),
},
)
missing_keys = {k for r in failed_keys_by_recipient.values() for k in r}
if missing_keys:
return {"status": "partial-fail", "keys": list(missing_keys)}
else:
return {"status": "OK"}
async def replicate(
self,
comm=None,
keys=None,
n=None,
workers=None,
branching_factor=2,
delete=True,
lock=True,
stimulus_id=None,
):
"""Replicate data throughout cluster
This performs a tree copy of the data throughout the network
individually on each piece of data.
Parameters
----------
keys: Iterable
list of keys to replicate
n: int
Number of replications we expect to see within the cluster
branching_factor: int, optional
The number of workers that can copy data in each generation.
The larger the branching factor, the more data we copy in
a single step, but the more a given worker risks being
swamped by data requests.
See also
--------
Scheduler.rebalance
"""
stimulus_id = stimulus_id or f"replicate-{time()}"
assert branching_factor > 0
async with self._lock if lock else empty_context:
if workers is not None:
workers = {self.workers[w] for w in self.workers_list(workers)}
workers = {ws for ws in workers if ws.status == Status.running}
else:
workers = self.running
if n is None:
n = len(workers)
else:
n = min(n, len(workers))
if n == 0:
raise ValueError("Can not use replicate to delete data")
tasks = {self.tasks[k] for k in keys}
missing_data = [ts.key for ts in tasks if not ts.who_has]
if missing_data:
return {"status": "partial-fail", "keys": missing_data}
# Delete extraneous data
if delete:
del_worker_tasks = defaultdict(set)
for ts in tasks:
del_candidates = tuple(ts.who_has & workers)
if len(del_candidates) > n:
for ws in random.sample(
del_candidates, len(del_candidates) - n
):
del_worker_tasks[ws].add(ts)
# Note: this never raises exceptions
await asyncio.gather(
*[
self.delete_worker_data(
ws.address, [t.key for t in tasks], stimulus_id
)
for ws, tasks in del_worker_tasks.items()
]
)
# Copy not-yet-filled data
while tasks:
gathers = defaultdict(dict)
for ts in list(tasks):
if ts.state == "forgotten":
# task is no longer needed by any client or dependent task
tasks.remove(ts)
continue
n_missing = n - len(ts.who_has & workers)
if n_missing <= 0:
# Already replicated enough
tasks.remove(ts)
continue
count = min(n_missing, branching_factor * len(ts.who_has))
assert count > 0
for ws in random.sample(tuple(workers - ts.who_has), count):
gathers[ws.address][ts.key] = [
wws.address for wws in ts.who_has
]
await asyncio.gather(
*(
# Note: this never raises exceptions
self.gather_on_worker(w, who_has)
for w, who_has in gathers.items()
)
)
for r, v in gathers.items():
self.log_event(r, {"action": "replicate-add", "who_has": v})
self.log_event(
"all",
{
"action": "replicate",
"workers": list(workers),
"key-count": len(keys),
"branching-factor": branching_factor,
},
)
def workers_to_close(
self,
comm=None,
memory_ratio: int | float | None = None,
n: int | None = None,
key: Callable[[WorkerState], Hashable] | bytes | None = None,
minimum: int | None = None,
target: int | None = None,
attribute: str = "address",
) -> list[str]:
"""
Find workers that we can close with low cost
This returns a list of workers that are good candidates to retire.
These workers are not running anything and are storing
relatively little data relative to their peers. If all workers are
idle then we still maintain enough workers to have enough RAM to store
our data, with a comfortable buffer.
This is for use with systems like ``distributed.deploy.adaptive``.
Parameters
----------
memory_ratio : Number
Amount of extra space we want to have for our stored data.
Defaults to 2, or that we want to have twice as much memory as we
currently have data.
n : int
Number of workers to close
minimum : int
Minimum number of workers to keep around
key : Callable(WorkerState)
An optional callable mapping a WorkerState object to a group
affiliation. Groups will be closed together. This is useful when
closing workers must be done collectively, such as by hostname.
target : int
Target number of workers to have after we close
attribute : str
The attribute of the WorkerState object to return, like "address"
or "name". Defaults to "address".
Examples
--------
>>> scheduler.workers_to_close()
['tcp://192.168.0.1:1234', 'tcp://192.168.0.2:1234']
Group workers by hostname prior to closing
>>> scheduler.workers_to_close(key=lambda ws: ws.host)
['tcp://192.168.0.1:1234', 'tcp://192.168.0.1:4567']
Remove two workers
>>> scheduler.workers_to_close(n=2)
Keep enough workers to have twice as much memory as we we need.
>>> scheduler.workers_to_close(memory_ratio=2)
Returns
-------
to_close: list of worker addresses that are OK to close
See Also
--------
Scheduler.retire_workers
"""
if target is not None and n is None:
n = len(self.workers) - target
if n is not None:
if n < 0:
n = 0
target = len(self.workers) - n
if n is None and memory_ratio is None:
memory_ratio = 2
with log_errors():
if not n and all([ws.processing for ws in self.workers.values()]):
return []
if key is None:
key = operator.attrgetter("address")
if isinstance(key, bytes) and dask.config.get(
"distributed.scheduler.pickle"
):
key = pickle.loads(key)
groups = groupby(key, self.workers.values())
limit_bytes = {
k: sum(ws.memory_limit for ws in v) for k, v in groups.items()
}
group_bytes = {k: sum(ws.nbytes for ws in v) for k, v in groups.items()}
limit = sum(limit_bytes.values())
total = sum(group_bytes.values())
def _key(group):
is_idle = not any([wws.processing for wws in groups[group]])
bytes = -group_bytes[group]
return is_idle, bytes
idle = sorted(groups, key=_key)
to_close = []
n_remain = len(self.workers)
while idle:
group = idle.pop()
if n is None and any([ws.processing for ws in groups[group]]):
break
if minimum and n_remain - len(groups[group]) < minimum:
break
limit -= limit_bytes[group]
if (
n is not None and n_remain - len(groups[group]) >= (target or 0)
) or (memory_ratio is not None and limit >= memory_ratio * total):
to_close.append(group)
n_remain -= len(groups[group])
else:
break
result = [getattr(ws, attribute) for g in to_close for ws in groups[g]]
if result:
logger.debug("Suggest closing workers: %s", result)
return result
@log_errors
async def retire_workers(
self,
workers: list[str] | None = None,
*,
names: list | None = None,
close_workers: bool = False,
remove: bool = True,
stimulus_id: str | None = None,
**kwargs: Any,
) -> dict[str, Any]:
"""Gracefully retire workers from cluster. Any key that is in memory exclusively
on the retired workers is replicated somewhere else.
Parameters
----------
workers: list[str] (optional)
List of worker addresses to retire.
names: list (optional)
List of worker names to retire.
Mutually exclusive with ``workers``.
If neither ``workers`` nor ``names`` are provided, we call
``workers_to_close`` which finds a good set.
close_workers: bool (defaults to False)
Whether or not to actually close the worker explicitly from here.
Otherwise we expect some external job scheduler to finish off the
worker.
remove: bool (defaults to True)
Whether or not to remove the worker metadata immediately or else
wait for the worker to contact us.
If close_workers=False and remove=False, this method just flushes the tasks
in memory out of the workers and then returns.
If close_workers=True and remove=False, this method will return while the
workers are still in the cluster, although they won't accept new tasks.
If close_workers=False or for whatever reason a worker doesn't accept the
close command, it will be left permanently unable to accept new tasks and
it is expected to be closed in some other way.
**kwargs: dict
Extra options to pass to workers_to_close to determine which
workers we should drop
Returns
-------
Dictionary mapping worker ID/address to dictionary of information about
that worker for each retired worker.
If there are keys that exist in memory only on the workers being retired and it
was impossible to replicate them somewhere else (e.g. because there aren't
any other running workers), the workers holding such keys won't be retired and
won't appear in the returned dict.
See Also
--------
Scheduler.workers_to_close
"""
stimulus_id = stimulus_id or f"retire-workers-{time()}"
# This lock makes retire_workers, rebalance, and replicate mutually
# exclusive and will no longer be necessary once rebalance and replicate are
# migrated to the Active Memory Manager.
# Note that, incidentally, it also prevents multiple calls to retire_workers
# from running in parallel - this is unnecessary.
async with self._lock:
if names is not None:
if workers is not None:
raise TypeError("names and workers are mutually exclusive")
if names:
logger.info("Retire worker names %s", names)
# Support cases where names are passed through a CLI and become
# strings
names_set = {str(name) for name in names}
wss = {ws for ws in self.workers.values() if str(ws.name) in names_set}
elif workers is not None:
wss = {
self.workers[address]
for address in workers
if address in self.workers
}
else:
wss = {
self.workers[address] for address in self.workers_to_close(**kwargs)
}
if not wss:
return {}
stop_amm = False
amm: ActiveMemoryManagerExtension = self.extensions["amm"]
if not amm.running:
amm = ActiveMemoryManagerExtension(
self, policies=set(), register=False, start=True, interval=2.0
)
stop_amm = True
try:
coros = []
for ws in wss:
logger.info("Retiring worker %s", ws.address)
policy = RetireWorker(ws.address)
amm.add_policy(policy)
# Change Worker.status to closing_gracefully. Immediately set
# the same on the scheduler to prevent race conditions.
prev_status = ws.status
self.handle_worker_status_change(
Status.closing_gracefully, ws, stimulus_id
)
# FIXME: We should send a message to the nanny first;
# eventually workers won't be able to close their own nannies.
self.stream_comms[ws.address].send(
{
"op": "worker-status-change",
"status": ws.status.name,
"stimulus_id": stimulus_id,
}
)
coros.append(
self._track_retire_worker(
ws,
policy,
prev_status=prev_status,
close=close_workers,
remove=remove,
stimulus_id=stimulus_id,
)
)
# Give the AMM a kick, in addition to its periodic running. This is
# to avoid unnecessarily waiting for a potentially arbitrarily long
# time (depending on interval settings)
amm.run_once()
workers_info = dict(await asyncio.gather(*coros))
workers_info.pop(None, None)
finally:
if stop_amm:
amm.stop()
self.log_event("all", {"action": "retire-workers", "workers": workers_info})
self.log_event(list(workers_info), {"action": "retired"})
return workers_info
async def _track_retire_worker(
self,
ws: WorkerState,
policy: RetireWorker,
prev_status: Status,
close: bool,
remove: bool,
stimulus_id: str,
) -> tuple: # tuple[str | None, dict]
while not policy.done():
# Sleep 0.01s when there are 4 tasks or less
# Sleep 0.5s when there are 200 or more
poll_interval = max(0.01, min(0.5, len(ws.has_what) / 400))
await asyncio.sleep(poll_interval)
if policy.no_recipients:
# Abort retirement. This time we don't need to worry about race
# conditions and we can wait for a scheduler->worker->scheduler
# round-trip.
self.stream_comms[ws.address].send(
{
"op": "worker-status-change",
"status": prev_status.name,
"stimulus_id": stimulus_id,
}
)
return None, {}
logger.debug(
"All unique keys on worker %s have been replicated elsewhere", ws.address
)
if remove:
await self.remove_worker(
ws.address, safe=True, close=close, stimulus_id=stimulus_id
)
elif close:
self.close_worker(ws.address)
logger.info("Retired worker %s", ws.address)
return ws.address, ws.identity()
def add_keys(self, worker=None, keys=(), stimulus_id=None):
"""
Learn that a worker has certain keys
This should not be used in practice and is mostly here for legacy
reasons. However, it is sent by workers from time to time.
"""
if worker not in self.workers:
return "not found"
ws: WorkerState = self.workers[worker]
redundant_replicas = []
for key in keys:
ts: TaskState = self.tasks.get(key)
if ts is not None and ts.state == "memory":
if ws not in ts.who_has:
self.add_replica(ts, ws)
else:
redundant_replicas.append(key)
if redundant_replicas:
if not stimulus_id:
stimulus_id = f"redundant-replicas-{time()}"
self.worker_send(
worker,
{
"op": "remove-replicas",
"keys": redundant_replicas,
"stimulus_id": stimulus_id,
},
)
return "OK"
@log_errors
def update_data(
self,
*,
who_has: dict,
nbytes: dict,
client=None,
):
"""
Learn that new data has entered the network from an external source
See Also
--------
Scheduler.mark_key_in_memory
"""
who_has = {k: [self.coerce_address(vv) for vv in v] for k, v in who_has.items()}
logger.debug("Update data %s", who_has)
for key, workers in who_has.items():
ts = self.tasks.get(key)
if ts is None:
ts = self.new_task(key, None, "memory")
ts.state = "memory"
ts_nbytes = nbytes.get(key, -1)
if ts_nbytes >= 0:
ts.set_nbytes(ts_nbytes)
for w in workers:
ws = self.workers[w]
if ws not in ts.who_has:
self.add_replica(ts, ws)
self.report({"op": "key-in-memory", "key": key, "workers": list(workers)})
if client:
self.client_desires_keys(keys=list(who_has), client=client)
@overload
def report_on_key(self, key: str, *, client: str | None = None) -> None:
...
@overload
def report_on_key(self, *, ts: TaskState, client: str | None = None) -> None:
...
def report_on_key(self, key=None, *, ts=None, client=None):
if (ts is None) == (key is None):
raise ValueError( # pragma: nocover
f"ts and key are mutually exclusive; received {key=!r}, {ts=!r}"
)
if ts is None:
assert key is not None
ts = self.tasks.get(key)
else:
key = ts.key
if ts is not None:
report_msg = _task_to_report_msg(ts)
else:
report_msg = {"op": "cancelled-key", "key": key}
if report_msg is not None:
self.report(report_msg, ts=ts, client=client)
@log_errors
async def feed(
self, comm, function=None, setup=None, teardown=None, interval="1s", **kwargs
):
"""
Provides a data Comm to external requester
Caution: this runs arbitrary Python code on the scheduler. This should
eventually be phased out. It is mostly used by diagnostics.
"""
if not dask.config.get("distributed.scheduler.pickle"):
logger.warning(
"Tried to call 'feed' route with custom functions, but "
"pickle is disallowed. Set the 'distributed.scheduler.pickle'"
"config value to True to use the 'feed' route (this is mostly "
"commonly used with progress bars)"
)
return
interval = parse_timedelta(interval)
if function:
function = pickle.loads(function)
if setup:
setup = pickle.loads(setup)
if teardown:
teardown = pickle.loads(teardown)
state = setup(self) if setup else None
if inspect.isawaitable(state):
state = await state
try:
while self.status == Status.running:
if state is None:
response = function(self)
else:
response = function(self, state)
await comm.write(response)
await asyncio.sleep(interval)
except OSError:
pass
finally:
if teardown:
teardown(self, state)
def log_worker_event(
self, worker: str, topic: str | Collection[str], msg: Any
) -> None:
if isinstance(msg, dict):
msg["worker"] = worker
self.log_event(topic, msg)
def subscribe_worker_status(self, comm=None):
WorkerStatusPlugin(self, comm)
ident = self.identity()
for v in ident["workers"].values():
del v["metrics"]
del v["last_seen"]
return ident
def get_processing(self, workers=None):
if workers is not None:
workers = set(map(self.coerce_address, workers))
return {w: [ts.key for ts in self.workers[w].processing] for w in workers}
else:
return {
w: [ts.key for ts in ws.processing] for w, ws in self.workers.items()
}
def get_who_has(self, keys: Iterable[str] | None = None) -> dict[str, list[str]]:
if keys is not None:
return {
key: [ws.address for ws in self.tasks[key].who_has]
if key in self.tasks
else []
for key in keys
}
else:
return {
key: [ws.address for ws in ts.who_has] for key, ts in self.tasks.items()
}
def get_has_what(self, workers=None):
if workers is not None:
workers = map(self.coerce_address, workers)
return {
w: [ts.key for ts in self.workers[w].has_what]
if w in self.workers
else []
for w in workers
}
else:
return {w: [ts.key for ts in ws.has_what] for w, ws in self.workers.items()}
def get_ncores(self, workers=None):
if workers is not None:
workers = map(self.coerce_address, workers)
return {w: self.workers[w].nthreads for w in workers if w in self.workers}
else:
return {w: ws.nthreads for w, ws in self.workers.items()}
def get_ncores_running(self, workers=None):
ncores = self.get_ncores(workers=workers)
return {
w: n for w, n in ncores.items() if self.workers[w].status == Status.running
}
async def get_call_stack(self, keys=None):
if keys is not None:
stack = list(keys)
processing = set()
while stack:
key = stack.pop()
ts = self.tasks[key]
if ts.state == "waiting":
stack.extend([dts.key for dts in ts.dependencies])
elif ts.state == "processing":
processing.add(ts)
workers = defaultdict(list)
for ts in processing:
if ts.processing_on:
workers[ts.processing_on.address].append(ts.key)
else:
workers = {w: None for w in self.workers}
if not workers:
return {}
results = await asyncio.gather(
*(self.rpc(w).call_stack(keys=v) for w, v in workers.items())
)
response = {w: r for w, r in zip(workers, results) if r}
return response
async def benchmark_hardware(self) -> "dict[str, dict[str, float]]":
"""
Run a benchmark on the workers for memory, disk, and network bandwidths
Returns
-------
result: dict
A dictionary mapping the names "disk", "memory", and "network" to
dictionaries mapping sizes to bandwidths. These bandwidths are
averaged over many workers running computations across the cluster.
"""
out: dict[str, defaultdict[str, list[float]]] = {
name: defaultdict(list) for name in ["disk", "memory", "network"]
}
# disk
result = await self.broadcast(msg={"op": "benchmark_disk"})
for d in result.values():
for size, duration in d.items():
out["disk"][size].append(duration)
# memory
result = await self.broadcast(msg={"op": "benchmark_memory"})
for d in result.values():
for size, duration in d.items():
out["memory"][size].append(duration)
# network
workers = list(self.workers)
# On an adaptive cluster, if multiple workers are started on the same physical host,
# they are more likely to connect to the Scheduler in sequence, ending up next to
# each other in this list.
# The transfer speed within such clusters of workers will be effectively that of
# localhost. This could happen across different VMs and/or docker images, so
# implementing logic based on IP addresses would not necessarily help.
# Randomize the connections to even out the mean measures.
random.shuffle(workers)
futures = [
self.rpc(a).benchmark_network(address=b) for a, b in partition(2, workers)
]
responses = await asyncio.gather(*futures)
for d in responses:
for size, duration in d.items():
out["network"][size].append(duration)
result = {}
for mode in out:
result[mode] = {
size: sum(durations) / len(durations)
for size, durations in out[mode].items()
}
return result
@log_errors
def get_nbytes(self, keys=None, summary=True):
if keys is not None:
result = {k: self.tasks[k].nbytes for k in keys}
else:
result = {k: ts.nbytes for k, ts in self.tasks.items() if ts.nbytes >= 0}
if summary:
out = defaultdict(lambda: 0)
for k, v in result.items():
out[key_split(k)] += v
result = dict(out)
return result
def run_function(self, comm, function, args=(), kwargs=None, wait=True):
"""Run a function within this process
See Also
--------
Client.run_on_scheduler
"""
from distributed.worker import run
if not dask.config.get("distributed.scheduler.pickle"):
raise ValueError(
"Cannot run function as the scheduler has been explicitly disallowed from "
"deserializing arbitrary bytestrings using pickle via the "
"'distributed.scheduler.pickle' configuration setting."
)
kwargs = kwargs or {}
self.log_event("all", {"action": "run-function", "function": function})
return run(self, comm, function=function, args=args, kwargs=kwargs, wait=wait)
def set_metadata(self, keys: list[str], value: object = None) -> None:
metadata = self.task_metadata
for key in keys[:-1]:
if key not in metadata or not isinstance(metadata[key], (dict, list)):
metadata[key] = {}
metadata = metadata[key]
metadata[keys[-1]] = value
def get_metadata(self, keys: list[str], default=no_default):
metadata = self.task_metadata
try:
for key in keys:
metadata = metadata[key]
return metadata
except KeyError:
if default != no_default:
return default
else:
raise
def set_restrictions(self, worker: dict[str, Collection[str] | str]):
for key, restrictions in worker.items():
ts = self.tasks[key]
if isinstance(restrictions, str):
restrictions = {restrictions}
ts.worker_restrictions = set(restrictions)
@log_errors
def get_task_prefix_states(self):
state = {}
for tp in self.task_prefixes.values():
active_states = tp.active_states
if any(
active_states.get(s)
for s in {"memory", "erred", "released", "processing", "waiting"}
):
state[tp.name] = {
"memory": active_states["memory"],
"erred": active_states["erred"],
"released": active_states["released"],
"processing": active_states["processing"],
"waiting": active_states["waiting"],
}
return state
def get_task_status(self, keys=None):
return {
key: (self.tasks[key].state if key in self.tasks else None) for key in keys
}
def get_task_stream(self, start=None, stop=None, count=None):
from distributed.diagnostics.task_stream import TaskStreamPlugin
if TaskStreamPlugin.name not in self.plugins:
self.add_plugin(TaskStreamPlugin(self))
plugin = self.plugins[TaskStreamPlugin.name]
return plugin.collect(start=start, stop=stop, count=count)
def start_task_metadata(self, name=None):
plugin = CollectTaskMetaDataPlugin(scheduler=self, name=name)
self.add_plugin(plugin)
def stop_task_metadata(self, name=None):
plugins = [
p
for p in list(self.plugins.values())
if isinstance(p, CollectTaskMetaDataPlugin) and p.name == name
]
if len(plugins) != 1:
raise ValueError(
"Expected to find exactly one CollectTaskMetaDataPlugin "
f"with name {name} but found {len(plugins)}."
)
plugin = plugins[0]
self.remove_plugin(name=plugin.name)
return {"metadata": plugin.metadata, "state": plugin.state}
async def register_worker_plugin(self, comm, plugin, name=None):
"""Registers a worker plugin on all running and future workers"""
self.worker_plugins[name] = plugin
responses = await self.broadcast(
msg=dict(op="plugin-add", plugin=plugin, name=name)
)
return responses
async def unregister_worker_plugin(self, comm, name):
"""Unregisters a worker plugin"""
try:
self.worker_plugins.pop(name)
except KeyError:
raise ValueError(f"The worker plugin {name} does not exist")
responses = await self.broadcast(msg=dict(op="plugin-remove", name=name))
return responses
async def register_nanny_plugin(self, comm, plugin, name=None):
"""Registers a setup function, and call it on every worker"""
self.nanny_plugins[name] = plugin
responses = await self.broadcast(
msg=dict(op="plugin_add", plugin=plugin, name=name),
nanny=True,
)
return responses
async def unregister_nanny_plugin(self, comm, name):
"""Unregisters a worker plugin"""
try:
self.nanny_plugins.pop(name)
except KeyError:
raise ValueError(f"The nanny plugin {name} does not exist")
responses = await self.broadcast(
msg=dict(op="plugin_remove", name=name), nanny=True
)
return responses
def transition(
self,
key: str,
finish: TaskStateState,
stimulus_id: str,
**kwargs: Any,
) -> Recs:
"""Transition a key from its current state to the finish state
Examples
--------
>>> self.transition('x', 'waiting')
{'x': 'processing'}
Returns
-------
Dictionary of recommendations for future transitions
See Also
--------
Scheduler.transitions: transitive version of this function
"""
recommendations, client_msgs, worker_msgs = self._transition(
key, finish, stimulus_id, **kwargs
)
self.send_all(client_msgs, worker_msgs)
return recommendations
def transitions(self, recommendations: Recs, stimulus_id: str) -> None:
"""Process transitions until none are left
This includes feedback from previous transitions and continues until we
reach a steady state
"""
client_msgs: Msgs = {}
worker_msgs: Msgs = {}
self._transitions(recommendations, client_msgs, worker_msgs, stimulus_id)
self.send_all(client_msgs, worker_msgs)
async def get_story(self, keys_or_stimuli: Iterable[str]) -> list[Transition]:
"""RPC hook for :meth:`SchedulerState.story`.
Note that the msgpack serialization/deserialization round-trip will transform
the :class:`Transition` namedtuples into regular tuples.
"""
return self.story(*keys_or_stimuli)
def _reschedule(
self, key: str, worker: str | None = None, *, stimulus_id: str
) -> None:
"""Reschedule a task.
This function should only be used when the task has already been released in
some way on the worker it's assigned to — either via cancellation or a
Reschedule exception — and you are certain the worker will not send any further
updates about the task to the scheduler.
"""
try:
ts = self.tasks[key]
except KeyError:
logger.warning(
f"Attempting to reschedule task {key}, which was not "
"found on the scheduler. Aborting reschedule."
)
return
if ts.state != "processing":
return
if worker and ts.processing_on and ts.processing_on.address != worker:
return
# transition_processing_released will immediately suggest an additional
# transition to waiting if the task has any waiters or clients holding a future.
self.transitions({key: "released"}, stimulus_id=stimulus_id)
#####################
# Utility functions #
#####################
def add_resources(self, worker: str, resources=None):
ws: WorkerState = self.workers[worker]
if resources:
ws.resources.update(resources)
ws.used_resources = {}
for resource, quantity in ws.resources.items():
ws.used_resources[resource] = 0
dr = self.resources.get(resource, None)
if dr is None:
self.resources[resource] = dr = {}
dr[worker] = quantity
return "OK"
def remove_resources(self, worker):
ws: WorkerState = self.workers[worker]
for resource in ws.resources:
dr: dict = self.resources.get(resource, None)
if dr is None:
self.resources[resource] = dr = {}
del dr[worker]
def coerce_address(self, addr, resolve=True):
"""
Coerce possible input addresses to canonical form.
*resolve* can be disabled for testing with fake hostnames.
Handles strings, tuples, or aliases.
"""
# XXX how many address-parsing routines do we have?
if addr in self.aliases:
addr = self.aliases[addr]
if isinstance(addr, tuple):
addr = unparse_host_port(*addr)
if not isinstance(addr, str):
raise TypeError(f"addresses should be strings or tuples, got {addr!r}")
if resolve:
addr = resolve_address(addr)
else:
addr = normalize_address(addr)
return addr
def workers_list(self, workers):
"""
List of qualifying workers
Takes a list of worker addresses or hostnames.
Returns a list of all worker addresses that match
"""
if workers is None:
return list(self.workers)
out = set()
for w in workers:
if ":" in w:
out.add(w)
else:
out.update({ww for ww in self.workers if w in ww}) # TODO: quadratic
return list(out)
async def get_profile(
self,
comm=None,
workers=None,
scheduler=False,
server=False,
merge_workers=True,
start=None,
stop=None,
key=None,
):
if workers is None:
workers = self.workers
else:
workers = set(self.workers) & set(workers)
if scheduler:
return profile.get_profile(self.io_loop.profile, start=start, stop=stop)
results = await asyncio.gather(
*(
self.rpc(w).profile(start=start, stop=stop, key=key, server=server)
for w in workers
),
return_exceptions=True,
)
results = [r for r in results if not isinstance(r, Exception)]
if merge_workers:
response = profile.merge(*results)
else:
response = dict(zip(workers, results))
return response
async def get_profile_metadata(
self,
workers: "Iterable[str] | None" = None,
start: float = 0,
stop: "float | None" = None,
profile_cycle_interval: "str | float | None" = None,
):
dt = profile_cycle_interval or dask.config.get(
"distributed.worker.profile.cycle"
)
dt = parse_timedelta(dt, default="ms")
if workers is None:
workers = self.workers
else:
workers = set(self.workers) & set(workers)
results = await asyncio.gather(
*(self.rpc(w).profile_metadata(start=start, stop=stop) for w in workers),
return_exceptions=True,
)
results = [r for r in results if not isinstance(r, Exception)]
counts = [
(time, sum(pluck(1, group)))
for time, group in itertools.groupby(
merge_sorted(
*(v["counts"] for v in results),
),
lambda t: t[0] // dt * dt,
)
]
keys: dict[str, list[list]] = {
k: [] for v in results for t, d in v["keys"] for k in d
}
groups1 = [v["keys"] for v in results]
groups2 = list(merge_sorted(*groups1, key=first))
last = 0
for t, d in groups2:
tt = t // dt * dt
if tt > last:
last = tt
for v in keys.values():
v.append([tt, 0])
for k, v in d.items():
keys[k][-1][1] += v
return {"counts": counts, "keys": keys}
async def performance_report(
self, start: float, last_count: int, code="", mode=None
):
stop = time()
# Profiles
compute, scheduler, workers = await asyncio.gather(
*[
self.get_profile(start=start),
self.get_profile(scheduler=True, start=start),
self.get_profile(server=True, start=start),
]
)
from distributed import profile
def profile_to_figure(state):
data = profile.plot_data(state)
figure, source = profile.plot_figure(data, sizing_mode="stretch_both")
return figure
compute, scheduler, workers = map(
profile_to_figure, (compute, scheduler, workers)
)
# Task stream
task_stream = self.get_task_stream(start=start)
total_tasks = len(task_stream)
timespent: defaultdict[str, float] = defaultdict(float)
for d in task_stream:
for x in d["startstops"]:
timespent[x["action"]] += x["stop"] - x["start"]
tasks_timings = ""
for k in sorted(timespent.keys()):
tasks_timings += f"\n<li> {k} time: {format_time(timespent[k])} </li>"
from distributed.dashboard.components.scheduler import task_stream_figure
from distributed.diagnostics.task_stream import rectangles
rects = rectangles(task_stream)
source, task_stream = task_stream_figure(sizing_mode="stretch_both")
source.data.update(rects)
# Bandwidth
from distributed.dashboard.components.scheduler import (
BandwidthTypes,
BandwidthWorkers,
)
bandwidth_workers = BandwidthWorkers(self, sizing_mode="stretch_both")
bandwidth_workers.update()
bandwidth_types = BandwidthTypes(self, sizing_mode="stretch_both")
bandwidth_types.update()
# System monitor
from distributed.dashboard.components.shared import SystemMonitor
sysmon = SystemMonitor(self, last_count=last_count, sizing_mode="stretch_both")
sysmon.update()
# Scheduler logs
from distributed.dashboard.components.scheduler import (
_BOKEH_STYLES_KWARGS,
SchedulerLogs,
)
logs = SchedulerLogs(self, start=start)
from bokeh.models import Div, Tabs
import distributed
from distributed.dashboard.core import TabPanel
# HTML
ws: WorkerState
html = """
<h1> Dask Performance Report </h1>
<i> Select different tabs on the top for additional information </i>
<h2> Duration: {time} </h2>
<h2> Tasks Information </h2>
<ul>
<li> number of tasks: {ntasks} </li>
{tasks_timings}
</ul>
<h2> Scheduler Information </h2>
<ul>
<li> Address: {address} </li>
<li> Workers: {nworkers} </li>
<li> Threads: {threads} </li>
<li> Memory: {memory} </li>
<li> Dask Version: {dask_version} </li>
<li> Dask.Distributed Version: {distributed_version} </li>
</ul>
<h2> Calling Code </h2>
<pre>
{code}
</pre>
""".format(
time=format_time(stop - start),
ntasks=total_tasks,
tasks_timings=tasks_timings,
address=self.address,
nworkers=len(self.workers),
threads=sum(ws.nthreads for ws in self.workers.values()),
memory=format_bytes(sum(ws.memory_limit for ws in self.workers.values())),
code=code,
dask_version=dask.__version__,
distributed_version=distributed.__version__,
)
html = Div(text=html, **_BOKEH_STYLES_KWARGS)
html = TabPanel(child=html, title="Summary")
compute = TabPanel(child=compute, title="Worker Profile (compute)")
workers = TabPanel(child=workers, title="Worker Profile (administrative)")
scheduler = TabPanel(
child=scheduler, title="Scheduler Profile (administrative)"
)
task_stream = TabPanel(child=task_stream, title="Task Stream")
bandwidth_workers = TabPanel(
child=bandwidth_workers.root, title="Bandwidth (Workers)"
)
bandwidth_types = TabPanel(
child=bandwidth_types.root, title="Bandwidth (Types)"
)
system = TabPanel(child=sysmon.root, title="System")
logs = TabPanel(child=logs.root, title="Scheduler Logs")
tabs = Tabs(
tabs=[
html,
task_stream,
system,
logs,
compute,
workers,
scheduler,
bandwidth_workers,
bandwidth_types,
],
sizing_mode="stretch_both",
)
from bokeh.core.templates import get_env
from bokeh.plotting import output_file, save
with tmpfile(extension=".html") as fn:
output_file(filename=fn, title="Dask Performance Report", mode=mode)
template_directory = os.path.join(
os.path.dirname(os.path.abspath(__file__)), "dashboard", "templates"
)
template_environment = get_env()
template_environment.loader.searchpath.append(template_directory)
template = template_environment.get_template("performance_report.html")
save(tabs, filename=fn, template=template)
with open(fn) as f:
data = f.read()
return data
async def get_worker_logs(self, n=None, workers=None, nanny=False):
results = await self.broadcast(
msg={"op": "get_logs", "n": n}, workers=workers, nanny=nanny
)
return results
def log_event(self, topic: str | Collection[str], msg: Any) -> None:
event = (time(), msg)
if not isinstance(topic, str):
for t in topic:
self.events[t].append(event)
self.event_counts[t] += 1
self._report_event(t, event)
else:
self.events[topic].append(event)
self.event_counts[topic] += 1
self._report_event(topic, event)
for plugin in list(self.plugins.values()):
try:
plugin.log_event(topic, msg)
except Exception:
logger.info("Plugin failed with exception", exc_info=True)
def _report_event(self, name, event):
msg = {
"op": "event",
"topic": name,
"event": event,
}
client_msgs = {client: [msg] for client in self.event_subscriber[name]}
self.send_all(client_msgs, worker_msgs={})
def subscribe_topic(self, topic, client):
self.event_subscriber[topic].add(client)
def unsubscribe_topic(self, topic, client):
self.event_subscriber[topic].discard(client)
def get_events(self, topic=None):
if topic is not None:
return tuple(self.events[topic])
else:
return valmap(tuple, self.events)
async def get_worker_monitor_info(self, recent=False, starts=None):
if starts is None:
starts = {}
results = await asyncio.gather(
*(
self.rpc(w).get_monitor_info(recent=recent, start=starts.get(w, 0))
for w in self.workers
)
)
return dict(zip(self.workers, results))
###########
# Cleanup #
###########
async def check_worker_ttl(self):
now = time()
stimulus_id = f"check-worker-ttl-{now}"
for ws in self.workers.values():
if (ws.last_seen < now - self.worker_ttl) and (
ws.last_seen < now - 10 * heartbeat_interval(len(self.workers))
):
logger.warning(
"Worker failed to heartbeat within %s seconds. Closing: %s",
self.worker_ttl,
ws,
)
await self.remove_worker(address=ws.address, stimulus_id=stimulus_id)
def check_idle(self):
assert self.idle_timeout
if self.status in (Status.closing, Status.closed):
return
if self.transition_counter != self._idle_transition_counter:
self._idle_transition_counter = self.transition_counter
self.idle_since = None
return
if (
self.queued
or self.unrunnable
or any([ws.processing for ws in self.workers.values()])
):
self.idle_since = None
return
if not self.idle_since:
self.idle_since = time()
if time() > self.idle_since + self.idle_timeout:
assert self.idle_since
logger.info(
"Scheduler closing after being idle for %s",
format_time(self.idle_timeout),
)
self._ongoing_background_tasks.call_soon(self.close)
def adaptive_target(self, target_duration=None):
"""Desired number of workers based on the current workload
This looks at the current running tasks and memory use, and returns a
number of desired workers. This is often used by adaptive scheduling.
Parameters
----------
target_duration : str
A desired duration of time for computations to take. This affects
how rapidly the scheduler will ask to scale.
See Also
--------
distributed.deploy.Adaptive
"""
if target_duration is None:
target_duration = dask.config.get("distributed.adaptive.target-duration")
target_duration = parse_timedelta(target_duration)
# CPU
# TODO consider any user-specified default task durations for queued tasks
queued_occupancy = len(self.queued) * self.UNKNOWN_TASK_DURATION
cpu = math.ceil(
(self.total_occupancy + queued_occupancy) / target_duration
) # TODO: threads per worker
# Avoid a few long tasks from asking for many cores
tasks_ready = len(self.queued)
for ws in self.workers.values():
tasks_ready += len(ws.processing)
if tasks_ready > cpu:
break
else:
cpu = min(tasks_ready, cpu)
if (self.unrunnable or self.queued) and not self.workers:
cpu = max(1, cpu)
# add more workers if more than 60% of memory is used
limit = sum(ws.memory_limit for ws in self.workers.values())
used = sum(ws.nbytes for ws in self.workers.values())
memory = 0
if used > 0.6 * limit and limit > 0:
memory = 2 * len(self.workers)
target = max(memory, cpu)
if target >= len(self.workers):
return target
else: # Scale down?
to_close = self.workers_to_close()
return len(self.workers) - len(to_close)
def request_acquire_replicas(
self, addr: str, keys: Iterable[str], *, stimulus_id: str
) -> None:
"""Asynchronously ask a worker to acquire a replica of the listed keys from
other workers. This is a fire-and-forget operation which offers no feedback for
success or failure, and is intended for housekeeping and not for computation.
"""
who_has = {}
nbytes = {}
for key in keys:
ts = self.tasks[key]
assert ts.who_has
who_has[key] = [ws.address for ws in ts.who_has]
nbytes[key] = ts.nbytes
self.stream_comms[addr].send(
{
"op": "acquire-replicas",
"who_has": who_has,
"nbytes": nbytes,
"stimulus_id": stimulus_id,
},
)
def request_remove_replicas(
self, addr: str, keys: list[str], *, stimulus_id: str
) -> None:
"""Asynchronously ask a worker to discard its replica of the listed keys.
This must never be used to destroy the last replica of a key. This is a
fire-and-forget operation, intended for housekeeping and not for computation.
The replica disappears immediately from TaskState.who_has on the Scheduler side;
if the worker refuses to delete, e.g. because the task is a dependency of
another task running on it, it will (also asynchronously) inform the scheduler
to re-add itself to who_has. If the worker agrees to discard the task, there is
no feedback.
"""
ws = self.workers[addr]
# The scheduler immediately forgets about the replica and suggests the worker to
# drop it. The worker may refuse, at which point it will send back an add-keys
# message to reinstate it.
for key in keys:
ts = self.tasks[key]
if self.validate:
# Do not destroy the last copy
assert len(ts.who_has) > 1
self.remove_replica(ts, ws)
self.stream_comms[addr].send(
{
"op": "remove-replicas",
"keys": keys,
"stimulus_id": stimulus_id,
}
)
def _task_to_report_msg(ts: TaskState) -> dict[str, Any] | None:
if ts.state == "forgotten":
return {"op": "cancelled-key", "key": ts.key}
elif ts.state == "memory":
return {"op": "key-in-memory", "key": ts.key}
elif ts.state == "erred":
failing_ts = ts.exception_blame
assert failing_ts
return {
"op": "task-erred",
"key": ts.key,
"exception": failing_ts.exception,
"traceback": failing_ts.traceback,
}
else:
return None
def _task_to_client_msgs(ts: TaskState) -> dict[str, list[dict[str, Any]]]:
if ts.who_wants:
report_msg = _task_to_report_msg(ts)
if report_msg is not None:
return {cs.client_key: [report_msg] for cs in ts.who_wants}
return {}
def decide_worker(
ts: TaskState,
all_workers: Iterable[WorkerState],
valid_workers: set[WorkerState] | None,
objective: Callable[[WorkerState], Any],
) -> WorkerState | None:
"""
Decide which worker should take task *ts*.
We choose the worker that has the data on which *ts* depends.
If several workers have dependencies then we choose the less-busy worker.
Optionally provide *valid_workers* of where jobs are allowed to occur
(if all workers are allowed to take the task, pass None instead).
If the task requires data communication because no eligible worker has
all the dependencies already, then we choose to minimize the number
of bytes sent between workers. This is determined by calling the
*objective* function.
"""
assert all(dts.who_has for dts in ts.dependencies)
if ts.actor:
candidates = set(all_workers)
else:
candidates = {wws for dts in ts.dependencies for wws in dts.who_has}
if valid_workers is None:
if not candidates:
candidates = set(all_workers)
else:
candidates &= valid_workers
if not candidates:
candidates = valid_workers
if not candidates:
if ts.loose_restrictions:
return decide_worker(ts, all_workers, None, objective)
if not candidates:
return None
elif len(candidates) == 1:
return next(iter(candidates))
else:
return min(candidates, key=objective)
def validate_task_state(ts: TaskState) -> None:
"""Validate the given TaskState"""
assert ts.state in ALL_TASK_STATES, ts
if ts.waiting_on:
assert ts.waiting_on.issubset(ts.dependencies), (
"waiting not subset of dependencies",
str(ts.waiting_on),
str(ts.dependencies),
)
if ts.waiters:
assert ts.waiters.issubset(ts.dependents), (
"waiters not subset of dependents",
str(ts.waiters),
str(ts.dependents),
)
for dts in ts.waiting_on:
assert not dts.who_has, ("waiting on in-memory dep", str(ts), str(dts))
assert dts.state != "released", ("waiting on released dep", str(ts), str(dts))
for dts in ts.dependencies:
assert ts in dts.dependents, (
"not in dependency's dependents",
str(ts),
str(dts),
str(dts.dependents),
)
if ts.state in ("waiting", "queued", "processing", "no-worker"):
assert dts in ts.waiting_on or dts.who_has, (
"dep missing",
str(ts),
str(dts),
)
assert dts.state != "forgotten"
for dts in ts.waiters:
assert dts.state in ("waiting", "queued", "processing", "no-worker"), (
"waiter not in play",
str(ts),
str(dts),
)
for dts in ts.dependents:
assert ts in dts.dependencies, (
"not in dependent's dependencies",
str(ts),
str(dts),
str(dts.dependencies),
)
assert dts.state != "forgotten"
assert (ts.processing_on is not None) == (ts.state == "processing")
assert bool(ts.who_has) == (ts.state == "memory"), (ts, ts.who_has, ts.state)
if ts.state == "queued":
assert not ts.processing_on
assert not ts.who_has
assert all(dts.who_has for dts in ts.dependencies), (
"task queued without all deps",
str(ts),
str(ts.dependencies),
)
if ts.state == "processing":
assert all(dts.who_has for dts in ts.dependencies), (
"task processing without all deps",
str(ts),
str(ts.dependencies),
)
assert not ts.waiting_on
if ts.who_has:
assert ts.waiters or ts.who_wants, (
"unneeded task in memory",
str(ts),
str(ts.who_has),
)
if ts.run_spec: # was computed
assert ts.type
assert isinstance(ts.type, str)
assert not any([ts in dts.waiting_on for dts in ts.dependents])
for ws in ts.who_has:
assert ts in ws.has_what, (
"not in who_has' has_what",
str(ts),
str(ws),
str(ws.has_what),
)
for cs in ts.who_wants:
assert ts in cs.wants_what, (
"not in who_wants' wants_what",
str(ts),
str(cs),
str(cs.wants_what),
)
if ts.actor:
if ts.state == "memory":
assert sum(ts in ws.actors for ws in ts.who_has) == 1
if ts.state == "processing":
assert ts.processing_on
assert ts in ts.processing_on.actors
assert ts.state != "queued"
def validate_worker_state(ws: WorkerState) -> None:
for ts in ws.has_what:
assert ws in ts.who_has, (
"not in has_what' who_has",
str(ws),
str(ts),
str(ts.who_has),
)
for ts in ws.actors:
assert ts.state in ("memory", "processing")
def validate_state(
tasks: dict[str, TaskState],
workers: dict[str, WorkerState],
clients: dict[str, ClientState],
) -> None:
"""Validate a current runtime state.
This performs a sequence of checks on the entire graph, running in about linear
time. This raises assert errors if anything doesn't check out.
"""
for ts in tasks.values():
validate_task_state(ts)
for ws in workers.values():
validate_worker_state(ws)
for cs in clients.values():
for ts in cs.wants_what:
assert cs in ts.who_wants, (
"not in wants_what' who_wants",
str(cs),
str(ts),
str(ts.who_wants),
)
def heartbeat_interval(n: int) -> float:
"""Interval in seconds that we desire heartbeats based on number of workers"""
if n <= 10:
return 0.5
elif n < 50:
return 1
elif n < 200:
return 2
else:
# No more than 200 heartbeats a second scaled by workers
return n / 200 + 1
def _task_slots_available(ws: WorkerState, saturation_factor: float) -> int:
"""Number of tasks that can be sent to this worker without oversaturating it"""
assert not math.isinf(saturation_factor)
return max(math.ceil(saturation_factor * ws.nthreads), 1) - (
len(ws.processing) - len(ws.long_running)
)
def _worker_full(ws: WorkerState, saturation_factor: float) -> bool:
if math.isinf(saturation_factor):
return False
return _task_slots_available(ws, saturation_factor) <= 0
class KilledWorker(Exception):
def __init__(self, task: str, last_worker: WorkerState, allowed_failures: int):
super().__init__(task, last_worker, allowed_failures)
@property
def task(self) -> str:
return self.args[0]
@property
def last_worker(self) -> WorkerState:
return self.args[1]
@property
def allowed_failures(self) -> int:
return self.args[2]
def __str__(self) -> str:
return (
f"Attempted to run task {self.task} on {self.allowed_failures} different "
"workers, but all those workers died while running it. "
f"The last worker that attempt to run the task was {self.last_worker.address}. "
"Inspecting worker logs is often a good next step to diagnose what went wrong. "
"For more information see https://distributed.dask.org/en/stable/killed.html."
)
class WorkerStatusPlugin(SchedulerPlugin):
"""A plugin to share worker status with a remote observer
This is used in cluster managers to keep updated about the status of the scheduler.
"""
name: ClassVar[str] = "worker-status"
bcomm: BatchedSend
def __init__(self, scheduler: Scheduler, comm: Comm):
self.bcomm = BatchedSend(interval="5ms")
self.bcomm.start(comm)
scheduler.add_plugin(self)
def add_worker(self, scheduler: Scheduler, worker: str) -> None:
ident = scheduler.workers[worker].identity()
del ident["metrics"]
del ident["last_seen"]
try:
self.bcomm.send(["add", {"workers": {worker: ident}}])
except CommClosedError:
scheduler.remove_plugin(name=self.name)
def remove_worker(self, scheduler: Scheduler, worker: str) -> None:
try:
self.bcomm.send(["remove", worker])
except CommClosedError:
scheduler.remove_plugin(name=self.name)
def teardown(self) -> None:
self.bcomm.close()
class CollectTaskMetaDataPlugin(SchedulerPlugin):
scheduler: Scheduler
name: str
keys: set[str]
metadata: dict[str, Any]
state: dict[str, str]
def __init__(self, scheduler: Scheduler, name: str):
self.scheduler = scheduler
self.name = name
self.keys = set()
self.metadata = {}
self.state = {}
def update_graph(
self,
scheduler: Scheduler,
keys: set[str],
restrictions: dict[str, float],
**kwargs: Any,
) -> None:
self.keys.update(keys)
def transition(
self,
key: str,
start: TaskStateState,
finish: TaskStateState,
*args: Any,
**kwargs: Any,
) -> None:
if finish in ("memory", "erred"):
ts = self.scheduler.tasks.get(key)
if ts is not None and ts.key in self.keys:
self.metadata[key] = ts.metadata
self.state[key] = finish
self.keys.discard(key)
|