| 1 | package lease |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "strings" |
| 8 | "sync/atomic" |
| 9 | "testing" |
| 10 | "time" |
| 11 | ) |
| 12 | |
| 13 | // The parallel round (offline-move speedup): copies and deletes run up to |
| 14 | // mirrorWorkers(m) at once. These tests hold the round to what the serial |
| 15 | // one promised: the same end state, receipts for exactly what landed, the |
| 16 | // first failure stopping the round, and no skip a receipt cannot prove. |
| 17 | // Run them under -race: the round's own state is shared by its workers. |
| 18 | |
| 19 | // fillReplica stores n objects under instances/h/: every third one |
| 20 | // multipart (the copy lands with a new ETag, so it needs a receipt), the |
| 21 | // rest single-part. Sizes differ so byte totals catch a double count. |
| 22 | func fillReplica(f *fakeMirror, n int, gen string) (keys []string, bytes int64) { |
| 23 | for i := 0; i < n; i++ { |
| 24 | key := fmt.Sprintf("instances/h/ltx/0/%s-%04d.ltx", gen, i) |
| 25 | size := int64(1000 + i) |
| 26 | if i%3 == 0 { |
| 27 | f.putMultipart("src", key, fmt.Sprintf("%s-body-%d", gen, i), size) |
| 28 | } else { |
| 29 | f.putSized("src", key, fmt.Sprintf("%s-etag-%d", gen, i), size) |
| 30 | } |
| 31 | keys = append(keys, key) |
| 32 | bytes += size |
| 33 | } |
| 34 | return keys, bytes |
| 35 | } |
| 36 | |
| 37 | // stagedBodies is moves/h/ by key relative to it, body for body. |
| 38 | func stagedBodies(f *fakeMirror) map[string]string { |
| 39 | f.mu.Lock() |
| 40 | defer f.mu.Unlock() |
| 41 | out := map[string]string{} |
| 42 | for k, v := range f.buckets["dst"] { |
| 43 | if strings.HasPrefix(k, "moves/h/") { |
| 44 | out[strings.TrimPrefix(k, "moves/h/")] = v.body |
| 45 | } |
| 46 | } |
| 47 | return out |
| 48 | } |
| 49 | |
| 50 | // sourceBodies is instances/h/ the same way, lock.json aside. |
| 51 | func sourceBodies(f *fakeMirror) map[string]string { |
| 52 | f.mu.Lock() |
| 53 | defer f.mu.Unlock() |
| 54 | out := map[string]string{} |
| 55 | for k, v := range f.buckets["src"] { |
| 56 | if rel, ok := strings.CutPrefix(k, "instances/h/"); ok && rel != "lock.json" { |
| 57 | out[rel] = v.body |
| 58 | } |
| 59 | } |
| 60 | return out |
| 61 | } |
| 62 | |
| 63 | func sameBodies(t *testing.T, got, want map[string]string) { |
| 64 | t.Helper() |
| 65 | if len(got) != len(want) { |
| 66 | t.Fatalf("staged %d objects, source has %d", len(got), len(want)) |
| 67 | } |
| 68 | for k, v := range want { |
| 69 | if got[k] != v { |
| 70 | t.Fatalf("staged %s = %q, source %q", k, got[k], v) |
| 71 | } |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | func copiesIn(ops []string) int { |
| 76 | n := 0 |
| 77 | for _, op := range ops { |
| 78 | if strings.HasPrefix(op, "copy ") { |
| 79 | n++ |
| 80 | } |
| 81 | } |
| 82 | return n |
| 83 | } |
| 84 | |
| 85 | // A parallel round ends where a serial one does: every object staged body |
| 86 | // for body, the same counts and bytes, receipts for exactly the multipart |
| 87 | // objects, and a caught-up round after it. It really runs side by side, |
| 88 | // and never more than the store allows. |
| 89 | func TestMirrorPrefixParallelMatchesSerial(t *testing.T) { |
| 90 | const n = 120 |
| 91 | var rounds [2]MirrorRound |
| 92 | var books [2]map[string]copyReceipt |
| 93 | for i, workers := range []int{1, 6} { |
| 94 | f := newFakeMirror() |
| 95 | f.workers = workers |
| 96 | _, bytes := fillReplica(f, n, "a") |
| 97 | f.put("src", "instances/h/lock.json", "renewal") |
| 98 | // Slow copies, so the workers overlap. |
| 99 | f.onCopy = func(context.Context, string) error { time.Sleep(time.Millisecond); return nil } |
| 100 | r := mirrorRound(t, f, testReceipts) |
| 101 | if r.Changed != n || r.BytesCopied != bytes || r.Deferred != 0 { |
| 102 | t.Fatalf("workers=%d: round = %+v, want %d copies of %d bytes", workers, r, n, bytes) |
| 103 | } |
| 104 | sameBodies(t, stagedBodies(f), sourceBodies(f)) |
| 105 | if f.peak > workers || (workers > 1 && f.peak < 2) { |
| 106 | t.Fatalf("workers=%d: peak concurrency %d", workers, f.peak) |
| 107 | } |
| 108 | book := readReceipts(t, f) |
| 109 | if len(book) != n/3 { |
| 110 | t.Fatalf("workers=%d: %d receipts, want one per multipart object (%d)", workers, len(book), n/3) |
| 111 | } |
| 112 | if again := mirrorRound(t, f, testReceipts); again.Changed != 0 { |
| 113 | t.Fatalf("workers=%d: caught-up round = %+v", workers, again) |
| 114 | } |
| 115 | rounds[i], books[i] = r, book |
| 116 | } |
| 117 | if rounds[0].Changed != rounds[1].Changed || rounds[0].BytesCopied != rounds[1].BytesCopied { |
| 118 | t.Fatalf("serial %+v and parallel %+v rounds differ", rounds[0], rounds[1]) |
| 119 | } |
| 120 | for k, v := range books[0] { |
| 121 | if books[1][k] != v { |
| 122 | t.Fatalf("receipt %s: serial %+v, parallel %+v", k, v, books[1][k]) |
| 123 | } |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | // The versioned mode under workers: every copied object's version lands in |
| 128 | // Versions, and the receipts carry them, so the next round names them all |
| 129 | // without copying. |
| 130 | func TestMirrorPrefixParallelVersionedRound(t *testing.T) { |
| 131 | f := &fakeVersionedMirror{fakeMirror: newFakeMirror()} |
| 132 | f.workers = 8 |
| 133 | keys, _ := fillReplica(f.fakeMirror, 64, "a") |
| 134 | r1, err := MirrorPrefix(context.Background(), f, "src", "instances/h/", "dst", "moves/h/", testReceipts) |
| 135 | if err != nil || len(r1.Versions) != len(keys) { |
| 136 | t.Fatalf("round 1 = %d versions, %v; want %d", len(r1.Versions), err, len(keys)) |
| 137 | } |
| 138 | seen := map[string]bool{} |
| 139 | for _, v := range r1.Versions { |
| 140 | if seen[v] { |
| 141 | t.Fatalf("version %s handed to two copies", v) |
| 142 | } |
| 143 | seen[v] = true |
| 144 | } |
| 145 | r2, err := MirrorPrefix(context.Background(), f, "src", "instances/h/", "dst", "moves/h/", testReceipts) |
| 146 | if err != nil || r2.Changed != 0 || len(r2.Versions) != len(keys) { |
| 147 | t.Fatalf("round 2 = %+v, %v; want nothing copied and every version named", r2, err) |
| 148 | } |
| 149 | for k, v := range r1.Versions { |
| 150 | if r2.Versions[k] != v { |
| 151 | t.Fatalf("%s: round 2 names %q, round 1 copied %q", k, r2.Versions[k], v) |
| 152 | } |
| 153 | } |
| 154 | } |
| 155 | |
| 156 | // The first failure stops the round: it is the error, copies still in |
| 157 | // flight see their context cancelled, nothing new starts, and every copy |
| 158 | // that landed keeps its receipt, so the next round copies only what did |
| 159 | // not land. |
| 160 | func TestMirrorPrefixParallelFailureMidRound(t *testing.T) { |
| 161 | f := newFakeMirror() |
| 162 | f.workers = 4 |
| 163 | // All multipart: a landed copy is skipped next round only by its receipt. |
| 164 | for i := 0; i < 40; i++ { |
| 165 | f.putMultipart("src", fmt.Sprintf("instances/h/ltx/1/%04d.ltx", i), fmt.Sprintf("b%d", i), 100) |
| 166 | } |
| 167 | bad := "instances/h/ltx/1/0010.ltx" |
| 168 | var cancelled atomic.Int32 |
| 169 | later := make(chan struct{}, 64) // a copy after the bad one has started |
| 170 | f.onCopy = func(ctx context.Context, key string) error { |
| 171 | switch { |
| 172 | case key == bad: |
| 173 | // Fail only once a later copy is in flight, so the cancel |
| 174 | // has something to reach. |
| 175 | select { |
| 176 | case <-later: |
| 177 | case <-time.After(5 * time.Second): |
| 178 | } |
| 179 | return errors.New("injected copy failure") |
| 180 | case key > bad: |
| 181 | later <- struct{}{} |
| 182 | select { |
| 183 | case <-ctx.Done(): |
| 184 | cancelled.Add(1) |
| 185 | return ctx.Err() |
| 186 | case <-time.After(5 * time.Second): |
| 187 | return errors.New("the round did not cancel a copy in flight") |
| 188 | } |
| 189 | } |
| 190 | return nil |
| 191 | } |
| 192 | r, err := MirrorPrefix(context.Background(), f, "src", "instances/h/", "dst", "moves/h/", testReceipts) |
| 193 | if err == nil || !strings.Contains(err.Error(), bad) || errors.Is(err, context.Canceled) { |
| 194 | t.Fatalf("round error = %v, want the injected failure on %s", err, bad) |
| 195 | } |
| 196 | landed := stagedBodies(f) |
| 197 | if r.Changed != len(landed) || r.BytesCopied != int64(len(landed))*100 { |
| 198 | t.Fatalf("round = %+v, but %d objects landed", r, len(landed)) |
| 199 | } |
| 200 | if len(landed) == 0 || len(landed) > 10 { |
| 201 | t.Fatalf("%d objects landed, want those before the failure only", len(landed)) |
| 202 | } |
| 203 | if cancelled.Load() == 0 { |
| 204 | t.Fatal("no copy in flight was cancelled") |
| 205 | } |
| 206 | // Nothing past the workers in flight at the failure ever started. |
| 207 | if started := copiesIn(f.ops) + int(cancelled.Load()) + 1; started > 10+1+f.workers { |
| 208 | t.Fatalf("%d copies started, want the round to stop dispatching at the failure", started) |
| 209 | } |
| 210 | book := readReceipts(t, f) |
| 211 | for rel := range landed { |
| 212 | if _, ok := book[rel]; !ok { |
| 213 | t.Fatalf("%s landed without a receipt", rel) |
| 214 | } |
| 215 | } |
| 216 | if len(book) != len(landed) { |
| 217 | t.Fatalf("%d receipts for %d landed copies", len(book), len(landed)) |
| 218 | } |
| 219 | f.onCopy = nil |
| 220 | before := len(f.ops) |
| 221 | next := mirrorRound(t, f, testReceipts) |
| 222 | if next.Changed != 40-len(landed) { |
| 223 | t.Fatalf("next round = %+v, want exactly the %d that did not land", next, 40-len(landed)) |
| 224 | } |
| 225 | for _, op := range f.ops[before:] { |
| 226 | for rel := range landed { |
| 227 | if strings.HasPrefix(op, "copy instances/h/"+rel+" ") { |
| 228 | t.Fatalf("a landed copy was recopied: %s", op) |
| 229 | } |
| 230 | } |
| 231 | } |
| 232 | sameBodies(t, stagedBodies(f), sourceBodies(f)) |
| 233 | } |
| 234 | |
| 235 | // A caller's context that ends mid-round is the round's error, never a |
| 236 | // silent partial success. |
| 237 | func TestMirrorPrefixParallelContextEnds(t *testing.T) { |
| 238 | f := newFakeMirror() |
| 239 | f.workers = 2 |
| 240 | fillReplica(f, 30, "a") |
| 241 | ctx, cancel := context.WithCancel(context.Background()) |
| 242 | var n atomic.Int32 |
| 243 | f.onCopy = func(context.Context, string) error { |
| 244 | if n.Add(1) == 3 { |
| 245 | cancel() |
| 246 | } |
| 247 | return nil |
| 248 | } |
| 249 | if _, err := MirrorPrefix(ctx, f, "src", "instances/h/", "dst", "moves/h/", testReceipts); !errors.Is(err, context.Canceled) { |
| 250 | t.Fatalf("round error = %v, want the context's", err) |
| 251 | } |
| 252 | if len(stagedBodies(f)) == 30 { |
| 253 | t.Fatal("every copy ran after the context ended") |
| 254 | } |
| 255 | } |
| 256 | |
| 257 | // The pre-copy (offline move stage): rounds while the instance serves, then |
| 258 | // the seal's round. The seal finds what stage copied and copies only what is |
| 259 | // new since; its receipts prove every multipart skip. |
| 260 | func TestMirrorPrefixReceiptsSkipAfterAPreCopy(t *testing.T) { |
| 261 | f := newFakeMirror() |
| 262 | _, staged := fillReplica(f, 90, "a") |
| 263 | if r := mirrorRound(t, f, testReceipts); r.BytesCopied != staged { |
| 264 | t.Fatalf("stage round = %+v, want %d bytes", r, staged) |
| 265 | } |
| 266 | // The instance kept writing: new LTX files. |
| 267 | _, fresh := fillReplica(f, 6, "b") |
| 268 | before := len(f.ops) |
| 269 | seal := mirrorRound(t, f, testReceipts) |
| 270 | if seal.BytesCopied != fresh || seal.Changed != 6 || seal.Deferred != 0 { |
| 271 | t.Fatalf("seal round = %+v, want only the 6 new objects (%d bytes)", seal, fresh) |
| 272 | } |
| 273 | for _, op := range f.ops[before:] { |
| 274 | if strings.HasPrefix(op, "copy ") && !strings.Contains(op, "/b-") { |
| 275 | t.Fatalf("the seal recopied a pre-copied object: %s", op) |
| 276 | } |
| 277 | } |
| 278 | sameBodies(t, stagedBodies(f), sourceBodies(f)) |
| 279 | } |
| 280 | |
| 281 | // Between stage and seal, litestream's retention and compaction delete |
| 282 | // replica files and write new ones. The seal's round deletes what the source |
| 283 | // no longer has and drops those receipts, so the staged listing is the |
| 284 | // source's again (the seal's sameListing). |
| 285 | func TestMirrorPrefixSourceDeletionsBetweenStageAndSeal(t *testing.T) { |
| 286 | f := newFakeMirror() |
| 287 | keys, _ := fillReplica(f, 30, "a") |
| 288 | mirrorRound(t, f, testReceipts) |
| 289 | // Compaction: twelve level-0 files folded into one level-1 file. |
| 290 | for _, k := range keys[:12] { |
| 291 | f.remove("src", k) |
| 292 | } |
| 293 | f.putMultipart("src", "instances/h/ltx/1/compacted.ltx", "compacted", 12000) |
| 294 | r := mirrorRound(t, f, testReceipts) |
| 295 | if r.Changed != 13 || r.BytesCopied != 12000 { |
| 296 | t.Fatalf("seal round = %+v, want 12 deletes and the one copy", r) |
| 297 | } |
| 298 | sameBodies(t, stagedBodies(f), sourceBodies(f)) |
| 299 | for rel := range readReceipts(t, f) { |
| 300 | if _, ok := sourceBodies(f)[rel]; !ok { |
| 301 | t.Fatalf("a receipt outlived its source object: %s", rel) |
| 302 | } |
| 303 | } |
| 304 | if again := mirrorRound(t, f, testReceipts); again.Changed != 0 { |
| 305 | t.Fatalf("round after = %+v", again) |
| 306 | } |
| 307 | } |
| 308 | |
| 309 | // A staged prefix left by an earlier move is never trusted as such. A stale |
| 310 | // object skips only when its ETag equals the source's (the same bytes) or a |
| 311 | // receipt for this exact pair still matches both sides; anything else is |
| 312 | // copied again, and every key the source lacks is deleted. |
| 313 | func TestMirrorPrefixStaleStagedPrefixIsNotTrusted(t *testing.T) { |
| 314 | f := newFakeMirror() |
| 315 | f.putMultipart("src", "instances/h/ltx/1/a.ltx", "new-a", 500) |
| 316 | f.putMultipart("src", "instances/h/ltx/1/b.ltx", "new-b", 500) |
| 317 | f.putSized("src", "instances/h/ltx/0/c.ltx", "new-c", 50) |
| 318 | f.putSized("src", "instances/h/ltx/0/d.ltx", "same-d", 50) |
| 319 | // The stale prefix: a under the old move's copy of other bytes, b with |
| 320 | // the same size as the source and a stale receipt, c changed, d byte |
| 321 | // identical, and an extra the source never had. |
| 322 | f.putSized("dst", "moves/h/ltx/1/a.ltx", "old-a", 500) |
| 323 | f.putSized("dst", "moves/h/ltx/1/b.ltx", "old-b", 500) |
| 324 | f.putSized("dst", "moves/h/ltx/0/c.ltx", "old-c", 50) |
| 325 | f.putSized("dst", "moves/h/ltx/0/d.ltx", "same-d", 50) |
| 326 | f.putSized("dst", "moves/h/ltx/0/extra.ltx", "old-x", 50) |
| 327 | // An old book for this pair whose receipt for b names the source as it |
| 328 | // is now but a destination ETag the staged object does not carry. |
| 329 | stale := (`{"v":1,"src_bucket":"src","src_prefix":"instances/h/","dst_bucket":"dst","dst_prefix":"moves/h/",` + |
| 330 | `"objects":{"ltx/1/b.ltx":{"src_etag":"new-b-27","src_size":500,"dst_etag":"new-b"},"ltx/0/extra.ltx":{"src_etag":"x","src_size":50,"dst_etag":"old-x"}}}`) |
| 331 | f.store("src", testReceipts.Key, fakeObject{MirrorObject{ETag: "book", Size: int64(len(stale))}, stale}) |
| 332 | |
| 333 | r := mirrorRound(t, f, testReceipts) |
| 334 | if r.Changed != 4 || r.BytesCopied != 1050 { |
| 335 | t.Fatalf("round = %+v, want a, b and c copied and the extra deleted", r) |
| 336 | } |
| 337 | sameBodies(t, stagedBodies(f), sourceBodies(f)) |
| 338 | book := readReceipts(t, f) |
| 339 | if _, ok := book["ltx/0/extra.ltx"]; ok { |
| 340 | t.Fatal("the stale receipt of a deleted key survived") |
| 341 | } |
| 342 | if got := book["ltx/1/b.ltx"]; got.DstETag != "new-b" || got.SrcETag != "new-b-27" { |
| 343 | t.Fatalf("b's receipt = %+v", got) |
| 344 | } |
| 345 | |
| 346 | // A book written for another pair (another move's target) is no |
| 347 | // evidence at all, even when its receipt would match. |
| 348 | g := newFakeMirror() |
| 349 | g.putMultipart("src", "instances/h/ltx/1/a.ltx", "a", 500) |
| 350 | g.putSized("dst", "moves/h/ltx/1/a.ltx", "a", 500) |
| 351 | other := `{"v":1,"src_bucket":"src","src_prefix":"instances/h/","dst_bucket":"other-dst","dst_prefix":"moves/h/",` + |
| 352 | `"objects":{"ltx/1/a.ltx":{"src_etag":"a-27","src_size":500,"dst_etag":"a"}}}` |
| 353 | g.store("src", testReceipts.Key, fakeObject{MirrorObject{ETag: "book", Size: int64(len(other))}, other}) |
| 354 | if r := mirrorRound(t, g, testReceipts); r.Changed != 1 { |
| 355 | t.Fatalf("round over another pair's book = %+v, want the object copied again", r) |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | // forEachParallel on its own: an error stops dispatch, and an already-ended |
| 360 | // context starts nothing. |
| 361 | func TestForEachParallel(t *testing.T) { |
| 362 | var ran atomic.Int32 |
| 363 | err := forEachParallel(context.Background(), 1, 10, func(_ context.Context, i int) error { |
| 364 | ran.Add(1) |
| 365 | if i == 3 { |
| 366 | return errors.New("boom") |
| 367 | } |
| 368 | return nil |
| 369 | }) |
| 370 | if err == nil || err.Error() != "boom" || ran.Load() > 5 { |
| 371 | t.Fatalf("err %v after %d calls, want boom after at most 5", err, ran.Load()) |
| 372 | } |
| 373 | ctx, cancel := context.WithCancel(context.Background()) |
| 374 | cancel() |
| 375 | ran.Store(0) |
| 376 | err = forEachParallel(ctx, 4, 10, func(context.Context, int) error { ran.Add(1); return nil }) |
| 377 | if !errors.Is(err, context.Canceled) || ran.Load() > 4 { |
| 378 | t.Fatalf("err %v after %d calls on an ended context", err, ran.Load()) |
| 379 | } |
| 380 | if err := forEachParallel(context.Background(), 8, 0, nil); err != nil { |
| 381 | t.Fatal(err) |
| 382 | } |
| 383 | } |
| 384 | |
| 385 | // Astra round 1, finding 3: no network request is made while the round's |
| 386 | // accounting is locked. A receipts read that stalls once a copy has landed |
| 387 | // (the lazy load a receipt write used to make under the lock) must not keep |
| 388 | // another copy's failure from stopping the round: the book is read before |
| 389 | // any copy starts. |
| 390 | func TestMirrorPrefixAReceiptsReadNeverBlocksTheFirstFailure(t *testing.T) { |
| 391 | f := newFakeMirror() |
| 392 | f.workers = 2 |
| 393 | f.putMultipart("src", "instances/h/ltx/1/a.ltx", "a", 100) // lands, needs a receipt |
| 394 | f.putSized("src", "instances/h/ltx/0/b.ltx", "b", 10) // fails |
| 395 | var landed atomic.Bool |
| 396 | release := make(chan struct{}) |
| 397 | t.Cleanup(func() { close(release) }) |
| 398 | f.onCopy = func(_ context.Context, key string) error { |
| 399 | if strings.HasSuffix(key, "b.ltx") { |
| 400 | // Fail only once a's copy has landed and its receipt is due. |
| 401 | for deadline := time.Now().Add(5 * time.Second); !landed.Load() && time.Now().Before(deadline); time.Sleep(time.Millisecond) { |
| 402 | } |
| 403 | time.Sleep(20 * time.Millisecond) |
| 404 | return errors.New("injected copy failure") |
| 405 | } |
| 406 | landed.Store(true) |
| 407 | return nil |
| 408 | } |
| 409 | f.onGet = func(string) { |
| 410 | if landed.Load() { |
| 411 | <-release // a receipts read that stalls after a copy |
| 412 | } |
| 413 | } |
| 414 | done := make(chan error, 1) |
| 415 | go func() { |
| 416 | _, err := MirrorPrefix(context.Background(), f, "src", "instances/h/", "dst", "moves/h/", testReceipts) |
| 417 | done <- err |
| 418 | }() |
| 419 | select { |
| 420 | case err := <-done: |
| 421 | if err == nil || !strings.Contains(err.Error(), "injected copy failure") { |
| 422 | t.Fatalf("round error = %v, want the injected failure", err) |
| 423 | } |
| 424 | case <-time.After(3 * time.Second): |
| 425 | t.Fatal("a receipts read made under the round's lock kept the first failure from stopping the round") |
| 426 | } |
| 427 | } |
| 428 | |
amadan