File: sqlite.go

package info (click to toggle)
golang-github-containers-image 5.28.0-4
  • links: PTS, VCS
  • area: main
  • in suites: sid
  • size: 5,104 kB
  • sloc: sh: 194; makefile: 73
file content (553 lines) | stat: -rw-r--r-- 25,669 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
// Package boltdb implements a BlobInfoCache backed by SQLite.
package sqlite

import (
	"database/sql"
	"errors"
	"fmt"
	"sync"
	"time"

	"github.com/containers/image/v5/internal/blobinfocache"
	"github.com/containers/image/v5/pkg/blobinfocache/internal/prioritize"
	"github.com/containers/image/v5/types"
	_ "github.com/mattn/go-sqlite3" // Registers the "sqlite3" backend backend for database/sql
	"github.com/opencontainers/go-digest"
	"github.com/sirupsen/logrus"
)

const (
	// NOTE: There is no versioning data inside the file; this is a “cache”, so on an incompatible format upgrade
	// we can simply start over with a different filename; update blobInfoCacheFilename.
	// That also means we don’t have to worry about co-existing readers/writers which know different versions of the schema
	// (which would require compatibility in both directions).

	// Assembled sqlite options used when opening the database.
	sqliteOptions = "?" +
		// Deal with timezone automatically.
		// go-sqlite3 always _records_ timestamps as a text: time in local time + a time zone offset.
		// _loc affects how the values are _parsed_: (which timezone is assumed for numeric timestamps or for text which does not specify an offset, or)
		// if the time zone offset matches the specified time zone, the timestamp is assumed to be in that time zone / location;
		// (otherwise an unnamed time zone carrying just a hard-coded offset, but no location / DST rules is used).
		"_loc=auto" +
		// Force an fsync after each transaction (https://www.sqlite.org/pragma.html#pragma_synchronous).
		"&_sync=FULL" +
		// Allow foreign keys (https://www.sqlite.org/pragma.html#pragma_foreign_keys).
		// We don’t currently use any foreign keys, but this is a good choice long-term (not default in SQLite only for historical reasons).
		"&_foreign_keys=1" +
		// Use BEGIN EXCLUSIVE (https://www.sqlite.org/lang_transaction.html);
		// i.e. obtain a write lock for _all_ transactions at the transaction start (never use a read lock,
		// never upgrade from a read to a write lock - that can fail if multiple read lock owners try to do that simultaneously).
		//
		// This, together with go-sqlite3’s default for _busy_timeout=5000, means that we should never see a “database is locked” error,
		// the database should block on the exclusive lock when starting a transaction, and the problematic case of two simultaneous
		// holders of a read lock trying to upgrade to a write lock (and one necessarily failing) is prevented.
		// Compare https://github.com/mattn/go-sqlite3/issues/274 .
		//
		// Ideally the BEGIN / BEGIN EXCLUSIVE decision could be made per-transaction, compare https://github.com/mattn/go-sqlite3/pull/1167
		// or https://github.com/mattn/go-sqlite3/issues/400 .
		// The currently-proposed  workaround is to create two different SQL “databases” (= connection pools) with different _txlock settings,
		// which seems rather wasteful.
		"&_txlock=exclusive"
)

// cache is a BlobInfoCache implementation which uses a SQLite file at the specified path.
type cache struct {
	path string

	// The database/sql package says “It is rarely necessary to close a DB.”, and steers towards a long-term *sql.DB connection pool.
	// That’s probably very applicable for database-backed services, where the database is the primary data store. That’s not necessarily
	// the case for callers of c/image, where image operations might be a small proportion of hte total runtime, and the cache is fairly
	// incidental even to the image operations. It’s also hard for us to use that model, because the public BlobInfoCache object doesn’t have
	// a Close method, so creating a lot of single-use caches could leak data.
	//
	// Instead, the private BlobInfoCache2 interface provides Open/Close methods, and they are called by c/image/copy.Image.
	// This amortizes the cost of opening/closing the SQLite state over a single image copy, while keeping no long-term resources open.
	// Some rough benchmarks in https://github.com/containers/image/pull/2092 suggest relative costs on the order of "25" for a single
	// *sql.DB left open long-term, "27" for a *sql.DB open for a single image copy, and "40" for opening/closing a *sql.DB for every
	// single transaction; so the Open/Close per image copy seems a reasonable compromise (especially compared to the previous implementation,
	// somewhere around "700").

	lock sync.Mutex
	// The following fields can only be accessed with lock held.
	refCount int     // number of outstanding Open() calls
	db       *sql.DB // nil if not set (may happen even if refCount > 0 on errors)
}

// New returns BlobInfoCache implementation which uses a SQLite file at path.
//
// Most users should call blobinfocache.DefaultCache instead.
func New(path string) (types.BlobInfoCache, error) {
	return new2(path)
}

func new2(path string) (*cache, error) {
	db, err := rawOpen(path)
	if err != nil {
		return nil, fmt.Errorf("initializing blob info cache at %q: %w", path, err)
	}
	defer db.Close()

	// We don’t check the schema before every operation, because that would be costly
	// and because we assume schema changes will be handled by using a different path.
	if err := ensureDBHasCurrentSchema(db); err != nil {
		return nil, err
	}

	return &cache{
		path:     path,
		refCount: 0,
		db:       nil,
	}, nil
}

// rawOpen returns a new *sql.DB for path.
// The caller should arrange for it to be .Close()d.
func rawOpen(path string) (*sql.DB, error) {
	// This exists to centralize the use of sqliteOptions.
	return sql.Open("sqlite3", path+sqliteOptions)
}

// Open() sets up the cache for future accesses, potentially acquiring costly state. Each Open() must be paired with a Close().
// Note that public callers may call the types.BlobInfoCache operations without Open()/Close().
func (sqc *cache) Open() {
	sqc.lock.Lock()
	defer sqc.lock.Unlock()

	if sqc.refCount == 0 {
		db, err := rawOpen(sqc.path)
		if err != nil {
			logrus.Warnf("Error opening (previously-succesfully-opened) blob info cache at %q: %v", sqc.path, err)
			db = nil // But still increase sqc.refCount, because a .Close() will happen
		}
		sqc.db = db
	}
	sqc.refCount++
}

// Close destroys state created by Open().
func (sqc *cache) Close() {
	sqc.lock.Lock()
	defer sqc.lock.Unlock()

	switch sqc.refCount {
	case 0:
		logrus.Errorf("internal error using pkg/blobinfocache/sqlite.cache: Close() without a matching Open()")
		return
	case 1:
		if sqc.db != nil {
			sqc.db.Close()
			sqc.db = nil
		}
	}
	sqc.refCount--
}

type void struct{} // So that we don’t have to write struct{}{} all over the place

// transaction calls fn within a read-write transaction in sqc.
func transaction[T any](sqc *cache, fn func(tx *sql.Tx) (T, error)) (T, error) {
	db, closeDB, err := func() (*sql.DB, func(), error) { // A scope for defer
		sqc.lock.Lock()
		defer sqc.lock.Unlock()

		if sqc.db != nil {
			return sqc.db, func() {}, nil
		}
		db, err := rawOpen(sqc.path)
		if err != nil {
			return nil, nil, fmt.Errorf("opening blob info cache at %q: %w", sqc.path, err)
		}
		return db, func() { db.Close() }, nil
	}()
	if err != nil {
		var zeroRes T // A zero value of T
		return zeroRes, err
	}
	defer closeDB()

	return dbTransaction(db, fn)
}

// dbTransaction calls fn within a read-write transaction in db.
func dbTransaction[T any](db *sql.DB, fn func(tx *sql.Tx) (T, error)) (T, error) {
	// Ideally we should be able to distinguish between read-only and read-write transctions, see the _txlock=exclusive dicussion.

	var zeroRes T // A zero value of T

	tx, err := db.Begin()
	if err != nil {
		return zeroRes, fmt.Errorf("beginning transaction: %w", err)
	}
	succeeded := false
	defer func() {
		if !succeeded {
			if err := tx.Rollback(); err != nil {
				logrus.Errorf("Rolling back transaction: %v", err)
			}
		}
	}()

	res, err := fn(tx)
	if err != nil {
		return zeroRes, err
	}
	if err := tx.Commit(); err != nil {
		return zeroRes, fmt.Errorf("committing transaction: %w", err)
	}

	succeeded = true
	return res, nil
}

// querySingleValue executes a SELECT which is expected to return at most one row with a single column.
// It returns (value, true, nil) on success, or (value, false, nil) if no row was returned.
func querySingleValue[T any](tx *sql.Tx, query string, params ...any) (T, bool, error) {
	var value T
	if err := tx.QueryRow(query, params...).Scan(&value); err != nil {
		var zeroValue T // A zero value of T
		if errors.Is(err, sql.ErrNoRows) {
			return zeroValue, false, nil
		}
		return zeroValue, false, err
	}
	return value, true, nil
}

// ensureDBHasCurrentSchema adds the necessary tables and indices to a database.
// This is typically used when creating a previously-nonexistent database.
// We don’t really anticipate schema migrations; with c/image usually vendored, not using
// shared libraries, migrating a schema on an existing database would affect old-version users.
// Instead, schema changes are likely to be implemented by using a different cache file name,
// and leaving existing caches around for old users.
func ensureDBHasCurrentSchema(db *sql.DB) error {
	// Considered schema design alternatives:
	//
	// (Overall, considering the overall network latency and disk I/O costs of many-megabyte layer pulls which are happening while referring
	// to the blob info cache, it seems reasonable to prioritize readability over microoptimization of this database.)
	//
	// * This schema uses the text representation of digests.
	//
	//   We use the fairly wasteful text with hexadecimal digits because digest.Digest does not define a binary representation;
	//   and the way digest.Digest.Hex() is deprecated in favor of digest.Digest.Encoded(), and the way digest.Algorithm
	//   is documented to “define the string encoding” suggests that assuming a hexadecimal representation and turning that
	//   into binary ourselves is not a good idea in general; we would have to special-case the currently-known algorithm
	//   — and that would require us to implement two code paths, one of them basically never exercised / never tested.
	//
	// * There are two separate items for recording the uncompressed digest and digest compressors.
	//   Alternatively, we could have a single "digest facts" table with NULLable columns.
	//
	//   The way the BlobInfoCache API works, we are only going to write one value at a time, so
	//   sharing a table would not be any more efficient for writes (same number of lookups, larger row tuples).
	//   Reads in candidateLocations would not be more efficient either, the searches in DigestCompressors and DigestUncompressedPairs
	//   do not coincide (we want a compressor for every candidate, but the uncompressed digest only for the primary digest; and then
	//   we search in DigestUncompressedPairs by uncompressed digest, not by the primary key).
	//
	//   Also, using separate items allows the single-item writes to be done using a simple INSERT OR REPLACE, instead of having to
	//   do a more verbose ON CONFLICT(…) DO UPDATE SET … = ….
	//
	// * Joins (the two that exist in appendReplacementCandidates) are based on the text representation of digests.
	//
	//   Using integer primary keys might make the joins themselves a bit more efficient, but then we would need to involve an extra
	//   join to translate from/to the user-provided digests anyway. If anything, that extra join (potentialy more btree lookups)
	//   is probably costlier than comparing a few more bytes of data.
	//
	//   Perhaps more importantly, storing digest texts directly makes the database dumps much easier to read for humans without
	//   having to do extra steps to decode the integers into digest values (either by running sqlite commands with joins, or mentally).
	//
	items := []struct{ itemName, command string }{
		{
			"DigestUncompressedPairs",
			`CREATE TABLE IF NOT EXISTS DigestUncompressedPairs(` +
				// index implied by PRIMARY KEY
				`anyDigest			TEXT PRIMARY KEY NOT NULL,` +
				// DigestUncompressedPairs_index_uncompressedDigest
				`uncompressedDigest	TEXT NOT NULL
			)`,
		},
		{
			"DigestUncompressedPairs_index_uncompressedDigest",
			`CREATE INDEX IF NOT EXISTS DigestUncompressedPairs_index_uncompressedDigest ON DigestUncompressedPairs(uncompressedDigest)`,
		},
		{
			"DigestCompressors",
			`CREATE TABLE IF NOT EXISTS DigestCompressors(` +
				// index implied by PRIMARY KEY
				`digest		TEXT PRIMARY KEY NOT NULL,` +
				// May include blobinfocache.Uncompressed (not blobinfocache.UnknownCompression).
				`compressor	TEXT NOT NULL
			)`,
		},
		{
			"KnownLocations",
			`CREATE TABLE IF NOT EXISTS KnownLocations(
				transport	TEXT NOT NULL,
				scope 		TEXT NOT NULL,
				digest		TEXT NOT NULL,
				location	TEXT NOT NULL,` +
				// TIMESTAMP is parsed by SQLITE as a NUMERIC affinity, but go-sqlite3 stores text in the (Go formatting semantics)
				// format "2006-01-02 15:04:05.999999999-07:00".
				// See also the _loc option in the sql.Open data source name.
				`time		TIMESTAMP NOT NULL,` +
				// Implies an index.
				// We also search by (transport, scope, digest), that doesn’t need an extra index
				// because it is a prefix of the implied primary-key index.
				`PRIMARY KEY (transport, scope, digest, location)
			)`,
		},
	}

	_, err := dbTransaction(db, func(tx *sql.Tx) (void, error) {
		// If the the last-created item exists, assume nothing needs to be done.
		lastItemName := items[len(items)-1].itemName
		_, found, err := querySingleValue[int](tx, "SELECT 1 FROM sqlite_schema WHERE name=?", lastItemName)
		if err != nil {
			return void{}, fmt.Errorf("checking if SQLite schema item %q exists: %w", lastItemName, err)
		}
		if !found {
			// Item does not exist, assuming a fresh database.
			for _, i := range items {
				if _, err := tx.Exec(i.command); err != nil {
					return void{}, fmt.Errorf("creating item %s: %w", i.itemName, err)
				}
			}
		}
		return void{}, nil
	})
	return err
}

// uncompressedDigest implements types.BlobInfoCache.UncompressedDigest within a transaction.
func (sqc *cache) uncompressedDigest(tx *sql.Tx, anyDigest digest.Digest) (digest.Digest, error) {
	uncompressedString, found, err := querySingleValue[string](tx, "SELECT uncompressedDigest FROM DigestUncompressedPairs WHERE anyDigest = ?", anyDigest.String())
	if err != nil {
		return "", err
	}
	if found {
		d, err := digest.Parse(uncompressedString)
		if err != nil {
			return "", err
		}
		return d, nil

	}
	// A record as uncompressedDigest implies that anyDigest must already refer to an uncompressed digest.
	// This way we don't have to waste storage space with trivial (uncompressed, uncompressed) mappings
	// when we already record a (compressed, uncompressed) pair.
	_, found, err = querySingleValue[int](tx, "SELECT 1 FROM DigestUncompressedPairs WHERE uncompressedDigest = ?", anyDigest.String())
	if err != nil {
		return "", err
	}
	if found {
		return anyDigest, nil
	}
	return "", nil
}

// UncompressedDigest returns an uncompressed digest corresponding to anyDigest.
// May return anyDigest if it is known to be uncompressed.
// Returns "" if nothing is known about the digest (it may be compressed or uncompressed).
func (sqc *cache) UncompressedDigest(anyDigest digest.Digest) digest.Digest {
	res, err := transaction(sqc, func(tx *sql.Tx) (digest.Digest, error) {
		return sqc.uncompressedDigest(tx, anyDigest)
	})
	if err != nil {
		return "" // FIXME? Log err (but throttle the log volume on repeated accesses)?
	}
	return res
}

// RecordDigestUncompressedPair records that the uncompressed version of anyDigest is uncompressed.
// It’s allowed for anyDigest == uncompressed.
// WARNING: Only call this for LOCALLY VERIFIED data; don’t record a digest pair just because some remote author claims so (e.g.
// because a manifest/config pair exists); otherwise the cache could be poisoned and allow substituting unexpected blobs.
// (Eventually, the DiffIDs in image config could detect the substitution, but that may be too late, and not all image formats contain that data.)
func (sqc *cache) RecordDigestUncompressedPair(anyDigest digest.Digest, uncompressed digest.Digest) {
	_, _ = transaction(sqc, func(tx *sql.Tx) (void, error) {
		previousString, gotPrevious, err := querySingleValue[string](tx, "SELECT uncompressedDigest FROM DigestUncompressedPairs WHERE anyDigest = ?", anyDigest.String())
		if err != nil {
			return void{}, fmt.Errorf("looking for uncompressed digest for %q", anyDigest)
		}
		if gotPrevious {
			previous, err := digest.Parse(previousString)
			if err != nil {
				return void{}, err
			}
			if previous != uncompressed {
				logrus.Warnf("Uncompressed digest for blob %s previously recorded as %s, now %s", anyDigest, previous, uncompressed)
			}
		}
		if _, err := tx.Exec("INSERT OR REPLACE INTO DigestUncompressedPairs(anyDigest, uncompressedDigest) VALUES (?, ?)",
			anyDigest.String(), uncompressed.String()); err != nil {
			return void{}, fmt.Errorf("recording uncompressed digest %q for %q: %w", uncompressed, anyDigest, err)
		}
		return void{}, nil
	}) // FIXME? Log error (but throttle the log volume on repeated accesses)?
}

// RecordKnownLocation records that a blob with the specified digest exists within the specified (transport, scope) scope,
// and can be reused given the opaque location data.
func (sqc *cache) RecordKnownLocation(transport types.ImageTransport, scope types.BICTransportScope, digest digest.Digest, location types.BICLocationReference) {
	_, _ = transaction(sqc, func(tx *sql.Tx) (void, error) {
		if _, err := tx.Exec("INSERT OR REPLACE INTO KnownLocations(transport, scope, digest, location, time) VALUES (?, ?, ?, ?, ?)",
			transport.Name(), scope.Opaque, digest.String(), location.Opaque, time.Now()); err != nil { // Possibly overwriting an older entry.
			return void{}, fmt.Errorf("recording known location %q for (%q, %q, %q): %w",
				location.Opaque, transport.Name(), scope.Opaque, digest.String(), err)
		}
		return void{}, nil
	}) // FIXME? Log error (but throttle the log volume on repeated accesses)?
}

// RecordDigestCompressorName records a compressor for the blob with the specified digest,
// or Uncompressed or UnknownCompression.
// WARNING: Only call this with LOCALLY VERIFIED data; don’t record a compressor for a
// digest just because some remote author claims so (e.g. because a manifest says so);
// otherwise the cache could be poisoned and cause us to make incorrect edits to type
// information in a manifest.
func (sqc *cache) RecordDigestCompressorName(anyDigest digest.Digest, compressorName string) {
	_, _ = transaction(sqc, func(tx *sql.Tx) (void, error) {
		previous, gotPrevious, err := querySingleValue[string](tx, "SELECT compressor FROM DigestCompressors WHERE digest = ?", anyDigest.String())
		if err != nil {
			return void{}, fmt.Errorf("looking for compressor of for %q", anyDigest)
		}
		if gotPrevious && previous != compressorName {
			logrus.Warnf("Compressor for blob with digest %s previously recorded as %s, now %s", anyDigest, previous, compressorName)
		}
		if compressorName == blobinfocache.UnknownCompression {
			if _, err := tx.Exec("DELETE FROM DigestCompressors WHERE digest = ?", anyDigest.String()); err != nil {
				return void{}, fmt.Errorf("deleting compressor for digest %q: %w", anyDigest, err)
			}
		} else {
			if _, err := tx.Exec("INSERT OR REPLACE INTO DigestCompressors(digest, compressor) VALUES (?, ?)",
				anyDigest.String(), compressorName); err != nil {
				return void{}, fmt.Errorf("recording compressor %q for %q: %w", compressorName, anyDigest, err)
			}
		}
		return void{}, nil
	}) // FIXME? Log error (but throttle the log volume on repeated accesses)?
}

// appendReplacementCandidates creates prioritize.CandidateWithTime values for (transport, scope, digest), and returns the result of appending them to candidates.
func (sqc *cache) appendReplacementCandidates(candidates []prioritize.CandidateWithTime, tx *sql.Tx, transport types.ImageTransport, scope types.BICTransportScope, digest digest.Digest, requireCompressionInfo bool) ([]prioritize.CandidateWithTime, error) {
	var rows *sql.Rows
	var err error
	if requireCompressionInfo {
		rows, err = tx.Query("SELECT location, time, compressor FROM KnownLocations JOIN DigestCompressors "+
			"ON KnownLocations.digest = DigestCompressors.digest "+
			"WHERE transport = ? AND scope = ? AND KnownLocations.digest = ?",
			transport.Name(), scope.Opaque, digest.String())
	} else {
		rows, err = tx.Query("SELECT location, time, IFNULL(compressor, ?) FROM KnownLocations "+
			"LEFT JOIN DigestCompressors ON KnownLocations.digest = DigestCompressors.digest "+
			"WHERE transport = ? AND scope = ? AND KnownLocations.digest = ?",
			blobinfocache.UnknownCompression,
			transport.Name(), scope.Opaque, digest.String())
	}
	if err != nil {
		return nil, fmt.Errorf("looking up candidate locations: %w", err)
	}
	defer rows.Close()

	for rows.Next() {
		var location string
		var time time.Time
		var compressorName string
		if err := rows.Scan(&location, &time, &compressorName); err != nil {
			return nil, fmt.Errorf("scanning candidate: %w", err)
		}
		candidates = append(candidates, prioritize.CandidateWithTime{
			Candidate: blobinfocache.BICReplacementCandidate2{
				Digest:         digest,
				CompressorName: compressorName,
				Location:       types.BICLocationReference{Opaque: location},
			},
			LastSeen: time,
		})
	}
	if err := rows.Err(); err != nil {
		return nil, fmt.Errorf("iterating through locations: %w", err)
	}
	return candidates, nil
}

// CandidateLocations2 returns a prioritized, limited, number of blobs and their locations
// that could possibly be reused within the specified (transport scope) (if they still
// exist, which is not guaranteed).
//
// If !canSubstitute, the returned cadidates will match the submitted digest exactly; if
// canSubstitute, data from previous RecordDigestUncompressedPair calls is used to also look
// up variants of the blob which have the same uncompressed digest.
//
// The CompressorName fields in returned data must never be UnknownCompression.
func (sqc *cache) CandidateLocations2(transport types.ImageTransport, scope types.BICTransportScope, digest digest.Digest, canSubstitute bool) []blobinfocache.BICReplacementCandidate2 {
	return sqc.candidateLocations(transport, scope, digest, canSubstitute, true)
}

func (sqc *cache) candidateLocations(transport types.ImageTransport, scope types.BICTransportScope, primaryDigest digest.Digest, canSubstitute, requireCompressionInfo bool) []blobinfocache.BICReplacementCandidate2 {
	var uncompressedDigest digest.Digest // = ""
	res, err := transaction(sqc, func(tx *sql.Tx) ([]prioritize.CandidateWithTime, error) {
		res := []prioritize.CandidateWithTime{}
		res, err := sqc.appendReplacementCandidates(res, tx, transport, scope, primaryDigest, requireCompressionInfo)
		if err != nil {
			return nil, err
		}
		if canSubstitute {
			uncompressedDigest, err = sqc.uncompressedDigest(tx, primaryDigest)
			if err != nil {
				return nil, err
			}

			// FIXME? We could integrate this with appendReplacementCandidates into a single join instead of N+1 queries.
			// (In the extreme, we could turn _everything_ this function does into a single query.
			// And going even further, even DestructivelyPrioritizeReplacementCandidates could be turned into SQL.)
			// For now, we prioritize simplicity, and sharing both code and implementation structure with the other cache implementations.
			rows, err := tx.Query("SELECT anyDigest FROM DigestUncompressedPairs WHERE uncompressedDigest = ?", uncompressedDigest.String())
			if err != nil {
				return nil, fmt.Errorf("querying for other digests: %w", err)
			}
			defer rows.Close()
			for rows.Next() {
				var otherDigestString string
				if err := rows.Scan(&otherDigestString); err != nil {
					return nil, fmt.Errorf("scanning other digest: %w", err)
				}
				otherDigest, err := digest.Parse(otherDigestString)
				if err != nil {
					return nil, err
				}
				if otherDigest != primaryDigest && otherDigest != uncompressedDigest {
					res, err = sqc.appendReplacementCandidates(res, tx, transport, scope, otherDigest, requireCompressionInfo)
					if err != nil {
						return nil, err
					}
				}
			}
			if err := rows.Err(); err != nil {
				return nil, fmt.Errorf("iterating through other digests: %w", err)
			}

			if uncompressedDigest != primaryDigest {
				res, err = sqc.appendReplacementCandidates(res, tx, transport, scope, uncompressedDigest, requireCompressionInfo)
				if err != nil {
					return nil, err
				}
			}
		}
		return res, nil
	})
	if err != nil {
		return []blobinfocache.BICReplacementCandidate2{} // FIXME? Log err (but throttle the log volume on repeated accesses)?
	}
	return prioritize.DestructivelyPrioritizeReplacementCandidates(res, primaryDigest, uncompressedDigest)

}

// CandidateLocations returns a prioritized, limited, number of blobs and their locations that could possibly be reused
// within the specified (transport scope) (if they still exist, which is not guaranteed).
//
// If !canSubstitute, the returned candidates will match the submitted digest exactly; if canSubstitute,
// data from previous RecordDigestUncompressedPair calls is used to also look up variants of the blob which have the same
// uncompressed digest.
func (sqc *cache) CandidateLocations(transport types.ImageTransport, scope types.BICTransportScope, digest digest.Digest, canSubstitute bool) []types.BICReplacementCandidate {
	return blobinfocache.CandidateLocationsFromV2(sqc.candidateLocations(transport, scope, digest, canSubstitute, false))
}