File: _worker_extension.py

package info (click to toggle)
dask.distributed 2022.12.1%2Bds.1-3
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 10,164 kB
  • sloc: python: 81,938; javascript: 1,549; makefile: 228; sh: 100
file content (655 lines) | stat: -rw-r--r-- 21,921 bytes parent folder | download
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
from __future__ import annotations

import asyncio
import contextlib
import functools
import logging
import os
import time
from collections import defaultdict
from collections.abc import Callable, Iterator
from concurrent.futures import ThreadPoolExecutor
from typing import TYPE_CHECKING, Any, BinaryIO, TypeVar, overload

import toolz

from dask.utils import parse_bytes

from distributed.core import PooledRPCCall
from distributed.protocol import to_serialize
from distributed.shuffle._arrow import (
    deserialize_schema,
    dump_batch,
    list_of_buffers_to_table,
    load_arrow,
)
from distributed.shuffle._comms import CommShardsBuffer
from distributed.shuffle._disk import DiskShardsBuffer
from distributed.shuffle._limiter import ResourceLimiter
from distributed.shuffle._shuffle import ShuffleId
from distributed.utils import log_errors, sync

if TYPE_CHECKING:
    import pandas as pd
    import pyarrow as pa

    from distributed.worker import Worker

T = TypeVar("T")

logger = logging.getLogger(__name__)


class ShuffleClosedError(RuntimeError):
    pass


class Shuffle:
    """State for a single active shuffle

    This object is responsible for splitting, sending, receiving and combining
    data shards.

    It is entirely agnostic to the distributed system and can perform a shuffle
    with other `Shuffle` instances using `rpc` and `broadcast`.

    The user of this needs to guarantee that only `Shuffle`s of the same unique
    `ShuffleID` interact.

    Parameters
    ----------
    worker_for:
        A mapping partition_id -> worker_address.
    output_workers:
        A set of all participating worker (addresses).
    column:
        The data column we split the input partition by.
    schema:
        The schema of the payload data.
    id:
        A unique `ShuffleID` this belongs to.
    local_address:
        The local address this Shuffle can be contacted by using `rpc`.
    directory:
        The scratch directory to buffer data in.
    nthreads:
        How many background threads to use for compute.
    loop:
        The event loop.
    rpc:
        A callable returning a PooledRPCCall to contact other Shuffle instances.
        Typically a ConnectionPool.
    broadcast:
        A function that ensures a RPC is evaluated on all `Shuffle` instances of
        a given `ShuffleID`.
    memory_limiter_disk:
    memory_limiter_comm:
        A ``ResourceLimiter`` limiting the total amount of memory used in either
        buffer.
    """

    def __init__(
        self,
        worker_for: dict[int, str],
        output_workers: set,
        column: str,
        schema: pa.Schema,
        id: ShuffleId,
        local_address: str,
        directory: str,
        nthreads: int,
        rpc: Callable[[str], PooledRPCCall],
        broadcast: Callable,
        memory_limiter_disk: ResourceLimiter,
        memory_limiter_comms: ResourceLimiter,
    ):

        import pandas as pd

        self.broadcast = broadcast
        self.rpc = rpc
        self.column = column
        self.id = id
        self.schema = schema
        self.output_workers = output_workers
        self.executor = ThreadPoolExecutor(nthreads)
        partitions_of = defaultdict(list)
        self.local_address = local_address
        for part, addr in worker_for.items():
            partitions_of[addr].append(part)
        self.partitions_of = dict(partitions_of)
        self.worker_for = pd.Series(worker_for, name="_workers").astype("category")
        self.closed = False

        def _dump_batch(batch: pa.Buffer, file: BinaryIO) -> None:
            return dump_batch(batch, file, self.schema)

        self._disk_buffer = DiskShardsBuffer(
            dump=_dump_batch,
            load=load_arrow,
            directory=directory,
            memory_limiter=memory_limiter_disk,
        )

        self._comm_buffer = CommShardsBuffer(
            send=self.send, memory_limiter=memory_limiter_comms
        )
        # TODO: reduce number of connections to number of workers
        # MultiComm.max_connections = min(10, n_workers)

        self.diagnostics: dict[str, float] = defaultdict(float)
        self.output_partitions_left = len(self.partitions_of.get(local_address, ()))
        self.transferred = False
        self.total_recvd = 0
        self.start_time = time.time()
        self._exception: Exception | None = None
        self._closed_event = asyncio.Event()

    def __repr__(self) -> str:
        return f"<Shuffle id: {self.id} on {self.local_address}>"

    @contextlib.contextmanager
    def time(self, name: str) -> Iterator[None]:
        start = time.time()
        yield
        stop = time.time()
        self.diagnostics[name] += stop - start

    async def barrier(self) -> None:
        self.raise_if_closed()
        # TODO: Consider broadcast pinging once when the shuffle starts to warm
        # up the comm pool on scheduler side
        await self.broadcast(msg={"op": "shuffle_inputs_done", "shuffle_id": self.id})

    async def send(self, address: str, shards: list[bytes]) -> None:
        self.raise_if_closed()
        return await self.rpc(address).shuffle_receive(
            data=to_serialize(shards),
            shuffle_id=self.id,
        )

    async def offload(self, func: Callable[..., T], *args: Any) -> T:
        self.raise_if_closed()
        with self.time("cpu"):
            return await asyncio.get_running_loop().run_in_executor(
                self.executor,
                func,
                *args,
            )

    def heartbeat(self) -> dict[str, Any]:
        comm_heartbeat = self._comm_buffer.heartbeat()
        comm_heartbeat["read"] = self.total_recvd
        return {
            "disk": self._disk_buffer.heartbeat(),
            "comm": comm_heartbeat,
            "diagnostics": self.diagnostics,
            "start": self.start_time,
        }

    async def receive(self, data: list[bytes]) -> None:
        await self._receive(data)

    async def _receive(self, data: list[bytes]) -> None:
        self.raise_if_closed()

        try:
            self.total_recvd += sum(map(len, data))
            groups = await self.offload(self._repartition_buffers, data)
            await self._write_to_disk(groups)
        except Exception as e:
            self._exception = e
            raise

    def _repartition_buffers(self, data: list[bytes]) -> dict[str, list[bytes]]:
        table = list_of_buffers_to_table(data, self.schema)
        groups = split_by_partition(table, self.column)
        assert len(table) == sum(map(len, groups.values()))
        del data
        return {
            k: [batch.serialize() for batch in v.to_batches()]
            for k, v in groups.items()
        }

    async def _write_to_disk(self, data: dict[str, list[bytes]]) -> None:
        self.raise_if_closed()
        await self._disk_buffer.write(data)

    def raise_if_closed(self) -> None:
        if self.closed:
            if self._exception:
                raise self._exception
            raise ShuffleClosedError(
                f"Shuffle {self.id} has been closed on {self.local_address}"
            )

    async def add_partition(self, data: pd.DataFrame) -> None:
        self.raise_if_closed()
        if self.transferred:
            raise RuntimeError(f"Cannot add more partitions to shuffle {self}")

        def _() -> dict[str, list[bytes]]:
            out = split_by_worker(
                data,
                self.column,
                self.worker_for,
            )
            out = {
                k: [b.serialize().to_pybytes() for b in t.to_batches()]
                for k, t in out.items()
            }
            return out

        out = await self.offload(_)
        await self._write_to_comm(out)

    async def _write_to_comm(self, data: dict[str, list[bytes]]) -> None:
        self.raise_if_closed()
        await self._comm_buffer.write(data)

    async def get_output_partition(self, i: int) -> pd.DataFrame:
        self.raise_if_closed()
        assert self.transferred, "`get_output_partition` called before barrier task"

        assert self.worker_for[i] == self.local_address, (
            f"Output partition {i} belongs on {self.worker_for[i]}, "
            f"not {self.local_address}. "
        )
        # ^ NOTE: this check isn't necessary, just a nice validation to prevent incorrect
        # data in the case something has gone very wrong

        assert (
            self.output_partitions_left > 0
        ), f"No outputs remaining, but requested output partition {i} on {self.local_address}."
        await self.flush_receive()
        try:
            df = self._read_from_disk(i)
            with self.time("cpu"):
                out = df.to_pandas()
        except KeyError:
            out = self.schema.empty_table().to_pandas()
        self.output_partitions_left -= 1
        return out

    def _read_from_disk(self, id: int | str) -> pa.Table:
        self.raise_if_closed()
        return self._disk_buffer.read(id)

    async def inputs_done(self) -> None:
        self.raise_if_closed()
        assert not self.transferred, "`inputs_done` called multiple times"
        self.transferred = True
        await self._flush_comm()
        try:
            self._comm_buffer.raise_on_exception()
        except Exception as e:
            self._exception = e
            raise

    async def _flush_comm(self) -> None:
        self.raise_if_closed()
        await self._comm_buffer.flush()

    def done(self) -> bool:
        return self.transferred and self.output_partitions_left == 0

    async def flush_receive(self) -> None:
        self.raise_if_closed()
        await self._disk_buffer.flush()

    async def close(self) -> None:
        if self.closed:
            await self._closed_event.wait()
            return

        self.closed = True
        await self._comm_buffer.close()
        await self._disk_buffer.close()
        try:
            self.executor.shutdown(cancel_futures=True)
        except Exception:
            self.executor.shutdown()
        self._closed_event.set()

    def fail(self, exception: Exception) -> None:
        if not self.closed:
            self._exception = exception


class ShuffleWorkerExtension:
    """Interface between a Worker and a Shuffle.

    This extension is responsible for

    - Lifecycle of Shuffle instances
    - ensuring connectivity between remote shuffle instances
    - ensuring connectivity and integration with the scheduler
    - routing concurrent calls to the appropriate `Shuffle` based on its `ShuffleID`
    - collecting instrumentation of ongoing shuffles and route to scheduler/worker
    """

    worker: Worker
    shuffles: dict[ShuffleId, Shuffle]
    memory_limiter_comms: ResourceLimiter
    memory_limiter_disk: ResourceLimiter
    closed: bool

    def __init__(self, worker: Worker) -> None:
        # Attach to worker
        worker.handlers["shuffle_receive"] = self.shuffle_receive
        worker.handlers["shuffle_inputs_done"] = self.shuffle_inputs_done
        worker.handlers["shuffle_fail"] = self.shuffle_fail
        worker.stream_handlers["shuffle-fail"] = self.shuffle_fail
        worker.extensions["shuffle"] = self

        # Initialize
        self.worker = worker
        self.shuffles = {}
        self.memory_limiter_comms = ResourceLimiter(parse_bytes("100 MiB"))
        self.memory_limiter_disk = ResourceLimiter(parse_bytes("1 GiB"))
        self.closed = False

    # Handlers
    ##########
    # NOTE: handlers are not threadsafe, but they're called from async comms, so that's okay

    def heartbeat(self) -> dict:
        return {id: shuffle.heartbeat() for id, shuffle in self.shuffles.items()}

    async def shuffle_receive(
        self,
        shuffle_id: ShuffleId,
        data: list[bytes],
    ) -> None:
        """
        Handler: Receive an incoming shard of data from a peer worker.
        Using an unknown ``shuffle_id`` is an error.
        """
        shuffle = await self._get_shuffle(shuffle_id)
        await shuffle.receive(data)

    async def shuffle_inputs_done(self, shuffle_id: ShuffleId) -> None:
        """
        Handler: Inform the extension that all input partitions have been handed off to extensions.
        Using an unknown ``shuffle_id`` is an error.
        """
        with log_errors():
            shuffle = await self._get_shuffle(shuffle_id)
            await shuffle.inputs_done()
            if shuffle.done():
                # If the shuffle has no output partitions, remove it now;
                # `get_output_partition` will never be called.
                # This happens when there are fewer output partitions than workers.
                assert shuffle._disk_buffer.empty
                logger.info(f"Shuffle inputs done {shuffle}")
                await self._register_complete(shuffle)
                del self.shuffles[shuffle_id]

    async def shuffle_fail(self, shuffle_id: ShuffleId, message: str) -> None:
        try:
            shuffle = self.shuffles[shuffle_id]
        except KeyError:
            return
        exception = RuntimeError(message)
        shuffle.fail(exception)
        await shuffle.close()
        del self.shuffles[shuffle_id]

    def add_partition(
        self,
        data: pd.DataFrame,
        shuffle_id: ShuffleId,
        npartitions: int,
        column: str,
    ) -> None:
        shuffle = self.get_shuffle(
            shuffle_id, empty=data, npartitions=npartitions, column=column
        )
        sync(self.worker.loop, shuffle.add_partition, data=data)

    async def _barrier(self, shuffle_id: ShuffleId) -> None:
        """
        Task: Note that the barrier task has been reached (`add_partition` called for all input partitions)

        Using an unknown ``shuffle_id`` is an error. Calling this before all partitions have been
        added is undefined.
        """
        # Tell all peers that we've reached the barrier
        # Note that this will call `shuffle_inputs_done` on our own worker as well
        shuffle = await self._get_shuffle(shuffle_id)
        await shuffle.barrier()

    async def _register_complete(self, shuffle: Shuffle) -> None:
        await shuffle.close()
        # All the relevant work has already succeeded if we reached this point,
        # so we do not need to check if the extension is closed.
        await self.worker.scheduler.shuffle_register_complete(
            id=shuffle.id,
            worker=self.worker.address,
        )

    @overload
    async def _get_shuffle(
        self,
        shuffle_id: ShuffleId,
    ) -> Shuffle:
        ...

    @overload
    async def _get_shuffle(
        self,
        shuffle_id: ShuffleId,
        empty: pd.DataFrame,
        column: str,
        npartitions: int,
    ) -> Shuffle:
        ...

    async def _get_shuffle(
        self,
        shuffle_id: ShuffleId,
        empty: pd.DataFrame | None = None,
        column: str | None = None,
        npartitions: int | None = None,
    ) -> Shuffle:
        "Get a shuffle by ID; raise ValueError if it's not registered."
        import pyarrow as pa

        try:
            shuffle = self.shuffles[shuffle_id]
        except KeyError:
            try:
                result = await self.worker.scheduler.shuffle_get(
                    id=shuffle_id,
                    schema=pa.Schema.from_pandas(empty).serialize().to_pybytes()
                    if empty is not None
                    else None,
                    npartitions=npartitions,
                    column=column,
                    worker=self.worker.address,
                )
                if result["status"] == "ERROR":
                    raise RuntimeError(result["message"])
                assert result["status"] == "OK"
            except KeyError:
                # Even the scheduler doesn't know about this shuffle
                # Let's hand this back to the scheduler and let it figure
                # things out
                logger.info(
                    "Worker Shuffle unable to get information from scheduler, rescheduling"
                )
                from distributed.worker import Reschedule

                raise Reschedule()
            else:
                if self.closed:
                    raise ShuffleClosedError(
                        f"{self.__class__.__name__} already closed on {self.worker.address}"
                    )
                if shuffle_id not in self.shuffles:
                    shuffle = Shuffle(
                        column=result["column"],
                        worker_for=result["worker_for"],
                        output_workers=result["output_workers"],
                        schema=deserialize_schema(result["schema"]),
                        id=shuffle_id,
                        directory=os.path.join(
                            self.worker.local_directory, f"shuffle-{shuffle_id}"
                        ),
                        nthreads=self.worker.state.nthreads,
                        local_address=self.worker.address,
                        rpc=self.worker.rpc,
                        broadcast=functools.partial(
                            self._broadcast_to_participants, shuffle_id
                        ),
                        memory_limiter_disk=self.memory_limiter_disk,
                        memory_limiter_comms=self.memory_limiter_comms,
                    )
                    self.shuffles[shuffle_id] = shuffle
                return self.shuffles[shuffle_id]
        else:
            if shuffle._exception:
                raise shuffle._exception
            return shuffle

    async def _broadcast_to_participants(self, id: ShuffleId, msg: dict) -> dict:
        participating_workers = (
            await self.worker.scheduler.shuffle_get_participating_workers(id=id)
        )
        return await self.worker.scheduler.broadcast(
            msg=msg, workers=participating_workers
        )

    async def close(self) -> None:
        assert not self.closed

        self.closed = True
        while self.shuffles:
            _, shuffle = self.shuffles.popitem()
            await shuffle.close()

    #############################
    # Methods for worker thread #
    #############################

    def barrier(self, shuffle_id: ShuffleId) -> None:
        sync(self.worker.loop, self._barrier, shuffle_id)

    @overload
    def get_shuffle(
        self,
        shuffle_id: ShuffleId,
        empty: pd.DataFrame,
        column: str,
        npartitions: int,
    ) -> Shuffle:
        ...

    @overload
    def get_shuffle(
        self,
        shuffle_id: ShuffleId,
    ) -> Shuffle:
        ...

    def get_shuffle(
        self,
        shuffle_id: ShuffleId,
        empty: pd.DataFrame | None = None,
        column: str | None = None,
        npartitions: int | None = None,
    ) -> Shuffle:
        return sync(
            self.worker.loop,
            self._get_shuffle,
            shuffle_id,
            empty,
            column,
            npartitions,
        )

    def get_output_partition(
        self, shuffle_id: ShuffleId, output_partition: int
    ) -> pd.DataFrame:
        """
        Task: Retrieve a shuffled output partition from the ShuffleExtension.

        Calling this for a ``shuffle_id`` which is unknown or incomplete is an error.
        """
        shuffle = self.get_shuffle(shuffle_id)
        output = sync(self.worker.loop, shuffle.get_output_partition, output_partition)
        # key missing if another thread got to it first
        if shuffle.done() and shuffle_id in self.shuffles:
            shuffle = self.shuffles.pop(shuffle_id)
            sync(self.worker.loop, self._register_complete, shuffle)
        return output


def split_by_worker(
    df: pd.DataFrame,
    column: str,
    worker_for: pd.Series,
) -> dict[Any, pa.Table]:
    """
    Split data into many arrow batches, partitioned by destination worker
    """
    import numpy as np
    import pyarrow as pa

    df = df.merge(
        right=worker_for.cat.codes.rename("_worker"),
        left_on=column,
        right_index=True,
        how="inner",
    )
    nrows = len(df)
    if not nrows:
        return {}
    # assert len(df) == nrows  # Not true if some outputs aren't wanted
    # FIXME: If we do not preserve the index something is corrupting the
    # bytestream such that it cannot be deserialized anymore
    t = pa.Table.from_pandas(df, preserve_index=True)
    t = t.sort_by("_worker")
    codes = np.asarray(t.select(["_worker"]))[0]
    t = t.drop(["_worker"])
    del df

    splits = np.where(codes[1:] != codes[:-1])[0] + 1
    splits = np.concatenate([[0], splits])

    shards = [
        t.slice(offset=a, length=b - a) for a, b in toolz.sliding_window(2, splits)
    ]
    shards.append(t.slice(offset=splits[-1], length=None))

    unique_codes = codes[splits]
    out = {
        # FIXME https://github.com/pandas-dev/pandas-stubs/issues/43
        worker_for.cat.categories[code]: shard
        for code, shard in zip(unique_codes, shards)
    }
    assert sum(map(len, out.values())) == nrows
    return out


def split_by_partition(t: pa.Table, column: str) -> dict[Any, pa.Table]:
    """
    Split data into many arrow batches, partitioned by final partition
    """
    import numpy as np

    partitions = t.select([column]).to_pandas()[column].unique()
    partitions.sort()
    t = t.sort_by(column)

    partition = np.asarray(t.select([column]))[0]
    splits = np.where(partition[1:] != partition[:-1])[0] + 1
    splits = np.concatenate([[0], splits])

    shards = [
        t.slice(offset=a, length=b - a) for a, b in toolz.sliding_window(2, splits)
    ]
    shards.append(t.slice(offset=splits[-1], length=None))
    assert len(t) == sum(map(len, shards))
    assert len(partitions) == len(shards)
    return dict(zip(partitions, shards))