| 1 | package rollout |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "runtime" |
| 8 | "strings" |
| 9 | "sync" |
| 10 | "sync/atomic" |
| 11 | "testing" |
| 12 | "time" |
| 13 | |
| 14 | "github.com/carlosframework/platform/internal/activator" |
| 15 | "github.com/carlosframework/platform/internal/appconfig" |
| 16 | "github.com/carlosframework/platform/internal/registry" |
| 17 | "github.com/carlosframework/platform/internal/release" |
| 18 | "github.com/carlosframework/platform/internal/rolloutwords" |
| 19 | "github.com/carlosframework/platform/internal/status" |
| 20 | ) |
| 21 | |
| 22 | func TestAReleaseRollsOutCanaryFirstThenPaced(t *testing.T) { |
| 23 | h := newHarness(t, "a", "b", "c") |
| 24 | h.stage("v2") |
| 25 | h.install("v2", "a", "b", "c") |
| 26 | h.pass() |
| 27 | waitFor(t, "every host on v2", func() bool { return h.servingAll("v2", "a", "b", "c") }) |
| 28 | calls, _, maxFlight := h.act.snapshot() |
| 29 | if len(calls) != 3 || !calls[0].a.Canary || calls[1].a.Canary { |
| 30 | t.Fatalf("calls = %+v, want the canary first and then two plain swaps", calls) |
| 31 | } |
| 32 | if maxFlight > 1 { |
| 33 | t.Fatalf("%d swaps ran at once with swap_concurrency 1", maxFlight) |
| 34 | } |
| 35 | for _, host := range []string{"a", "b", "c"} { |
| 36 | if r := h.rows.get(host); r.GoodVersion != "v2" || r.GoodDigest != "sha-v2" || r.ServingVersion != "v2" { |
| 37 | t.Fatalf("%s row = %+v, want good and serving v2", host, r) |
| 38 | } |
| 39 | } |
| 40 | if !h.events.has(textHead(textSwapped("v2", time.Hour), time.Hour.String())) { |
| 41 | t.Fatal("no swapped event") |
| 42 | } |
| 43 | } |
| 44 | |
| 45 | // Review focus 1, the index's cross-plan item. |
| 46 | func TestTwoReleasesTwentySecondsApartNeverSwitchToTheFirst(t *testing.T) { |
| 47 | h := newHarness(t, "a", "b", "c") |
| 48 | h.stage("v2", "v3") |
| 49 | gate := h.act.hold("a") // the canary onto v2 sits in its settle |
| 50 | h.install("v2", "a", "b", "c") |
| 51 | h.pass() |
| 52 | waitFor(t, "the v2 canary to start", func() bool { _, started, _ := h.act.snapshot(); return len(started) == 1 }) |
| 53 | h.advance(20 * time.Second) |
| 54 | h.install("v3", "a", "b", "c") |
| 55 | h.pass() |
| 56 | close(gate) |
| 57 | waitFor(t, "every host on v3", func() bool { return h.servingAll("v3", "a", "b", "c") }) |
| 58 | calls, _, maxFlight := h.act.snapshot() |
| 59 | for _, c := range calls { |
| 60 | if c.a.Build == "v2" && c.outcome == activator.SwapDone { |
| 61 | t.Fatalf("switched %s to the superseded v2", c.a.Host) |
| 62 | } |
| 63 | } |
| 64 | if calls[0].a.Build != "v2" || calls[0].outcome != activator.SwapAbandoned { |
| 65 | t.Fatalf("first call = %+v, want the v2 canary abandoned at its re-check", calls[0]) |
| 66 | } |
| 67 | if len(h.book.List()) != 0 { |
| 68 | t.Fatal("an abandoned attempt is not a build failure") |
| 69 | } |
| 70 | if maxFlight > 1 { |
| 71 | t.Fatal("swap_concurrency 1 exceeded") |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | func TestAFailedCanaryHaltsTheCohort(t *testing.T) { |
| 76 | h := newHarness(t, "a", "b", "c") |
| 77 | h.stage("v2") |
| 78 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 79 | if a.Build == "v2" { |
| 80 | return activator.SwapFailed |
| 81 | } |
| 82 | return activator.SwapDone |
| 83 | } |
| 84 | h.install("v2", "a", "b", "c") |
| 85 | h.pass() |
| 86 | waitFor(t, "a failure record", func() bool { _, ok := h.book.Lookup("bab", "blog", "v2"); return ok }) |
| 87 | h.pass() |
| 88 | time.Sleep(50 * time.Millisecond) |
| 89 | if calls, _, _ := h.act.snapshot(); len(calls) != 1 { |
| 90 | t.Fatalf("%d attempts onto v2, want the halt at the first failure", len(calls)) |
| 91 | } |
| 92 | if r := h.rows.get("a"); r.GoodVersion != "v1" { |
| 93 | t.Fatalf("good build moved to %q on a failure", r.GoodVersion) |
| 94 | } |
| 95 | if !h.events.has(textHead(textCohortHalted("v2", "\x00", status.ReasonHealth), "\x00")) { |
| 96 | t.Fatal("no cohort-halted event") |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | // Spec §8.1: the failure transition (the record and the halt) happens in |
| 101 | // the same step that frees the failed attempt's slot. The hook looks at the |
| 102 | // queue at every moment done() reads a row after the failure: whenever the |
| 103 | // failed attempt's slot is free, the record must already be in the book |
| 104 | // (which wake plans read) and admission must start nothing onto the build. |
| 105 | // Halting in that step but writing the record after the row read let a |
| 106 | // wake select the failed build in between. |
| 107 | func TestAFailureIsRecordedAndHaltedInTheStepThatFreesItsSlot(t *testing.T) { |
| 108 | h := newHarness(t, "a", "b") |
| 109 | h.stage("v2") |
| 110 | var failed atomic.Bool |
| 111 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 112 | if a.Build != "v2" { |
| 113 | return activator.SwapDone |
| 114 | } |
| 115 | failed.Store(true) |
| 116 | return activator.SwapFailed |
| 117 | } |
| 118 | var bad atomic.Int32 |
| 119 | h.rows.mu.Lock() |
| 120 | h.rows.getHook = func(host string) { |
| 121 | if !failed.Load() { |
| 122 | return |
| 123 | } |
| 124 | s := h.m.schedulerOf(host) |
| 125 | s.mu.Lock() |
| 126 | hs := s.q.hosts[host] |
| 127 | free := hs != nil && hs.running == nil |
| 128 | var ds []dispatch |
| 129 | if free { |
| 130 | ds = s.q.admit(currentView(s.q), h.clock()) |
| 131 | } |
| 132 | s.mu.Unlock() |
| 133 | if !free { |
| 134 | return |
| 135 | } |
| 136 | if _, ok := h.book.Lookup("bab", "blog", "v2"); !ok { |
| 137 | bad.Add(1) |
| 138 | } |
| 139 | for _, d := range ds { |
| 140 | if d.ID.Build == "v2" { |
| 141 | bad.Add(1) |
| 142 | } |
| 143 | } |
| 144 | } |
| 145 | h.rows.mu.Unlock() |
| 146 | h.install("v2", "a", "b") |
| 147 | h.pass() |
| 148 | waitFor(t, "a failure record", func() bool { _, ok := h.book.Lookup("bab", "blog", "v2"); return ok }) |
| 149 | h.pass() // reads the rows again, with the slot free |
| 150 | if n := bad.Load(); n != 0 { |
| 151 | t.Fatalf("%d moment(s) with the slot free and the record missing or a v2 attempt admissible", n) |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | // Spec §8.5: a successful attempt keeps its host's slot until its state |
| 156 | // (serving and good build) is published, so no newer attempt can be |
| 157 | // admitted and complete first and then be overwritten by this one. Freeing |
| 158 | // the slot before the write let that happen. |
| 159 | func TestASuccessKeepsItsSlotUntilItIsPublished(t *testing.T) { |
| 160 | h := newHarness(t, "a") |
| 161 | h.stage("v2") |
| 162 | var freed atomic.Int32 |
| 163 | h.rows.mu.Lock() |
| 164 | h.rows.setGoodHook = func(host string) { |
| 165 | s := h.m.schedulerOf(host) |
| 166 | s.mu.Lock() |
| 167 | if hs := s.q.hosts[host]; hs != nil && hs.running == nil { |
| 168 | freed.Add(1) |
| 169 | } |
| 170 | s.mu.Unlock() |
| 171 | } |
| 172 | h.rows.mu.Unlock() |
| 173 | h.install("v2", "a") |
| 174 | h.pass() |
| 175 | waitFor(t, "the good build moved to v2", func() bool { return h.rows.get("a").GoodVersion == "v2" }) |
| 176 | if n := freed.Load(); n != 0 { |
| 177 | t.Fatalf("%d good-build write(s) made after the attempt's slot was freed", n) |
| 178 | } |
| 179 | } |
| 180 | |
| 181 | // A failure is published with no blocking read between the failure and the |
| 182 | // record (spec §8.1): the pointer the build failed under is read before the |
| 183 | // attempt runs. Reading it in done(), before the record, left every sibling's |
| 184 | // wake free to plan the failed build for as long as that read took. |
| 185 | func TestNoPointerReadSitsBetweenAFailureAndItsRecord(t *testing.T) { |
| 186 | h := newHarness(t, "a") |
| 187 | h.stage("v2") |
| 188 | var failed, late atomic.Bool |
| 189 | h.act.outcome = func(activator.Attempt) activator.SwapOutcome { |
| 190 | failed.Store(true) |
| 191 | return activator.SwapFailed |
| 192 | } |
| 193 | h.mu.Lock() |
| 194 | h.pointerHook = func() { |
| 195 | if failed.Load() { |
| 196 | if _, ok := h.book.Lookup("bab", "blog", "v2"); !ok { |
| 197 | late.Store(true) |
| 198 | } |
| 199 | } |
| 200 | } |
| 201 | h.mu.Unlock() |
| 202 | h.install("v2", "a") |
| 203 | h.pass() |
| 204 | waitFor(t, "a failure record", func() bool { _, ok := h.book.Lookup("bab", "blog", "v2"); return ok }) |
| 205 | if late.Load() { |
| 206 | t.Fatal("a pointer read ran after the failure and before its record was published") |
| 207 | } |
| 208 | } |
| 209 | |
| 210 | // Spec §6.3: a build installed while the entry is unstable (a hold |
| 211 | // draining, a wake running) still supersedes the attempt in flight, so its |
| 212 | // gates refuse the older build. Skipping wantBuild for such a host left the |
| 213 | // revision on the older build, and a hold's pinned wake exec'd it. Nothing |
| 214 | // is admitted onto the host until it is stable again. |
| 215 | func TestAnInstallDuringAnUnstableEntryStillSupersedes(t *testing.T) { |
| 216 | h := newHarness(t, "a") |
| 217 | h.pass() |
| 218 | s := h.m.schedulerOf("a") |
| 219 | h.act.mu.Lock() |
| 220 | delete(h.act.gens, "a") |
| 221 | h.act.owned["a"] = true // stopping or waking: a process, no current generation |
| 222 | h.act.mu.Unlock() |
| 223 | s.mu.Lock() |
| 224 | before := s.q.newest("a") |
| 225 | s.mu.Unlock() |
| 226 | h.stage("v2") |
| 227 | h.install("v2", "a") |
| 228 | h.pass() |
| 229 | s.mu.Lock() |
| 230 | after := s.q.newest("a") |
| 231 | s.mu.Unlock() |
| 232 | if after == before { |
| 233 | t.Fatal("an install during an unstable entry left the attempt revision unchanged") |
| 234 | } |
| 235 | time.Sleep(50 * time.Millisecond) |
| 236 | if calls, _, _ := h.act.snapshot(); len(calls) != 0 { |
| 237 | t.Fatalf("attempts %+v admitted onto an unstable entry", calls) |
| 238 | } |
| 239 | } |
| 240 | |
| 241 | // A row that cannot be read at the failure still gets its record: the |
| 242 | // failure is the build's, not the row's (spec §8.1). Skipping the write on |
| 243 | // a failed registry read left the failed build selectable by every wake. |
| 244 | func TestAFailureIsRecordedWhenItsRowIsUnreadable(t *testing.T) { |
| 245 | h := newHarness(t, "a") |
| 246 | h.stage("v2") |
| 247 | var failed atomic.Bool |
| 248 | h.act.outcome = func(activator.Attempt) activator.SwapOutcome { |
| 249 | failed.Store(true) |
| 250 | return activator.SwapFailed |
| 251 | } |
| 252 | h.rows.mu.Lock() |
| 253 | h.rows.getErr = func(string) error { |
| 254 | if failed.Load() { |
| 255 | return errors.New("registry: database is locked") |
| 256 | } |
| 257 | return nil |
| 258 | } |
| 259 | h.rows.mu.Unlock() |
| 260 | h.install("v2", "a") |
| 261 | h.pass() |
| 262 | waitFor(t, "a failure record despite the unreadable row", func() bool { |
| 263 | _, ok := h.book.Lookup("bab", "blog", "v2") |
| 264 | return ok |
| 265 | }) |
| 266 | } |
| 267 | |
| 268 | // Review focus 4. |
| 269 | func TestStagingRefusalWritesNoFailureRecord(t *testing.T) { |
| 270 | h := newHarness(t, "a") |
| 271 | h.stage("v2") |
| 272 | h.mismatch["v2"] = true |
| 273 | h.install("v2", "a") |
| 274 | h.pass() |
| 275 | waitFor(t, "the stage relay touch", func() bool { return h.touched.Load() > 0 }) |
| 276 | if len(h.book.List()) != 0 { |
| 277 | t.Fatal("a staging error wrote a failure record against the build") |
| 278 | } |
| 279 | if calls, _, _ := h.act.snapshot(); len(calls) != 0 { |
| 280 | t.Fatal("a refused attempt reached Swap") |
| 281 | } |
| 282 | if g, _ := h.act.Current("a"); g.Build != "v1" { |
| 283 | t.Fatal("the instance left its serving build") |
| 284 | } |
| 285 | if !h.events.has(textStagingRefused("v2")) { |
| 286 | t.Fatal("no staging event") |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | func TestAnUnstagedBuildIsAStagingRefusal(t *testing.T) { |
| 291 | h := newHarness(t, "a") |
| 292 | h.install("v9", "a") // never staged on this box |
| 293 | h.pass() |
| 294 | waitFor(t, "the stage relay touch", func() bool { return h.touched.Load() > 0 }) |
| 295 | if len(h.book.List()) != 0 { |
| 296 | t.Fatal("a missing staged file wrote a failure record") |
| 297 | } |
| 298 | } |
| 299 | |
| 300 | func TestRestartGoesAheadForAnInstanceServingAFailedBuild(t *testing.T) { |
| 301 | h := newHarness(t, "d") |
| 302 | h.stage("v2") |
| 303 | h.act.serve("d", "v2") |
| 304 | r := h.rows.get("d") |
| 305 | r.Version, r.GoodVersion, r.GoodDigest = "v2", "v2", "sha-v2" |
| 306 | h.rows.put(r) |
| 307 | h.book.Record(FailureRecord{Account: "bab", App: "blog", Build: "v2", Host: "e", At: t0}) |
| 308 | h.pass() |
| 309 | if !h.m.Restart("d", "gen:etag-1") { |
| 310 | t.Fatal("the restart was not queued") |
| 311 | } |
| 312 | waitFor(t, "the restart", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 1 }) |
| 313 | calls, _, _ := h.act.snapshot() |
| 314 | if c := calls[0].a; c.Cause != activator.CauseRestart || c.Build != "v2" || c.RestartID != "gen:etag-1" { |
| 315 | t.Fatalf("attempt = %+v", c) |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | func TestRecoveryUsesTheRetainedIdentityAndTheCurrentPolicy(t *testing.T) { |
| 320 | h := newHarness(t, "a") |
| 321 | h.swap = "hold" // switched to hold after v1 passed under overlap |
| 322 | h.pass() |
| 323 | h.m.Restart("a", "gen:etag-1") |
| 324 | waitFor(t, "the restart", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 1 }) |
| 325 | calls, _, _ := h.act.snapshot() |
| 326 | g, _ := h.act.Current("a") |
| 327 | if c := calls[0].a; c.Mode != activator.ModeHold || c.StagedPath != "/opt/v1" || c.Digest != "sha-v1" { |
| 328 | t.Fatalf("attempt = %+v, want hold, by the generation's captured path and digest (current gen %+v)", c, g) |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | func TestAReleaseOntoTheServingBuildMakesNoSwap(t *testing.T) { |
| 333 | h := newHarness(t, "a") |
| 334 | h.pass() |
| 335 | time.Sleep(50 * time.Millisecond) |
| 336 | if calls, _, _ := h.act.snapshot(); len(calls) != 0 { |
| 337 | t.Fatalf("calls = %+v", calls) |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | // Review focus 5. |
| 342 | func TestConfigDeliveredMidSwapIsAckedOnlyWhenActive(t *testing.T) { |
| 343 | h := newHarness(t, "a") |
| 344 | h.stage("v2") |
| 345 | gate := h.act.hold("a") |
| 346 | h.install("v2", "a") |
| 347 | h.pass() |
| 348 | waitFor(t, "the release swap to start", func() bool { _, s, _ := h.act.snapshot(); return len(s) == 1 }) |
| 349 | // The ack is written from the scheduler's goroutine: counted atomically. |
| 350 | var acks atomic.Int32 |
| 351 | ack := h.m.DeferAck(func(context.Context, appconfig.ConfigAck) { acks.Add(1) }) |
| 352 | h.m.ConfigDelivered("a", "sha256:c1") |
| 353 | ack(appconfig.ConfigAck{Host: "a"}) |
| 354 | if acks.Load() != 0 { |
| 355 | t.Fatal("acknowledged a config no generation has started with") |
| 356 | } |
| 357 | close(gate) |
| 358 | waitFor(t, "the config attempt", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 2 }) |
| 359 | calls, _, _ := h.act.snapshot() |
| 360 | if c := calls[1].a; c.Cause != activator.CauseConfig || c.ConfigRev != "sha256:c1" || c.Build != "v2" { |
| 361 | t.Fatalf("second attempt = %+v, want a config attempt onto v2", c) |
| 362 | } |
| 363 | waitFor(t, "the ack", func() bool { return acks.Load() == 1 }) |
| 364 | } |
| 365 | |
| 366 | // Spec §8.1: a successful config attempt's activation is published before |
| 367 | // its slot frees, and with no registry read in front of it. Published after |
| 368 | // the slot and the read, a newer config attempt could complete first and be |
| 369 | // overwritten (active went back to the older revision), and an unreadable |
| 370 | // row lost the activation, leaving delivered and active apart for good. The |
| 371 | // held ack is released (handed to the writer, its order fixed) in the same |
| 372 | // step; the write itself is the writer's, after (Task 11). |
| 373 | func TestConfigActivationIsPublishedBeforeTheSlotFrees(t *testing.T) { |
| 374 | h := newHarness(t, "a") |
| 375 | h.pass() |
| 376 | var acks, released, freed atomic.Int32 |
| 377 | h.m.acks.onSubmit = func(host string) { |
| 378 | released.Add(1) |
| 379 | s := h.m.schedulerOf("a") |
| 380 | s.mu.Lock() |
| 381 | if hs := s.q.hosts["a"]; hs != nil && hs.running == nil { |
| 382 | freed.Add(1) |
| 383 | } |
| 384 | s.mu.Unlock() |
| 385 | } |
| 386 | ack := h.m.DeferAck(func(context.Context, appconfig.ConfigAck) { acks.Add(1) }) |
| 387 | var swapped atomic.Bool |
| 388 | h.act.outcome = func(activator.Attempt) activator.SwapOutcome { |
| 389 | swapped.Store(true) |
| 390 | return activator.SwapDone |
| 391 | } |
| 392 | h.rows.mu.Lock() |
| 393 | h.rows.getErr = func(string) error { |
| 394 | if swapped.Load() { |
| 395 | return errors.New("registry: database is locked") |
| 396 | } |
| 397 | return nil |
| 398 | } |
| 399 | h.rows.mu.Unlock() |
| 400 | // The config attempt waits at the gate until the ack is held, so the ack |
| 401 | // is released by the activation and not written at once by a config sync |
| 402 | // call that came after it. |
| 403 | gate := h.act.hold("a") |
| 404 | h.m.ConfigDelivered("a", "sha256:c1") |
| 405 | ack(appconfig.ConfigAck{Host: "a"}) |
| 406 | if acks.Load() != 0 { |
| 407 | t.Fatal("acknowledged a config no generation has started with") |
| 408 | } |
| 409 | close(gate) |
| 410 | waitFor(t, "the config activated and acknowledged", func() bool { return acks.Load() == 1 }) |
| 411 | if released.Load() != 1 { |
| 412 | t.Fatalf("%d acks released, want 1", released.Load()) |
| 413 | } |
| 414 | if n := freed.Load(); n != 0 { |
| 415 | t.Fatal("the activation was published after the attempt's slot was freed") |
| 416 | } |
| 417 | } |
| 418 | |
| 419 | func TestAFailedConfigAttemptRetriesWithBackoff(t *testing.T) { |
| 420 | h := newHarness(t, "a") |
| 421 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 422 | if a.Cause == activator.CauseConfig { |
| 423 | return activator.SwapFailed |
| 424 | } |
| 425 | return activator.SwapDone |
| 426 | } |
| 427 | h.pass() |
| 428 | h.m.ConfigDelivered("a", "sha256:c1") |
| 429 | waitFor(t, "the first config attempt", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 1 }) |
| 430 | waitFor(t, "the pending event", func() bool { return h.events.has(configPendingHead) }) |
| 431 | h.advance(59 * time.Second) |
| 432 | h.pass() |
| 433 | time.Sleep(50 * time.Millisecond) |
| 434 | if calls, _, _ := h.act.snapshot(); len(calls) != 1 { |
| 435 | t.Fatal("retried inside the one-minute backoff") |
| 436 | } |
| 437 | h.advance(2 * time.Second) |
| 438 | h.pass() |
| 439 | waitFor(t, "the retry after the backoff", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 2 }) |
| 440 | if len(h.book.List()) != 0 { |
| 441 | t.Fatal("a failed config attempt wrote a failure record") |
| 442 | } |
| 443 | } |
| 444 | |
| 445 | func TestBootstrapRecordsAwakeInstancesOnly(t *testing.T) { |
| 446 | h := newHarness(t, "awake", "asleep") |
| 447 | for _, host := range []string{"awake", "asleep"} { |
| 448 | r := h.rows.get(host) |
| 449 | r.GoodVersion, r.GoodDigest = "", "" |
| 450 | h.rows.put(r) |
| 451 | } |
| 452 | h.act.mu.Lock() |
| 453 | delete(h.act.gens, "asleep") |
| 454 | h.act.mu.Unlock() |
| 455 | rs, _ := h.rows.List() |
| 456 | h.m.Bootstrap(rs) |
| 457 | if r := h.rows.get("awake"); r.Good() != (registry.GoodBuild{Version: "v1", Digest: "sha-v1"}) { |
| 458 | t.Fatalf("awake row good = %+v", r.Good()) |
| 459 | } |
| 460 | if r := h.rows.get("asleep"); r.GoodVersion != "" { |
| 461 | t.Fatal("nothing may be inferred for a sleeping instance (spec §8.5)") |
| 462 | } |
| 463 | } |
| 464 | |
| 465 | // `carlos promote --retry` re-signs the channel pointer with a fresh |
| 466 | // PromotedAt. The record keeps the PromotedAt it failed under, survives that |
| 467 | // same pointer being read again, and clears on the re-signed one. |
| 468 | func TestAReSignedPointerClearsTheRecord(t *testing.T) { |
| 469 | h := newHarness(t, "a", "b") |
| 470 | h.stage("v2") |
| 471 | h.failing.Store(true) |
| 472 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 473 | if a.Build == "v2" && h.failing.Load() { |
| 474 | return activator.SwapFailed |
| 475 | } |
| 476 | return activator.SwapDone |
| 477 | } |
| 478 | h.pointer("stable", "v2", promotedAt) |
| 479 | h.install("v2", "a", "b") |
| 480 | h.pass() |
| 481 | waitFor(t, "a failure record", func() bool { _, ok := h.book.Lookup("bab", "blog", "v2"); return ok }) |
| 482 | if r, _ := h.book.Lookup("bab", "blog", "v2"); !r.PromotedAt.Equal(promotedAt) { |
| 483 | t.Fatalf("record keeps PromotedAt %v, want the pointer's %v", r.PromotedAt, promotedAt) |
| 484 | } |
| 485 | h.advance(pointerReadEvery) |
| 486 | h.pass() |
| 487 | if _, ok := h.book.Lookup("bab", "blog", "v2"); !ok { |
| 488 | t.Fatal("the pointer the build failed under cleared its record") |
| 489 | } |
| 490 | h.failing.Store(false) |
| 491 | h.pointer("stable", "v2", promotedAt.Add(time.Minute)) |
| 492 | h.advance(pointerReadEvery) |
| 493 | h.pass() |
| 494 | waitFor(t, "every host on v2", func() bool { return h.servingAll("v2", "a", "b") }) |
| 495 | calls, _, _ := h.act.snapshot() |
| 496 | if c := calls[1].a; c.Cause != activator.CauseRetry || !c.Canary { |
| 497 | t.Fatalf("first attempt after the retry = %+v, want a retry canary", c) |
| 498 | } |
| 499 | } |
| 500 | |
| 501 | // `carlos instances retry <host>`: that host alone, as a canary. |
| 502 | func TestRetryHostClearsOneInstance(t *testing.T) { |
| 503 | h := newHarness(t, "a", "b") |
| 504 | h.stage("v2") |
| 505 | h.book.Record(FailureRecord{Account: "bab", App: "blog", Build: "v2", Host: "c", PromotedAt: promotedAt, At: t0}) |
| 506 | h.install("v2", "a", "b") |
| 507 | h.pass() |
| 508 | if !h.retry("b", "v2") { |
| 509 | t.Fatal("RetryHost did not clear b") |
| 510 | } |
| 511 | waitFor(t, "b on v2", func() bool { return h.servingAll("v2", "b") }) |
| 512 | if h.servingAll("v2", "a") { |
| 513 | t.Fatal("a host retry moved another host") |
| 514 | } |
| 515 | } |
| 516 | |
| 517 | // Carried from Task 3: a canonical manifest that cannot be read, or a store |
| 518 | // that cannot compare digests, must reach AttemptMode as a hold demand, never |
| 519 | // as "". Only release.ErrStagedDigestMismatch is a staging refusal. |
| 520 | func TestAnUnreadableCanonicalManifestRunsInHold(t *testing.T) { |
| 521 | for name, set := range map[string]func(h *harness){ |
| 522 | "demand error": func(h *harness) { h.demandErr["v2"] = errors.New("store: GetObject: connection reset") }, |
| 523 | "no digest": func(h *harness) { h.noDigest["v2"] = true }, |
| 524 | } { |
| 525 | t.Run(name, func(t *testing.T) { |
| 526 | h := newHarness(t, "a") |
| 527 | h.stage("v2") |
| 528 | set(h) |
| 529 | h.install("v2", "a") |
| 530 | h.pass() |
| 531 | waitFor(t, "the attempt", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 1 }) |
| 532 | want, err := release.FileSHA256(release.VersionPath(h.opt, "bab", "v2", "blog")) |
| 533 | if err != nil { |
| 534 | t.Fatal(err) |
| 535 | } |
| 536 | calls, _, _ := h.act.snapshot() |
| 537 | if c := calls[0].a; c.Mode != activator.ModeHold || c.Demand != release.DemandHold || c.Digest != want { |
| 538 | t.Fatalf("attempt = %+v, want hold on the staged file's own digest %s", c, want) |
| 539 | } |
| 540 | if h.touched.Load() != 0 || len(h.book.List()) != 0 { |
| 541 | t.Fatal("an unreadable manifest is neither a staging refusal nor a build failure") |
| 542 | } |
| 543 | }) |
| 544 | } |
| 545 | } |
| 546 | |
| 547 | // Carried from Task 4: a record keeps a real PromotedAt. With no pointer to |
| 548 | // read, the failure's own time stands in, never a zero time. |
| 549 | func TestAFailureWithNoReadablePointerKeepsItsOwnTime(t *testing.T) { |
| 550 | h := newHarness(t, "a") |
| 551 | h.stage("v2") |
| 552 | h.act.outcome = func(activator.Attempt) activator.SwapOutcome { return activator.SwapFailed } |
| 553 | h.install("v2", "a") |
| 554 | h.pass() |
| 555 | waitFor(t, "a failure record", func() bool { _, ok := h.book.Lookup("bab", "blog", "v2"); return ok }) |
| 556 | r, _ := h.book.Lookup("bab", "blog", "v2") |
| 557 | if r.PromotedAt.IsZero() || !r.PromotedAt.Equal(r.At) { |
| 558 | t.Fatalf("record %+v: want the failure's own time as PromotedAt", r) |
| 559 | } |
| 560 | } |
| 561 | |
| 562 | // Carried from Task 5: a second `carlos promote --retry` of the same build, |
| 563 | // after the retry failed too, clears the record again and the next attempt is |
| 564 | // a canary again. Nothing may rest on wantBuild reporting a change. |
| 565 | func TestASecondRetryOfTheSameBuildIsACanaryAgain(t *testing.T) { |
| 566 | h := newHarness(t, "a", "b") |
| 567 | h.stage("v2") |
| 568 | h.failing.Store(true) |
| 569 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 570 | if a.Build == "v2" && h.failing.Load() { |
| 571 | return activator.SwapFailed |
| 572 | } |
| 573 | return activator.SwapDone |
| 574 | } |
| 575 | h.pointer("stable", "v2", promotedAt) |
| 576 | h.install("v2", "a", "b") |
| 577 | h.pass() |
| 578 | waitFor(t, "the first failure", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 1 }) |
| 579 | waitFor(t, "a failure record", func() bool { _, ok := h.book.Lookup("bab", "blog", "v2"); return ok }) |
| 580 | |
| 581 | h.pointer("stable", "v2", promotedAt.Add(time.Minute)) |
| 582 | h.advance(pointerReadEvery) |
| 583 | h.pass() |
| 584 | waitFor(t, "the retry to fail too", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 2 }) |
| 585 | waitFor(t, "the record again", func() bool { |
| 586 | r, ok := h.book.Lookup("bab", "blog", "v2") |
| 587 | return ok && r.PromotedAt.Equal(promotedAt.Add(time.Minute)) |
| 588 | }) |
| 589 | |
| 590 | h.failing.Store(false) |
| 591 | h.pointer("stable", "v2", promotedAt.Add(2*time.Minute)) |
| 592 | h.advance(pointerReadEvery) |
| 593 | h.pass() |
| 594 | waitFor(t, "every host on v2", func() bool { return h.servingAll("v2", "a", "b") }) |
| 595 | calls, _, _ := h.act.snapshot() |
| 596 | for i, c := range calls[:3] { |
| 597 | if c.a.Host != "a" || !c.a.Canary { |
| 598 | t.Fatalf("call %d = %+v, want a canary on a", i, c.a) |
| 599 | } |
| 600 | } |
| 601 | if c := calls[2].a; c.Cause != activator.CauseRetry { |
| 602 | t.Fatalf("second retry's attempt = %+v, want cause retry", c) |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | // Carried from Task 6 and plan 3 (O7): a hold attempt that fails is held |
| 607 | // down in the book before its slot frees, so no wake planned in between can |
| 608 | // boot the failed hold build. The mode is the one Swap ran (a foreign-layout |
| 609 | // overlap runs as a hold, SwapResult.Mode), and the record's build comes from |
| 610 | // the dispatch: a failed hold returns no generation. |
| 611 | func TestAFailedHoldIsHeldDownBeforeItsSlotFrees(t *testing.T) { |
| 612 | h := newHarness(t, "a") |
| 613 | h.stage("v2") |
| 614 | h.act.mode = activator.ModeHold |
| 615 | var failed atomic.Bool |
| 616 | h.act.outcome = func(activator.Attempt) activator.SwapOutcome { |
| 617 | failed.Store(true) |
| 618 | return activator.SwapDown |
| 619 | } |
| 620 | var bad, seen atomic.Int32 |
| 621 | h.rows.mu.Lock() |
| 622 | h.rows.getHook = func(host string) { |
| 623 | if !failed.Load() { |
| 624 | return |
| 625 | } |
| 626 | s := h.m.schedulerOf(host) |
| 627 | s.mu.Lock() |
| 628 | free := s.q.hosts[host] != nil && s.q.hosts[host].running == nil |
| 629 | s.mu.Unlock() |
| 630 | if !free { |
| 631 | return |
| 632 | } |
| 633 | seen.Add(1) |
| 634 | if !h.book.HeldDown("bab", "blog", "v2", host) { |
| 635 | bad.Add(1) |
| 636 | } |
| 637 | } |
| 638 | h.rows.mu.Unlock() |
| 639 | h.install("v2", "a") |
| 640 | h.pass() |
| 641 | waitFor(t, "the held-down record", func() bool { return h.book.HeldDown("bab", "blog", "v2", "a") }) |
| 642 | waitFor(t, "the down event", func() bool { return h.events.has(textSwapDown("v2")) }) |
| 643 | if bad.Load() != 0 || seen.Load() == 0 { |
| 644 | t.Fatalf("%d of %d look(s) with the slot free found the host not held down", bad.Load(), seen.Load()) |
| 645 | } |
| 646 | r, _ := h.book.Lookup("bab", "blog", "v2") |
| 647 | if r.Mode != string(activator.ModeHold) || r.Build != "v2" || r.Host != "a" { |
| 648 | t.Fatalf("record = %+v, want build v2 held down in the mode Swap ran", r) |
| 649 | } |
| 650 | } |
| 651 | |
| 652 | // Carried from Task 4: every Pass prunes, a held-down instance whose route |
| 653 | // still names the failed build keeps the record, and a record no route |
| 654 | // installs any more goes. |
| 655 | func TestPassPrunesOnlyRecordsNoRouteInstalls(t *testing.T) { |
| 656 | h := newHarness(t, "a", "b") |
| 657 | h.book.Record(FailureRecord{Account: "bab", App: "blog", Build: "v2", Host: "a", HeldDown: []string{"a"}, PromotedAt: promotedAt, At: t0}) |
| 658 | h.book.Record(FailureRecord{Account: "bab", App: "blog", Build: "v0", Host: "b", PromotedAt: promotedAt, At: t0}) |
| 659 | h.install("v2", "a") |
| 660 | h.pass() |
| 661 | if _, ok := h.book.Lookup("bab", "blog", "v0"); ok { |
| 662 | t.Fatal("a record no route installs survived the pass") |
| 663 | } |
| 664 | if !h.book.HeldDown("bab", "blog", "v2", "a") { |
| 665 | t.Fatal("the pass pruned a held-down instance's record while its route still names the build") |
| 666 | } |
| 667 | h.install("v1", "a") // rolled back by hand |
| 668 | h.pass() |
| 669 | if len(h.book.List()) != 0 { |
| 670 | t.Fatalf("records = %+v after the last route left v2", h.book.List()) |
| 671 | } |
| 672 | } |
| 673 | |
| 674 | // Carried from Task 6, spec §8.2/§8.4: an install by hand that clears the |
| 675 | // record lifts the halt with it. A halt is in memory and outlived its record, |
| 676 | // so installing the build again later was refused for good. |
| 677 | func TestAnInstallByHandThatClearsTheRecordLiftsTheHalt(t *testing.T) { |
| 678 | h := newHarness(t, "a", "b") |
| 679 | h.stage("v2") |
| 680 | h.failing.Store(true) |
| 681 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 682 | if a.Build == "v2" && h.failing.Load() { |
| 683 | return activator.SwapFailed |
| 684 | } |
| 685 | return activator.SwapDone |
| 686 | } |
| 687 | h.install("v2", "a", "b") |
| 688 | h.pass() |
| 689 | waitFor(t, "a failure record", func() bool { _, ok := h.book.Lookup("bab", "blog", "v2"); return ok }) |
| 690 | h.install("v1", "a", "b") |
| 691 | h.pass() |
| 692 | if len(h.book.List()) != 0 { |
| 693 | t.Fatal("the record outlived every route installing the build") |
| 694 | } |
| 695 | h.failing.Store(false) |
| 696 | h.install("v2", "a", "b") |
| 697 | h.pass() |
| 698 | waitFor(t, "every host on v2", func() bool { return h.servingAll("v2", "a", "b") }) |
| 699 | calls, _, _ := h.act.snapshot() |
| 700 | if c := calls[1].a; !c.Canary { |
| 701 | t.Fatalf("first attempt after the hold lifted = %+v, want a fresh canary", c) |
| 702 | } |
| 703 | } |
| 704 | |
| 705 | // Carried from Task 6: `carlos instances retry` runs that host as a canary |
| 706 | // through its cohort's halt, and only that host. |
| 707 | func TestRetryHostRunsACanaryThroughAHalt(t *testing.T) { |
| 708 | h := newHarness(t, "a", "b") |
| 709 | h.stage("v2") |
| 710 | h.failing.Store(true) |
| 711 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 712 | if a.Build == "v2" && h.failing.Load() { |
| 713 | return activator.SwapFailed |
| 714 | } |
| 715 | return activator.SwapDone |
| 716 | } |
| 717 | h.install("v2", "a", "b") |
| 718 | h.pass() |
| 719 | waitFor(t, "a failure record", func() bool { _, ok := h.book.Lookup("bab", "blog", "v2"); return ok }) |
| 720 | h.failing.Store(false) |
| 721 | if !h.retry("b", "v2") { |
| 722 | t.Fatal("RetryHost did not clear b") |
| 723 | } |
| 724 | waitFor(t, "b on v2", func() bool { return h.servingAll("v2", "b") }) |
| 725 | calls, _, _ := h.act.snapshot() |
| 726 | if c := calls[1].a; c.Host != "b" || !c.Canary || c.Cause != activator.CauseRetry { |
| 727 | t.Fatalf("attempt after the host retry = %+v, want b's retry canary", c) |
| 728 | } |
| 729 | h.pass() |
| 730 | time.Sleep(50 * time.Millisecond) |
| 731 | if h.servingAll("v2", "a") { |
| 732 | t.Fatal("a host retry lifted the halt for another host") |
| 733 | } |
| 734 | } |
| 735 | |
| 736 | // Carried from Task 6: a failed config attempt backs off even when its row |
| 737 | // cannot be read afterwards; the backoff is set in the step that frees the |
| 738 | // slot, so nothing can re-admit the attempt before it. The row is unreadable |
| 739 | // only for the read that follows the failure: the next attempt's own read |
| 740 | // succeeds, so only the backoff can keep it from running. |
| 741 | func TestAFailedConfigAttemptBacksOffWithItsRowUnreadable(t *testing.T) { |
| 742 | h := newHarness(t, "a") |
| 743 | var failed, refused atomic.Bool |
| 744 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 745 | if a.Cause == activator.CauseConfig { |
| 746 | failed.Store(true) |
| 747 | return activator.SwapFailed |
| 748 | } |
| 749 | return activator.SwapDone |
| 750 | } |
| 751 | h.rows.mu.Lock() |
| 752 | h.rows.getErr = func(string) error { |
| 753 | if failed.Load() && refused.CompareAndSwap(false, true) { |
| 754 | return errors.New("registry: database is locked") |
| 755 | } |
| 756 | return nil |
| 757 | } |
| 758 | h.rows.mu.Unlock() |
| 759 | h.pass() |
| 760 | h.m.ConfigDelivered("a", "sha256:c1") |
| 761 | waitFor(t, "the first config attempt", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) >= 1 }) |
| 762 | h.pass() |
| 763 | time.Sleep(50 * time.Millisecond) |
| 764 | if calls, _, _ := h.act.snapshot(); len(calls) != 1 { |
| 765 | t.Fatalf("%d config attempts with no backoff after an unreadable row", len(calls)) |
| 766 | } |
| 767 | } |
| 768 | |
| 769 | // Carried from Task 7: activating an older revision than the delivered one |
| 770 | // leaves the delivered revision's failure count, and so its backoff, alone. |
| 771 | func TestAnOlderActivationKeepsTheConfigFailureCount(t *testing.T) { |
| 772 | h := newHarness(t, "a") |
| 773 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 774 | if a.Cause == activator.CauseConfig { |
| 775 | return activator.SwapFailed |
| 776 | } |
| 777 | return activator.SwapDone |
| 778 | } |
| 779 | h.pass() |
| 780 | h.m.ConfigDelivered("a", "sha256:c2") |
| 781 | waitFor(t, "the pending event", func() bool { return h.events.has(configPendingHead) }) |
| 782 | h.m.configActivated("a", "sha256:c1") |
| 783 | h.m.mu.Lock() |
| 784 | fails := h.m.cfg["a"].fails |
| 785 | h.m.mu.Unlock() |
| 786 | if fails != 1 { |
| 787 | t.Fatalf("fails = %d after an older revision activated, want 1", fails) |
| 788 | } |
| 789 | } |
| 790 | |
| 791 | // Carried from Task 6, spec §8.1: an instance that goes to sleep with a |
| 792 | // config attempt pending needs no attempt any more. Its next wake starts with |
| 793 | // the delivered file, so the revision is active, and the held ack goes out as |
| 794 | // a sleeper's does, marked Hibernating. After it wakes, no config attempt |
| 795 | // restarts a generation that already runs the revision. |
| 796 | func TestASleepersPendingConfigIsNoAttemptAfterItWakes(t *testing.T) { |
| 797 | h := newHarness(t, "a") |
| 798 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 799 | if a.Cause == activator.CauseConfig { |
| 800 | return activator.SwapFailed |
| 801 | } |
| 802 | return activator.SwapDone |
| 803 | } |
| 804 | h.pass() |
| 805 | var mu sync.Mutex |
| 806 | var acks []appconfig.ConfigAck |
| 807 | ack := h.m.DeferAck(func(_ context.Context, a appconfig.ConfigAck) { |
| 808 | mu.Lock() |
| 809 | defer mu.Unlock() |
| 810 | acks = append(acks, a) |
| 811 | }) |
| 812 | h.m.ConfigDelivered("a", "sha256:c1") |
| 813 | ack(appconfig.ConfigAck{Host: "a", Restarted: true}) |
| 814 | waitFor(t, "the failed config attempt", func() bool { return h.events.has(configPendingHead) }) |
| 815 | h.act.mu.Lock() |
| 816 | delete(h.act.gens, "a") // hibernated inside its backoff |
| 817 | h.act.mu.Unlock() |
| 818 | h.pass() |
| 819 | waitFor(t, "the held ack", func() bool { mu.Lock(); defer mu.Unlock(); return len(acks) == 1 }) |
| 820 | mu.Lock() |
| 821 | if !acks[0].Hibernating || acks[0].Restarted { |
| 822 | t.Fatalf("ack = %+v, want a sleeper's ack, Hibernating", acks[0]) |
| 823 | } |
| 824 | mu.Unlock() |
| 825 | h.act.serve("a", "v1") // woke, with the delivered file |
| 826 | h.advance(2 * time.Minute) |
| 827 | h.pass() |
| 828 | time.Sleep(50 * time.Millisecond) |
| 829 | if calls, _, _ := h.act.snapshot(); len(calls) != 1 { |
| 830 | t.Fatalf("calls = %+v: a woken sleeper ran its old pending config attempt", calls) |
| 831 | } |
| 832 | } |
| 833 | |
| 834 | // Spec §8.1, fix round 1: a config attempt pending for c1 while the instance |
| 835 | // was awake is no attempt any more once c2 is delivered to it asleep, even |
| 836 | // if it wakes before the scheduler ever saw it asleep (configSlept never |
| 837 | // fires). The wake starts with c2. The stale c1 attempt would restart a |
| 838 | // generation already running c2 and then mark c1 active. |
| 839 | func TestAConfigDeliveredToASleeperClearsTheOlderPendingAttempt(t *testing.T) { |
| 840 | h := newHarness(t, "a") |
| 841 | h.act.outcome = func(a activator.Attempt) activator.SwapOutcome { |
| 842 | if a.Cause == activator.CauseConfig { |
| 843 | return activator.SwapFailed |
| 844 | } |
| 845 | return activator.SwapDone |
| 846 | } |
| 847 | h.pass() |
| 848 | h.m.ConfigDelivered("a", "sha256:c1") |
| 849 | waitFor(t, "the failed c1 attempt", func() bool { return h.events.has(configPendingHead) }) |
| 850 | s := h.m.schedulerOf("a") |
| 851 | waitFor(t, "the slot to free with c1 pending", func() bool { |
| 852 | s.mu.Lock() |
| 853 | defer s.mu.Unlock() |
| 854 | hs := s.q.hosts["a"] |
| 855 | return hs != nil && hs.running == nil && hs.wantConfig == "sha256:c1" |
| 856 | }) |
| 857 | h.act.mu.Lock() |
| 858 | delete(h.act.gens, "a") // hibernated inside the backoff; no scheduler iteration runs |
| 859 | h.act.mu.Unlock() |
| 860 | h.m.ConfigDelivered("a", "sha256:c2") |
| 861 | h.act.serve("a", "v1") // woke at once, with c2 in the file |
| 862 | h.m.configActivated("a", "sha256:c2") |
| 863 | h.advance(2 * time.Minute) |
| 864 | h.pass() |
| 865 | time.Sleep(50 * time.Millisecond) |
| 866 | if calls, _, _ := h.act.snapshot(); len(calls) != 1 { |
| 867 | t.Fatalf("calls = %+v: a stale c1 attempt ran against a generation already on c2", calls) |
| 868 | } |
| 869 | h.m.mu.Lock() |
| 870 | tr := h.m.track("a") |
| 871 | delivered, active := tr.delivered, tr.active |
| 872 | h.m.mu.Unlock() |
| 873 | if delivered != "sha256:c2" || active != "sha256:c2" { |
| 874 | t.Fatalf("delivered %q, active %q: want both c2", delivered, active) |
| 875 | } |
| 876 | } |
| 877 | |
| 878 | // Spec §8.4: a record stays while any instance is on the failed build. A |
| 879 | // held-down instance whose row is "draining" during a Pass is not managed |
| 880 | // for that Pass, but it still names the build: the record and the halt stay. |
| 881 | func TestPruneKeepsTheRecordOfAHeldDownHostWhoseRowLeavesReady(t *testing.T) { |
| 882 | h := newHarness(t, "a") |
| 883 | h.stage("v2") |
| 884 | h.book.Record(FailureRecord{Account: "bab", App: "blog", Build: "v2", Host: "a", HeldDown: []string{"a"}, PromotedAt: promotedAt, At: t0}) |
| 885 | r := h.rows.get("a") |
| 886 | r.Version, r.Status = "v2", "draining" |
| 887 | h.rows.put(r) |
| 888 | h.pass() |
| 889 | if !h.book.HeldDown("bab", "blog", "v2", "a") { |
| 890 | t.Fatal("a Pass dropped the record of a held-down host whose row names the build") |
| 891 | } |
| 892 | r.Status = "ready" |
| 893 | h.rows.put(r) |
| 894 | h.pass() |
| 895 | if !h.book.HeldDown("bab", "blog", "v2", "a") { |
| 896 | t.Fatal("the held-down host lost its record when its row returned to ready") |
| 897 | } |
| 898 | } |
| 899 | |
| 900 | // Plan 3's deferral (O5): an attempt abandoned while it is still the host's |
| 901 | // newest (a hold whose stop withdrew, a lost re-check) waits before it runs |
| 902 | // again, rather than looping stop cycles. One abandoned by a newer desire |
| 903 | // does not wait (TestTwoReleasesTwentySecondsApartNeverSwitchToTheFirst). |
| 904 | func TestAnAbandonedAttemptStillNewestWaitsBeforeItRunsAgain(t *testing.T) { |
| 905 | h := newHarness(t, "a") |
| 906 | h.stage("v2") |
| 907 | h.act.outcome = func(activator.Attempt) activator.SwapOutcome { return activator.SwapAbandoned } |
| 908 | h.install("v2", "a") |
| 909 | h.pass() |
| 910 | waitFor(t, "the first attempt", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) >= 1 }) |
| 911 | h.pass() |
| 912 | time.Sleep(50 * time.Millisecond) |
| 913 | if calls, _, _ := h.act.snapshot(); len(calls) != 1 { |
| 914 | t.Fatalf("%d attempts: an abandoned attempt ran again at once", len(calls)) |
| 915 | } |
| 916 | h.advance(abandonDeferral + time.Second) |
| 917 | h.pass() |
| 918 | waitFor(t, "the attempt after the wait", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) == 2 }) |
| 919 | if len(h.book.List()) != 0 { |
| 920 | t.Fatal("an abandoned attempt wrote a failure record") |
| 921 | } |
| 922 | } |
| 923 | |
| 924 | // PF5 (plan 3's O8): member-facing events carry the reason word, never the |
| 925 | // swap's diagnostic text, which can name paths and internals. |
| 926 | func TestMemberEventsCarryOnlyTheReasonWord(t *testing.T) { |
| 927 | h := newHarness(t, "a") |
| 928 | h.stage("v2") |
| 929 | h.act.reason = "spawn: fork/exec /opt/carlos/versions/bab/v2/blog: permission denied" |
| 930 | h.act.outcome = func(activator.Attempt) activator.SwapOutcome { return activator.SwapFailed } |
| 931 | h.install("v2", "a") |
| 932 | h.pass() |
| 933 | waitFor(t, "the halt event", func() bool { |
| 934 | return h.events.has(textHead(textCohortHalted("v2", "\x00", status.ReasonHealth), "\x00")) |
| 935 | }) |
| 936 | waitFor(t, "the failed event", func() bool { return h.events.has(textHead(textSwapFailed("v2", "\x00", status.ReasonHealth), "\x00")) }) |
| 937 | for _, raw := range []string{"permission denied", "fork/exec", "blog:"} { |
| 938 | if h.events.has(raw) { |
| 939 | t.Fatalf("a member event carries the raw reason %q", raw) |
| 940 | } |
| 941 | } |
| 942 | if !h.events.has(textSwapFailed("v2", "v1", status.ReasonHealth)) || |
| 943 | !h.events.has(textCohortHalted("v2", "box-1", status.ReasonHealth)) { |
| 944 | t.Fatal("the events do not carry the reason word") |
| 945 | } |
| 946 | } |
| 947 | |
| 948 | // Astra R3 §5 item 4 (execution admission): an attempt is admitted from the |
| 949 | // pass's view of a row, and the row may change before the scheduler runs it. |
| 950 | // execute revalidates the row it reads: a row that has become a named-unit |
| 951 | // route, or that now belongs to another app, is no longer this scheduler's, |
| 952 | // so the attempt ends abandoned with nothing run (no Swap, no ReplaceWorker, |
| 953 | // no record). Before the fix the attempt went ahead on the changed row. |
| 954 | func TestARowThatLeavesTheSchedulerBeforeDispatchIsAbandoned(t *testing.T) { |
| 955 | for _, worker := range []bool{false, true} { |
| 956 | for name, change := range map[string]func(*registry.Route){ |
| 957 | "a named-unit route": func(r *registry.Route) { r.Unit = "blog.service" }, |
| 958 | "another app": func(r *registry.Route) { r.App = "shop" }, |
| 959 | } { |
| 960 | kind := "socket instance" |
| 961 | if worker { |
| 962 | kind = "worker" |
| 963 | } |
| 964 | t.Run(kind+"/"+name, func(t *testing.T) { |
| 965 | var h *harness |
| 966 | if worker { |
| 967 | h = workerHarness(t, "h") |
| 968 | } else { |
| 969 | h = newHarness(t, "h") |
| 970 | } |
| 971 | var ( |
| 972 | mu sync.Mutex |
| 973 | abandoned []string |
| 974 | ) |
| 975 | h.m.d.Logf = func(format string, args ...any) { |
| 976 | if line := fmt.Sprintf(format, args...); strings.Contains(line, "abandoned") { |
| 977 | mu.Lock() |
| 978 | abandoned = append(abandoned, line) |
| 979 | mu.Unlock() |
| 980 | } |
| 981 | } |
| 982 | h.stage("v2") |
| 983 | h.install("v2", "h") |
| 984 | // The row changes between the pass that queued the attempt and |
| 985 | // the scheduler's read of it. |
| 986 | h.rows.mu.Lock() |
| 987 | h.rows.getHook = func(host string) { |
| 988 | r := h.rows.get(host) |
| 989 | change(&r) |
| 990 | h.rows.put(r) |
| 991 | } |
| 992 | h.rows.mu.Unlock() |
| 993 | moves := func() int { |
| 994 | calls, _, _ := h.act.snapshot() |
| 995 | return len(calls) + len(h.act.replacements()) |
| 996 | } |
| 997 | h.pass() |
| 998 | waitFor(t, "the attempt to end", func() bool { |
| 999 | mu.Lock() |
| 1000 | n := len(abandoned) |
| 1001 | mu.Unlock() |
| 1002 | return (n > 0 || moves() > 0) && !h.m.attemptRunning("h") |
| 1003 | }) |
| 1004 | if n := moves(); n != 0 { |
| 1005 | t.Fatalf("%d Swap/ReplaceWorker calls on a row that left the scheduler, want none", n) |
| 1006 | } |
| 1007 | mu.Lock() |
| 1008 | got := abandoned[0] |
| 1009 | mu.Unlock() |
| 1010 | if !strings.Contains(got, "route no longer managed") { |
| 1011 | t.Fatalf("abandoned as %q, want the route no longer managed", got) |
| 1012 | } |
| 1013 | if recs := h.book.List(); len(recs) != 0 { |
| 1014 | t.Fatalf("records %+v for an attempt that never ran", recs) |
| 1015 | } |
| 1016 | }) |
| 1017 | } |
| 1018 | } |
| 1019 | } |
| 1020 | |
| 1021 | func TestReasonWords(t *testing.T) { |
| 1022 | for text, want := range map[string]string{ |
| 1023 | "health check failed": status.ReasonHealth, |
| 1024 | "swap: successor exited during its settle": status.ReasonSettle, |
| 1025 | "exited within 10s of becoming healthy": status.ReasonExit, |
| 1026 | "staging: digest mismatch": status.ReasonStaging, |
| 1027 | "version name reused with different content, ship under a new name": status.ReasonReused, |
| 1028 | "GET / answered 500": status.ReasonHealth, |
| 1029 | "": status.ReasonHealth, |
| 1030 | } { |
| 1031 | if got := reasonWord(text); got != want { |
| 1032 | t.Errorf("reasonWord(%q) = %q, want %q", text, got, want) |
| 1033 | } |
| 1034 | } |
| 1035 | } |
| 1036 | |
| 1037 | // configPendingHead is the config-pending event up to its reason: the tests |
| 1038 | // that wait for it do not know the reason or the backoff. |
| 1039 | var configPendingHead = textHead(rolloutwords.EventConfigPending, "%s") |
| 1040 | |
| 1041 | // Task 11 rule 3 (Astra R1 Flag triage T8): the held ack an activation |
| 1042 | // releases is handed to the manager's writer, not written by done(). A store |
| 1043 | // write that never returns cannot keep the attempt's slot, so the host's |
| 1044 | // next attempt is admitted while the writer is still blocked. |
| 1045 | func TestABlockedAckWriterDoesNotHoldTheSlot(t *testing.T) { |
| 1046 | h := newHarness(t, "a") |
| 1047 | h.pass() |
| 1048 | h.stage("v2") |
| 1049 | entered, block := make(chan struct{}, 4), make(chan struct{}) |
| 1050 | t.Cleanup(func() { close(block) }) |
| 1051 | ack := h.m.DeferAck(func(context.Context, appconfig.ConfigAck) { |
| 1052 | entered <- struct{}{} |
| 1053 | <-block |
| 1054 | }) |
| 1055 | h.m.ConfigDelivered("a", "sha256:c1") |
| 1056 | ack(appconfig.ConfigAck{Host: "a"}) |
| 1057 | select { |
| 1058 | case <-entered: |
| 1059 | case <-time.After(5 * time.Second): |
| 1060 | t.Fatal("the config attempt's ack was never written") |
| 1061 | } |
| 1062 | h.install("v2", "a") |
| 1063 | h.pass() |
| 1064 | waitFor(t, "the release attempt, admitted while the ack writer blocks", func() bool { return h.servingAll("v2", "a") }) |
| 1065 | } |
| 1066 | |
| 1067 | // Task 11 rule 2 (Astra R1 Flag triage T10): the failing pointer a wake's |
| 1068 | // plan reads runs under the wake's context. A store read that never answers |
| 1069 | // ends when the wake is cancelled, and the plan goes on with no PromotedAt |
| 1070 | // (the failure's own time stands in, the safe direction). |
| 1071 | func TestABlockedPointerReadEndsWithTheWake(t *testing.T) { |
| 1072 | h := newHarness(t, "a") |
| 1073 | h.stage("v2") |
| 1074 | h.install("v2", "a") |
| 1075 | reading := make(chan struct{}) |
| 1076 | var once sync.Once |
| 1077 | h.m.d.Pointer = func(ctx context.Context, _, _, _ string) (*release.ChannelPointer, error) { |
| 1078 | once.Do(func() { close(reading) }) |
| 1079 | <-ctx.Done() |
| 1080 | return nil, ctx.Err() |
| 1081 | } |
| 1082 | ctx, cancel := context.WithCancel(context.Background()) |
| 1083 | defer cancel() |
| 1084 | plans := make(chan activator.WakePlan, 1) |
| 1085 | go func() { plans <- h.m.Plan(ctx, h.rows.get("a")) }() |
| 1086 | <-reading |
| 1087 | cancel() |
| 1088 | select { |
| 1089 | case p := <-plans: |
| 1090 | if !p.FirstWake || !p.FailedUnder.IsZero() { |
| 1091 | t.Fatalf("plan %+v: want the first wake of v2 with no PromotedAt", p) |
| 1092 | } |
| 1093 | case <-time.After(2 * time.Second): |
| 1094 | t.Fatal("the plan's pointer read outlived the wake's context") |
| 1095 | } |
| 1096 | } |
| 1097 | |
| 1098 | // Task 11 Corrections (Task 5 review; spec §8.6): a planned agent restart |
| 1099 | // stops the activator first, which cancels a swap mid-flight, and then joins |
| 1100 | // the scheduler. Close returns only once that attempt's done() has written |
| 1101 | // its outcome (here the hold-ran marker of the hold build that ran) and |
| 1102 | // ended its intent, so the restart reconciles nothing. After Close no |
| 1103 | // attempt is admitted. |
| 1104 | func TestCloseJoinsAnAttemptCancelledByShutdown(t *testing.T) { |
| 1105 | h := newHarness(t, "a") |
| 1106 | dir := h.onDisk() |
| 1107 | h.swap = "hold" |
| 1108 | h.stage("v2", "v3") |
| 1109 | gate := h.act.hold("a") |
| 1110 | h.act.outcome = func(activator.Attempt) activator.SwapOutcome { |
| 1111 | time.Sleep(200 * time.Millisecond) // the swap winding down after its entry stopped |
| 1112 | return activator.SwapAbandoned |
| 1113 | } |
| 1114 | h.act.executed = func(activator.Attempt, activator.SwapOutcome) bool { return true } // the hold build ran |
| 1115 | h.install("v2", "a") |
| 1116 | h.pass() |
| 1117 | waitFor(t, "the hold swap to start", func() bool { _, s, _ := h.act.snapshot(); return len(s) == 1 }) |
| 1118 | if n := len(fileOf(t, dir).Intents); n != 1 { |
| 1119 | t.Fatalf("%d intents on disk while the swap runs, want 1", n) |
| 1120 | } |
| 1121 | close(gate) // svc.Shutdown stops the entry under the swap |
| 1122 | ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) |
| 1123 | defer cancel() |
| 1124 | h.m.Close(ctx) |
| 1125 | f := fileOf(t, dir) |
| 1126 | if len(f.Intents) != 0 { |
| 1127 | t.Fatalf("intents %+v on disk after Close: a restart would reconcile the attempt as a failure", f.Intents) |
| 1128 | } |
| 1129 | if mk, ok := f.HoldRan["a"]; !ok || mk.Build != "v2" { |
| 1130 | t.Fatalf("hold-ran markers %+v after Close, want a's for v2: the attempt's outcome", f.HoldRan) |
| 1131 | } |
| 1132 | h.advance(time.Hour) |
| 1133 | h.install("v3", "a") |
| 1134 | h.pass() |
| 1135 | time.Sleep(100 * time.Millisecond) |
| 1136 | if _, s, _ := h.act.snapshot(); len(s) != 1 { |
| 1137 | t.Fatalf("swaps started %v: an attempt was admitted after Close", s) |
| 1138 | } |
| 1139 | } |
| 1140 | |
| 1141 | // Task 10 review minor (1), restored in Task 11: the pointer pass requeues a |
| 1142 | // cleared build at once, before it reads the app's next channel, as the |
| 1143 | // armed pass did before the pointer work was shared (onCleared). Both |
| 1144 | // channels re-sign v2, so the second read always follows a clear. |
| 1145 | // |
| 1146 | // Only the pass's reads are counted (fix wave 1, item 6). The requeued |
| 1147 | // retry's attempt runs on the scheduler's goroutine and reads the pointer |
| 1148 | // too, for its FailedUnder (scheduler.execute, failedUnder); counted, it |
| 1149 | // made a third read about once in a hundred runs. |
| 1150 | func TestAPointerClearRequeuesBeforeTheNextChannelIsRead(t *testing.T) { |
| 1151 | h := newHarness(t, "a", "b") |
| 1152 | h.stage("v2") |
| 1153 | rb := h.rows.get("b") |
| 1154 | rb.Channel = "beta" |
| 1155 | h.rows.put(rb) |
| 1156 | h.book.Record(FailureRecord{Account: "bab", App: "blog", Build: "v2", Host: "c", PromotedAt: promotedAt, At: t0}) |
| 1157 | h.install("v2", "a", "b") |
| 1158 | h.pass() |
| 1159 | h.pointer("stable", "v2", promotedAt.Add(time.Minute)) |
| 1160 | h.pointer("beta", "v2", promotedAt.Add(time.Minute)) |
| 1161 | var reads atomic.Int32 |
| 1162 | var requeuedLate atomic.Bool |
| 1163 | h.mu.Lock() |
| 1164 | h.pointerHook = func() { |
| 1165 | if !calledFrom("(*Manager).applyPointers") { |
| 1166 | return |
| 1167 | } |
| 1168 | if reads.Add(1) != 2 { |
| 1169 | return |
| 1170 | } |
| 1171 | h.m.mu.Lock() |
| 1172 | n := len(h.m.retried) |
| 1173 | h.m.mu.Unlock() |
| 1174 | if n == 0 { |
| 1175 | requeuedLate.Store(true) |
| 1176 | } |
| 1177 | } |
| 1178 | h.mu.Unlock() |
| 1179 | h.advance(pointerReadEvery) |
| 1180 | h.pass() |
| 1181 | if reads.Load() != 2 { |
| 1182 | t.Fatalf("%d pointer reads, want 2 (stable and beta)", reads.Load()) |
| 1183 | } |
| 1184 | if requeuedLate.Load() { |
| 1185 | t.Fatal("the cleared build was requeued only after every channel was read") |
| 1186 | } |
| 1187 | waitFor(t, "every host on v2", func() bool { return h.servingAll("v2", "a", "b") }) |
| 1188 | } |
| 1189 | |
| 1190 | // Task 11 Corrections (plan review #10): acks are ordered at the transition. |
| 1191 | // The writer keeps one pending ack per host, the newest by sequence; an older |
| 1192 | // one is dropped, never written after a newer one; hosts are written in the |
| 1193 | // order they were first queued; each write is bounded by ackWriteBound. |
| 1194 | func TestAcksAreWrittenNewestLastPerHost(t *testing.T) { |
| 1195 | ctx, cancel := context.WithCancel(context.Background()) |
| 1196 | defer cancel() |
| 1197 | w := newAckWriter(ctx, func(string, ...any) {}) |
| 1198 | var mu sync.Mutex |
| 1199 | var got []string |
| 1200 | write := func(tag string) func(context.Context) { |
| 1201 | return func(ctx context.Context) { |
| 1202 | if dl, ok := ctx.Deadline(); !ok || time.Until(dl) > ackWriteBound { |
| 1203 | t.Errorf("ack %s written with deadline %v (%v), want at most ackWriteBound", tag, dl, ok) |
| 1204 | } |
| 1205 | mu.Lock() |
| 1206 | got = append(got, tag) |
| 1207 | mu.Unlock() |
| 1208 | } |
| 1209 | } |
| 1210 | entered, block := make(chan struct{}), make(chan struct{}) |
| 1211 | first := write("a1") |
| 1212 | w.submit("a", ackJob{seq: 1, write: func(ctx context.Context) { close(entered); <-block; first(ctx) }}) |
| 1213 | <-entered |
| 1214 | w.submit("a", ackJob{seq: 3, write: write("a3")}) |
| 1215 | w.submit("a", ackJob{seq: 2, write: write("a2")}) // older than the pending a3 |
| 1216 | w.submit("b", ackJob{seq: 4, write: write("b4")}) |
| 1217 | w.submit("a", ackJob{seq: 5, write: write("a5")}) // replaces a3, keeps a's place |
| 1218 | close(block) |
| 1219 | waitFor(t, "three writes", func() bool { mu.Lock(); defer mu.Unlock(); return len(got) == 3 }) |
| 1220 | w.submit("a", ackJob{seq: 2, write: write("a2-late")}) // older than one written |
| 1221 | time.Sleep(50 * time.Millisecond) |
| 1222 | mu.Lock() |
| 1223 | defer mu.Unlock() |
| 1224 | if want := []string{"a1", "a5", "b4"}; strings.Join(got, ",") != strings.Join(want, ",") { |
| 1225 | t.Fatalf("written %v, want %v", got, want) |
| 1226 | } |
| 1227 | } |
| 1228 | |
| 1229 | // Task 11 Corrections (spec §8.6): Close also joins a wake event under way. |
| 1230 | // A terminal OnGen writes the wake's outcome and ends its intent; Close does |
| 1231 | // not return while it is still running. |
| 1232 | func TestCloseWaitsForAnOnGenUnderWay(t *testing.T) { |
| 1233 | h := newHarness(t, "a") |
| 1234 | h.pass() |
| 1235 | entered, release := make(chan struct{}), make(chan struct{}) |
| 1236 | var once sync.Once |
| 1237 | h.rows.mu.Lock() |
| 1238 | h.rows.setServingHook = func(string) { |
| 1239 | once.Do(func() { close(entered); <-release }) |
| 1240 | } |
| 1241 | h.rows.mu.Unlock() |
| 1242 | g, _ := h.act.Current("a") |
| 1243 | go h.m.OnGen(activator.GenEvent{Kind: activator.GenSettled, Route: h.rows.get("a"), Gen: g, FirstWake: true}) |
| 1244 | <-entered |
| 1245 | closed := make(chan struct{}) |
| 1246 | go func() { |
| 1247 | ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) |
| 1248 | defer cancel() |
| 1249 | h.m.Close(ctx) |
| 1250 | close(closed) |
| 1251 | }() |
| 1252 | select { |
| 1253 | case <-closed: |
| 1254 | t.Fatal("Close returned while a wake event was still being recorded") |
| 1255 | case <-time.After(100 * time.Millisecond): |
| 1256 | } |
| 1257 | close(release) |
| 1258 | select { |
| 1259 | case <-closed: |
| 1260 | case <-time.After(5 * time.Second): |
| 1261 | t.Fatal("Close did not return once the wake event finished") |
| 1262 | } |
| 1263 | } |
| 1264 | |
| 1265 | // Task 11 Corrections: closing the writer writes what is pending, each |
| 1266 | // write still bounded, and an ack submitted after it is dropped. |
| 1267 | func TestClosingTheAckWriterWritesWhatIsPending(t *testing.T) { |
| 1268 | ctx, cancel := context.WithCancel(context.Background()) |
| 1269 | defer cancel() |
| 1270 | w := newAckWriter(ctx, func(string, ...any) {}) |
| 1271 | var mu sync.Mutex |
| 1272 | var got []string |
| 1273 | write := func(tag string) func(context.Context) { |
| 1274 | return func(context.Context) { mu.Lock(); got = append(got, tag); mu.Unlock() } |
| 1275 | } |
| 1276 | entered, block := make(chan struct{}), make(chan struct{}) |
| 1277 | w.submit("a", ackJob{seq: 1, write: func(context.Context) { close(entered); <-block }}) |
| 1278 | <-entered |
| 1279 | w.submit("b", ackJob{seq: 2, write: write("b2")}) |
| 1280 | w.submit("c", ackJob{seq: 3, write: write("c3")}) |
| 1281 | go func() { time.Sleep(50 * time.Millisecond); close(block) }() |
| 1282 | cctx, ccancel := context.WithTimeout(context.Background(), 5*time.Second) |
| 1283 | defer ccancel() |
| 1284 | w.close(cctx) |
| 1285 | w.submit("d", ackJob{seq: 4, write: write("d4")}) |
| 1286 | time.Sleep(50 * time.Millisecond) |
| 1287 | mu.Lock() |
| 1288 | defer mu.Unlock() |
| 1289 | if strings.Join(got, ",") != "b2,c3" { |
| 1290 | t.Fatalf("written %v, want b2 and c3 written before close returned, and nothing after", got) |
| 1291 | } |
| 1292 | } |
| 1293 | |
| 1294 | // joinCount is Close's count of operations under way (test-only). |
| 1295 | func (m *Manager) joinCount() int { |
| 1296 | m.joinMu.Lock() |
| 1297 | defer m.joinMu.Unlock() |
| 1298 | return m.inflight |
| 1299 | } |
| 1300 | |
| 1301 | // Task 11 fix round 1, review finding 1 (spec §8.6): a wake already past |
| 1302 | // doWake's context check when the activator's Shutdown begins can still plan |
| 1303 | // and write its intent. Its terminal event (afterWake's, here a refusal at |
| 1304 | // its lease: nothing ran) reaches OnGen only after svc.Shutdown has |
| 1305 | // returned, because Shutdown waits for e.ready and afterWake runs after it. |
| 1306 | // Close waits for that event, so the wake's End is on disk when Close |
| 1307 | // returns, and the next agent reconciles nothing. |
| 1308 | func TestCloseJoinsAPlannedWakeWhoseEventArrivesAfterShutdown(t *testing.T) { |
| 1309 | h := newHarness(t, "a") |
| 1310 | dir := h.onDisk() |
| 1311 | h.stage("v2") |
| 1312 | h.registerThenInstall("v2", "a") |
| 1313 | h.wakeHost("a") |
| 1314 | p := h.m.Plan(context.Background(), h.rows.get("a")) |
| 1315 | if !p.FirstWake || p.Boot.Version != "v2" { |
| 1316 | t.Fatalf("plan %+v: want a first wake of v2", p) |
| 1317 | } |
| 1318 | if _, ok := fileOf(t, dir).Intents["a"]; !ok { |
| 1319 | t.Fatal("no intent on disk after the plan") |
| 1320 | } |
| 1321 | rt := h.rows.get("a") |
| 1322 | go func() { |
| 1323 | time.Sleep(200 * time.Millisecond) // afterWake, after Shutdown saw e.ready close |
| 1324 | h.m.OnGen(activator.GenEvent{Kind: activator.GenCancelled, Route: rt, Mode: p.Mode, FirstWake: true, |
| 1325 | Token: p.Token, ConfigRev: p.ConfigRev}) |
| 1326 | }() |
| 1327 | ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) |
| 1328 | defer cancel() |
| 1329 | h.m.Close(ctx) |
| 1330 | if f := fileOf(t, dir); len(f.Intents) != 0 { |
| 1331 | t.Fatalf("intents %+v on disk after Close: a restart would reconcile the wake as a failure", f.Intents) |
| 1332 | } |
| 1333 | } |
| 1334 | |
| 1335 | // Task 11 fix round 1, review finding 1: a planned wake is counted from the |
| 1336 | // intent its plan wrote until the one event that ends that intent, whatever |
| 1337 | // its kind; a plan that wrote no intent is never counted; and an event |
| 1338 | // delivered twice, or one for a wake that wrote nothing, never takes the |
| 1339 | // count below what is still under way. |
| 1340 | func TestAPlannedWakeIsCountedUntilItsIntentEnds(t *testing.T) { |
| 1341 | ctx := context.Background() |
| 1342 | setup := func(t *testing.T) (*harness, string, activator.WakePlan) { |
| 1343 | h := newHarness(t, "a") |
| 1344 | dir := h.onDisk() |
| 1345 | h.stage("v2") |
| 1346 | h.registerThenInstall("v2", "a") |
| 1347 | h.wakeHost("a") |
| 1348 | p := h.m.Plan(ctx, h.rows.get("a")) |
| 1349 | if !p.FirstWake || p.Boot.Version != "v2" { |
| 1350 | t.Fatalf("plan %+v: want a first wake of v2", p) |
| 1351 | } |
| 1352 | if n := h.m.joinCount(); n != 1 { |
| 1353 | t.Fatalf("count %d after a plan that wrote its intent, want 1", n) |
| 1354 | } |
| 1355 | return h, dir, p |
| 1356 | } |
| 1357 | ev := func(p activator.WakePlan, h *harness, kind activator.GenEventKind, first, ran bool) activator.GenEvent { |
| 1358 | e := activator.GenEvent{Kind: kind, Route: h.rows.get("a"), Gen: activator.Gen{N: 3, Build: "v2", Digest: "sha-v2"}, |
| 1359 | Mode: p.Mode, FirstWake: first, Token: p.Token, ConfigRev: p.ConfigRev, Executed: ran} |
| 1360 | if kind == activator.GenFailed || kind == activator.GenFellBack { |
| 1361 | e.Failed = p.Boot |
| 1362 | } |
| 1363 | if kind == activator.GenFellBack { |
| 1364 | e.Gen = activator.Gen{N: 4, Build: "v1", Digest: "sha-v1"} |
| 1365 | } |
| 1366 | return e |
| 1367 | } |
| 1368 | type step struct { |
| 1369 | kind activator.GenEventKind |
| 1370 | first bool |
| 1371 | ran bool |
| 1372 | count int // the count once the event has returned |
| 1373 | } |
| 1374 | for _, tc := range []struct { |
| 1375 | name string |
| 1376 | steps []step |
| 1377 | }{ |
| 1378 | {"settled", []step{{activator.GenHealthy, true, false, 1}, {activator.GenSettled, true, false, 0}}}, |
| 1379 | {"failed", []step{{activator.GenFailed, true, false, 0}}}, |
| 1380 | {"fell back", []step{{activator.GenFellBack, true, false, 0}}}, |
| 1381 | {"refused at its lease", []step{{activator.GenCancelled, true, false, 0}}}, |
| 1382 | {"stopped in its observation", []step{{activator.GenHealthy, true, false, 1}, {activator.GenCancelled, true, true, 0}}}, |
| 1383 | {"the good build booted untried", []step{{activator.GenHealthy, false, false, 0}}}, |
| 1384 | {"its terminal event twice", []step{{activator.GenCancelled, true, false, 0}, {activator.GenCancelled, true, false, 0}}}, |
| 1385 | } { |
| 1386 | t.Run(tc.name, func(t *testing.T) { |
| 1387 | h, dir, p := setup(t) |
| 1388 | for i, s := range tc.steps { |
| 1389 | h.m.OnGen(ev(p, h, s.kind, s.first, s.ran)) |
| 1390 | if n := h.m.joinCount(); n != s.count { |
| 1391 | t.Fatalf("after event %d (%s): count %d, want %d", i, s.kind, n, s.count) |
| 1392 | } |
| 1393 | } |
| 1394 | if _, stands := fileOf(t, dir).Intents["a"]; stands { |
| 1395 | t.Fatal("the intent stands after its terminal event") |
| 1396 | } |
| 1397 | }) |
| 1398 | } |
| 1399 | t.Run("a replaced wake and its replacement each end once", func(t *testing.T) { |
| 1400 | h, _, old := setup(t) |
| 1401 | cur := h.m.Plan(ctx, h.rows.get("a")) |
| 1402 | if n := h.m.joinCount(); n != 2 { |
| 1403 | t.Fatalf("count %d with two planned wakes, want 2", n) |
| 1404 | } |
| 1405 | h.m.OnGen(ev(old, h, activator.GenCancelled, true, false)) |
| 1406 | if n := h.m.joinCount(); n != 1 { |
| 1407 | t.Fatalf("count %d after the replaced wake's event, want 1", n) |
| 1408 | } |
| 1409 | h.m.OnGen(ev(cur, h, activator.GenCancelled, true, false)) |
| 1410 | if n := h.m.joinCount(); n != 0 { |
| 1411 | t.Fatalf("count %d after both events, want 0", n) |
| 1412 | } |
| 1413 | }) |
| 1414 | t.Run("a plan whose intent could not be written", func(t *testing.T) { |
| 1415 | h := newHarness(t, "a") |
| 1416 | dir := h.onDisk() |
| 1417 | h.stage("v2") |
| 1418 | h.registerThenInstall("v2", "a") |
| 1419 | h.wakeHost("a") |
| 1420 | readOnly(t, dir) |
| 1421 | p := h.m.Plan(ctx, h.rows.get("a")) |
| 1422 | if p.Refuse != BookUnreadableRefusal { |
| 1423 | t.Fatalf("plan %+v: want the book refusal", p) |
| 1424 | } |
| 1425 | if n := h.m.joinCount(); n != 0 { |
| 1426 | t.Fatalf("count %d after a plan that wrote nothing, want 0", n) |
| 1427 | } |
| 1428 | h.m.OnGen(activator.GenEvent{Kind: activator.GenCancelled, Route: h.rows.get("a"), FirstWake: true, Token: p.Token}) |
| 1429 | if n := h.m.joinCount(); n != 0 { |
| 1430 | t.Fatalf("count %d after an event for a wake that wrote nothing, want 0", n) |
| 1431 | } |
| 1432 | }) |
| 1433 | t.Run("a plan refused while the attempt's intent stands", func(t *testing.T) { |
| 1434 | h := newHarness(t, "a") |
| 1435 | h.onDisk() |
| 1436 | h.stage("v2") |
| 1437 | h.registerThenInstall("v2", "a") |
| 1438 | h.wakeHost("a") |
| 1439 | if err := h.book.Begin("a", Intent{Account: "bab", App: "blog", Build: "v3", Mode: "hold", ID: "a1"}); err != nil { |
| 1440 | t.Fatal(err) |
| 1441 | } |
| 1442 | if p := h.m.Plan(ctx, h.rows.get("a")); p.Refuse == "" { |
| 1443 | t.Fatalf("plan %+v: want a refusal", p) |
| 1444 | } |
| 1445 | if n := h.m.joinCount(); n != 0 { |
| 1446 | t.Fatalf("count %d after a refused plan, want 0", n) |
| 1447 | } |
| 1448 | }) |
| 1449 | t.Run("a plan under way", func(t *testing.T) { |
| 1450 | h := newHarness(t, "a") |
| 1451 | h.onDisk() |
| 1452 | h.stage("v2") |
| 1453 | h.registerThenInstall("v2", "a") |
| 1454 | h.wakeHost("a") |
| 1455 | entered, release := make(chan struct{}), make(chan struct{}) |
| 1456 | var once sync.Once |
| 1457 | h.mu.Lock() |
| 1458 | h.demandHook = func(ver string) { |
| 1459 | if ver == "v2" { |
| 1460 | once.Do(func() { close(entered); <-release }) |
| 1461 | } |
| 1462 | } |
| 1463 | h.mu.Unlock() |
| 1464 | planned := make(chan activator.WakePlan, 1) |
| 1465 | go func() { planned <- h.m.Plan(ctx, h.rows.get("a")) }() |
| 1466 | <-entered |
| 1467 | if n := h.m.joinCount(); n != 1 { |
| 1468 | t.Fatalf("count %d while a plan reads the demand, want 1", n) |
| 1469 | } |
| 1470 | closed := make(chan struct{}) |
| 1471 | go func() { |
| 1472 | cctx, cancel := context.WithTimeout(ctx, 5*time.Second) |
| 1473 | defer cancel() |
| 1474 | h.m.Close(cctx) |
| 1475 | close(closed) |
| 1476 | }() |
| 1477 | select { |
| 1478 | case <-closed: |
| 1479 | t.Fatal("Close returned while a plan was under way") |
| 1480 | case <-time.After(100 * time.Millisecond): |
| 1481 | } |
| 1482 | close(release) |
| 1483 | p := <-planned |
| 1484 | if n := h.m.joinCount(); n != 1 || !p.FirstWake { |
| 1485 | t.Fatalf("count %d once the plan (%+v) returned, want 1: its wake's", n, p) |
| 1486 | } |
| 1487 | h.m.OnGen(ev(p, h, activator.GenCancelled, true, false)) |
| 1488 | select { |
| 1489 | case <-closed: |
| 1490 | case <-time.After(5 * time.Second): |
| 1491 | t.Fatal("Close did not return once the wake ended") |
| 1492 | } |
| 1493 | }) |
| 1494 | } |
| 1495 | |
| 1496 | // calledFrom reports whether a function whose name ends in fn is on the |
| 1497 | // calling goroutine's stack. |
| 1498 | func calledFrom(fn string) bool { |
| 1499 | pc := make([]uintptr, 64) |
| 1500 | frames := runtime.CallersFrames(pc[:runtime.Callers(2, pc)]) |
| 1501 | for { |
| 1502 | f, more := frames.Next() |
| 1503 | if strings.HasSuffix(f.Function, fn) { |
| 1504 | return true |
| 1505 | } |
| 1506 | if !more { |
| 1507 | return false |
| 1508 | } |
| 1509 | } |
| 1510 | } |
| 1511 | |
amadan