| 1 | package rollout |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "runtime" |
| 6 | "strings" |
| 7 | "sync" |
| 8 | "sync/atomic" |
| 9 | "testing" |
| 10 | "time" |
| 11 | |
| 12 | "github.com/carlosframework/platform/internal/activator" |
| 13 | "github.com/carlosframework/platform/internal/release" |
| 14 | "github.com/carlosframework/platform/internal/status" |
| 15 | ) |
| 16 | |
| 17 | func TestStatusReportsHeldAndRollingInstances(t *testing.T) { |
| 18 | h := newHarness(t, "a", "b") |
| 19 | h.stage("v2") |
| 20 | gate := h.act.hold("a") |
| 21 | h.install("v2", "a", "b") |
| 22 | h.pass() |
| 23 | waitFor(t, "a's canary", func() bool { _, s, _ := h.act.snapshot(); return len(s) == 1 }) |
| 24 | // Recorded after the pass: a pass prunes the records of builds no route |
| 25 | // installs, and v3 is installed nowhere. |
| 26 | if _, err := h.book.Record(FailureRecord{Account: "bab", App: "blog", Build: "v3", Host: "z", Reason: "health check failed", At: t0}); err != nil { |
| 27 | t.Fatal(err) |
| 28 | } |
| 29 | r := h.m.Status() |
| 30 | if got := r.Hosts["a"]; got.Attempt != status.AttemptSettling || got.AttemptBuild != "v2" || got.Installed != "v2" || got.Serving != "v1" || !got.Awake { |
| 31 | t.Fatalf("a = %+v, want the canary settling onto v2", got) |
| 32 | } |
| 33 | if got := r.Hosts["b"]; got.Attempt != status.AttemptQueued || got.AttemptBuild != "v2" { |
| 34 | t.Fatalf("b = %+v, want queued behind the canary", got) |
| 35 | } |
| 36 | if app := r.Apps["bab/blog"]; app.Settling != 1 || app.Queued != 1 || app.Running != 0 { |
| 37 | t.Fatalf("app = %+v", app) |
| 38 | } |
| 39 | if len(r.Failures) != 1 || r.Failures[0].Build != "v3" || r.Failures[0].Reason != status.ReasonHealth { |
| 40 | t.Fatalf("failures = %+v", r.Failures) |
| 41 | } |
| 42 | close(gate) |
| 43 | } |
| 44 | |
| 45 | func TestAFailedFirstWakeShowsAsHeld(t *testing.T) { |
| 46 | h := newHarness(t, "a") |
| 47 | h.act.mu.Lock() |
| 48 | delete(h.act.gens, "a") // asleep: v2 is a first wake's build |
| 49 | h.act.mu.Unlock() |
| 50 | h.stage("v2") |
| 51 | h.install("v2", "a") |
| 52 | h.pass() |
| 53 | h.act.serve("a", "v1") // the first wake failed and the good build serves in its place |
| 54 | cur, _ := h.act.Current("a") |
| 55 | h.m.OnGen(activator.GenEvent{Kind: activator.GenFellBack, Route: h.rows.get("a"), FirstWake: true, |
| 56 | Mode: activator.ModeOverlap, Gen: cur, Failed: release.StagedIdentity{Version: "v2"}, |
| 57 | Reason: "exited within 10s of becoming healthy"}) |
| 58 | h.pass() |
| 59 | got := h.m.Status().Hosts["a"] |
| 60 | if got.Attempt != status.AttemptFailed || got.AttemptBuild != "v2" || got.HeldBy != "v2" || got.Reason != status.ReasonExit || got.Serving != "v1" { |
| 61 | t.Fatalf("a = %+v", got) |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | // A hold build that failed on the instance is held down: its wakes are |
| 66 | // refused, which the report calls down. |
| 67 | func TestAHeldDownInstanceShowsAsDown(t *testing.T) { |
| 68 | h := newHarness(t, "a") |
| 69 | h.stage("v2") |
| 70 | if _, err := h.book.Record(FailureRecord{Account: "bab", App: "blog", Build: "v2", Host: "a", Mode: "hold", |
| 71 | Reason: "swap: successor exited during its settle", At: t0, HeldDown: []string{"a"}}); err != nil { |
| 72 | t.Fatal(err) |
| 73 | } |
| 74 | h.install("v2", "a") |
| 75 | h.pass() |
| 76 | got := h.m.Status().Hosts["a"] |
| 77 | if got.Attempt != status.AttemptDown || got.HeldBy != "v2" || got.Reason != status.ReasonSettle { |
| 78 | t.Fatalf("a = %+v", got) |
| 79 | } |
| 80 | } |
| 81 | |
| 82 | // Spec §6.4, §9.1: an intentional stop (hibernation, maintenance) during a |
| 83 | // first wake's 10 s ends the observation with no verdict. GenCancelled closes |
| 84 | // what GenHealthy opened, so the host stops showing as settling and the |
| 85 | // deploy watch can complete; nothing is recorded against the build. With no |
| 86 | // closing event, observing[host] stayed set for ever. |
| 87 | func TestACancelledObservationStopsSettlingAndRecordsNothing(t *testing.T) { |
| 88 | h := newHarness(t, "a") |
| 89 | h.act.mu.Lock() |
| 90 | delete(h.act.gens, "a") // asleep: its first wake is the only thing that runs v2 |
| 91 | h.act.mu.Unlock() |
| 92 | h.stage("v2") |
| 93 | h.install("v2", "a") |
| 94 | h.pass() |
| 95 | // The wake's plan: its events carry the plan's token (final review F4). |
| 96 | p := h.m.Plan(context.Background(), h.rows.get("a")) |
| 97 | g := activator.Gen{N: 9, Build: "v2", Digest: "sha-v2"} |
| 98 | h.m.OnGen(activator.GenEvent{Kind: activator.GenHealthy, Route: h.rows.get("a"), Gen: g, FirstWake: true, Token: p.Token}) |
| 99 | if got := h.m.Status().Hosts["a"]; got.Attempt != status.AttemptSettling { |
| 100 | t.Fatalf("under observation: a = %+v, want settling", got) |
| 101 | } |
| 102 | h.m.OnGen(activator.GenEvent{Kind: activator.GenCancelled, Route: h.rows.get("a"), Gen: g, FirstWake: true, Token: p.Token}) |
| 103 | if got := h.m.Status().Hosts["a"]; got.Attempt == status.AttemptSettling { |
| 104 | t.Fatalf("after the cancelled observation: a = %+v, still settling", got) |
| 105 | } |
| 106 | if _, ok := h.book.Lookup("bab", "blog", "v2"); ok { |
| 107 | t.Fatal("a cancelled observation was recorded as a failure of the build") |
| 108 | } |
| 109 | if r := h.rows.get("a"); r.GoodVersion != "v1" { |
| 110 | t.Fatalf("good build = %q: a cancelled observation is no settle", r.GoodVersion) |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | // Plan 6's deploy watch counts a sleeper as "installed, awaiting first |
| 115 | // wake" once its box reports B installed. No attempt runs for a sleeper, so |
| 116 | // the pass that first sees B must republish. |
| 117 | func TestANewlyInstalledBuildIsPublishedAtOnce(t *testing.T) { |
| 118 | h := newHarness(t, "a") |
| 119 | var changes atomic.Int32 |
| 120 | h.m.d.OnChange = func() { changes.Add(1) } |
| 121 | h.act.mu.Lock() |
| 122 | delete(h.act.gens, "a") // asleep |
| 123 | h.act.mu.Unlock() |
| 124 | h.stage("v2") |
| 125 | h.install("v2", "a") |
| 126 | h.pass() |
| 127 | if changes.Load() == 0 { |
| 128 | t.Fatal("a newly installed build waited for the five-minute status tick") |
| 129 | } |
| 130 | if got := h.m.Status().Hosts["a"]; got.Installed != "v2" || got.Awake { |
| 131 | t.Fatalf("a = %+v, want v2 installed and asleep", got) |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | // OnChange is a nudge called after every transition, with no manager or |
| 136 | // scheduler lock held: this one reads Status, which takes both, so a call |
| 137 | // under either lock deadlocks the test. |
| 138 | func TestOnChangeFollowsEveryTransitionWithNoLockHeld(t *testing.T) { |
| 139 | h := newHarness(t, "a") |
| 140 | var changes atomic.Int32 |
| 141 | h.m.d.OnChange = func() { |
| 142 | h.m.Status() |
| 143 | changes.Add(1) |
| 144 | } |
| 145 | h.stage("v2") |
| 146 | h.install("v2", "a") |
| 147 | h.pass() |
| 148 | waitFor(t, "a on v2", func() bool { return h.servingAll("v2", "a") }) |
| 149 | waitFor(t, "the attempt's change", func() bool { return changes.Load() >= 2 }) // the pass's, then done()'s |
| 150 | before := changes.Load() |
| 151 | h.m.OnGen(activator.GenEvent{Kind: activator.GenSettled, Route: h.rows.get("a"), Gen: activator.Gen{N: 1, Build: "v2"}}) |
| 152 | if changes.Load() <= before { |
| 153 | t.Fatal("OnGen did not report its transition") |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | func TestStatusReportsAConfigRevisionDeliveredButNotActive(t *testing.T) { |
| 158 | h := newHarness(t, "a") |
| 159 | h.pass() |
| 160 | gate := h.act.hold("a") |
| 161 | h.m.ConfigDelivered("a", "rev1") |
| 162 | waitFor(t, "the config attempt", func() bool { _, s, _ := h.act.snapshot(); return len(s) == 1 }) |
| 163 | got := h.m.Status().Hosts["a"] |
| 164 | // The config attempt is its cohort's canary, so it reads as settling. |
| 165 | if !got.ConfigPending || got.Attempt != status.AttemptSettling || got.AttemptBuild != "v1" || got.ConfigError != "" { |
| 166 | t.Fatalf("a = %+v, want the config attempt in flight with its revision pending", got) |
| 167 | } |
| 168 | close(gate) |
| 169 | waitFor(t, "the config to activate", func() bool { return !h.m.Status().Hosts["a"].ConfigPending }) |
| 170 | } |
| 171 | |
| 172 | // Carried from Task 13: a host whose route left the box is forgotten by the |
| 173 | // next Pass, queue slot and config track alike, so Status never reports a |
| 174 | // departed host as queued. One with an attempt running is kept until the |
| 175 | // attempt finishes, and a row only passing through "moving" keeps its place. |
| 176 | func TestPassForgetsHostsThatLeftTheBox(t *testing.T) { |
| 177 | h := newHarness(t, "a", "b", "c") |
| 178 | h.stage("v2") |
| 179 | gate := h.act.hold("a") |
| 180 | h.install("v2", "a", "b", "c") |
| 181 | h.pass() |
| 182 | waitFor(t, "a's canary", func() bool { _, s, _ := h.act.snapshot(); return len(s) == 1 }) |
| 183 | h.m.ConfigDelivered("b", "rev1") |
| 184 | if r := h.m.Status(); r.Hosts["b"].Attempt != status.AttemptQueued || !r.Hosts["b"].ConfigPending { |
| 185 | t.Fatalf("b = %+v, want queued with its config pending", r.Hosts["b"]) |
| 186 | } |
| 187 | moving := h.rows.get("c") |
| 188 | moving.Status = "moving" |
| 189 | h.rows.put(moving) |
| 190 | h.rows.mu.Lock() |
| 191 | delete(h.rows.rows, "a") |
| 192 | delete(h.rows.rows, "b") |
| 193 | h.rows.mu.Unlock() |
| 194 | h.pass() |
| 195 | r := h.m.Status() |
| 196 | if _, ok := r.Hosts["b"]; ok { |
| 197 | t.Fatalf("b = %+v after its route left the box", r.Hosts["b"]) |
| 198 | } |
| 199 | h.m.mu.Lock() |
| 200 | _, track := h.m.cfg["b"] |
| 201 | _, mapped := h.m.hostApp["b"] |
| 202 | h.m.mu.Unlock() |
| 203 | if track || mapped { |
| 204 | t.Fatalf("b's config track (%v) or app mapping (%v) outlived its route", track, mapped) |
| 205 | } |
| 206 | if got := r.Hosts["a"]; got.Attempt != status.AttemptSettling { |
| 207 | t.Fatalf("a = %+v: a host with its attempt running was forgotten", got) |
| 208 | } |
| 209 | if got := r.Hosts["c"]; got.Attempt != status.AttemptQueued || got.Installed != "v2" { |
| 210 | t.Fatalf("c = %+v: a row passing through moving lost its place", got) |
| 211 | } |
| 212 | close(gate) |
| 213 | waitFor(t, "a's attempt to finish", func() bool { calls, _, _ := h.act.snapshot(); return len(calls) >= 1 }) |
| 214 | waitFor(t, "a forgotten once its attempt finished", func() bool { |
| 215 | h.pass() |
| 216 | _, ok := h.m.Status().Hosts["a"] |
| 217 | return !ok |
| 218 | }) |
| 219 | } |
| 220 | |
| 221 | // A host whose process is starting or stopping (the activator owns it and has |
| 222 | // no current generation) is transient: no serving build and not awake, though |
| 223 | // the last pass cached one. One that is awake or asleep is not (Astra R3 |
| 224 | // section 5 item 10). Nothing is installed, so no attempt runs: the harness's |
| 225 | // swap goroutines would otherwise write the activator's generations under the |
| 226 | // test. |
| 227 | func TestStatusReportsAStartingOrStoppingHostAsTransient(t *testing.T) { |
| 228 | h := newHarness(t, "a", "b", "c") |
| 229 | h.pass() |
| 230 | if got := h.m.Status().Hosts["a"]; got.Transient || got.Installed != "v1" || got.Serving != "v1" || !got.Awake { |
| 231 | t.Fatalf("a before = %+v, want awake on v1 and not transient", got) |
| 232 | } |
| 233 | h.act.mu.Lock() |
| 234 | delete(h.act.gens, "a") |
| 235 | h.act.owned["a"] = true // stopping or waking: a process, no current generation |
| 236 | delete(h.act.gens, "b") // asleep: owned nowhere |
| 237 | h.act.mu.Unlock() |
| 238 | // The scheduler's cached serving build is deliberately not refreshed: the |
| 239 | // report must not repeat it for a host the activator says is mid-change. |
| 240 | r := h.m.Status() |
| 241 | a := r.Hosts["a"] |
| 242 | if !a.Transient || a.Serving != "" || a.Awake || a.Installed != "v1" { |
| 243 | t.Fatalf("a = %+v, want transient, not awake, no serving build", a) |
| 244 | } |
| 245 | if b := r.Hosts["b"]; b.Transient { |
| 246 | t.Fatalf("b = %+v: an asleep host is not transient", b) |
| 247 | } |
| 248 | if c := r.Hosts["c"]; c.Transient || c.Serving != "v1" || !c.Awake { |
| 249 | t.Fatalf("c = %+v: an awake host is not transient", c) |
| 250 | } |
| 251 | } |
| 252 | |
| 253 | // Status asks the activator with no Manager or scheduler lock held (the lock |
| 254 | // order on Manager), for both of its questions: Current for every host, and |
| 255 | // Owns for one with no current generation (b, asleep here). The probe only |
| 256 | // TryLocks, so a violation fails the test instead of deadlocking it. No |
| 257 | // attempt runs (nothing is installed), so no other goroutine holds these |
| 258 | // locks for long. The probe is the harness's (probeAct), wired in before |
| 259 | // any scheduler started; its hook is installed atomically (final review |
| 260 | // round 1, finding 7). |
| 261 | func TestStatusAsksTheActivatorWithNoLockHeld(t *testing.T) { |
| 262 | h := newHarness(t, "a", "b") |
| 263 | h.pass() |
| 264 | h.m.mu.Lock() |
| 265 | scheds := make([]*scheduler, 0, len(h.m.apps)) |
| 266 | for _, s := range h.m.apps { |
| 267 | scheds = append(scheds, s) |
| 268 | } |
| 269 | h.m.mu.Unlock() |
| 270 | if len(scheds) == 0 { |
| 271 | t.Fatal("the pass built no scheduler") |
| 272 | } |
| 273 | h.act.mu.Lock() |
| 274 | delete(h.act.gens, "b") // asleep: Status asks Owns about it |
| 275 | h.act.mu.Unlock() |
| 276 | lp := &lockProbe{m: h.m, scheds: scheds, t: t, asked: map[string]*atomic.Int64{"Owns": {}, "Current": {}}} |
| 277 | h.probe.onAsk(lp.check) |
| 278 | defer h.probe.onAsk(nil) |
| 279 | h.m.Status() |
| 280 | if lp.asked["Current"].Load() == 0 || lp.asked["Owns"].Load() == 0 { |
| 281 | t.Fatalf("Status asked Current %d and Owns %d time(s): want both asked", lp.asked["Current"].Load(), lp.asked["Owns"].Load()) |
| 282 | } |
| 283 | } |
| 284 | |
| 285 | // lockProbe fails the test when Owns or Current is asked by Status while the |
| 286 | // Manager's or a scheduler's mutex is held. |
| 287 | type lockProbe struct { |
| 288 | m *Manager |
| 289 | scheds []*scheduler |
| 290 | t *testing.T |
| 291 | asked map[string]*atomic.Int64 // by method; built before the hook is installed, never written after |
| 292 | } |
| 293 | |
| 294 | func (p *lockProbe) check(method, _ string) { |
| 295 | if !askedByStatus() { |
| 296 | return // the scheduler's own goroutine asks too; only Status is under test |
| 297 | } |
| 298 | p.asked[method].Add(1) |
| 299 | // A lock that Status holds stays held for the whole ask, while a |
| 300 | // scheduler goroutine's own hold is brief, so a violation is a lock that |
| 301 | // cannot be taken for a long while. TryLock, never Lock: a violation |
| 302 | // fails the test instead of deadlocking it. |
| 303 | free := func(mu *sync.Mutex) bool { |
| 304 | for end := time.Now().Add(300 * time.Millisecond); time.Now().Before(end); time.Sleep(time.Millisecond) { |
| 305 | if mu.TryLock() { |
| 306 | mu.Unlock() |
| 307 | return true |
| 308 | } |
| 309 | } |
| 310 | return false |
| 311 | } |
| 312 | if !free(&p.m.mu) { |
| 313 | p.t.Errorf("activator asked (%s) with Manager.mu held", method) |
| 314 | } |
| 315 | for _, s := range p.scheds { |
| 316 | if !free(&s.mu) { |
| 317 | p.t.Errorf("activator asked (%s) with a scheduler.mu held", method) |
| 318 | } |
| 319 | } |
| 320 | } |
| 321 | |
| 322 | // askedByStatus reports whether the calling goroutine is inside |
| 323 | // Manager.Status. |
| 324 | func askedByStatus() bool { |
| 325 | pcs := make([]uintptr, 32) |
| 326 | frames := runtime.CallersFrames(pcs[:runtime.Callers(2, pcs)]) |
| 327 | for { |
| 328 | f, more := frames.Next() |
| 329 | if strings.HasSuffix(f.Function, "(*Manager).Status") { |
| 330 | return true |
| 331 | } |
| 332 | if !more { |
| 333 | return false |
| 334 | } |
| 335 | } |
| 336 | } |
| 337 | |
| 338 | // A host that joins a scheduler after Status read the names was never asked of |
| 339 | // the activator, so nothing says it is settled: that report calls it |
| 340 | // transient, and the next one asks. |
| 341 | func TestStatusCallsAHostThatJoinedAfterTheNamesWereReadTransient(t *testing.T) { |
| 342 | h := newHarness(t, "a") |
| 343 | h.pass() |
| 344 | h.m.mu.Lock() |
| 345 | var s *scheduler |
| 346 | for _, s = range h.m.apps { |
| 347 | } |
| 348 | h.m.mu.Unlock() |
| 349 | var once atomic.Bool |
| 350 | h.probe.onAsk(func(method, _ string) { |
| 351 | if method != "Current" || !askedByStatus() { // not the scheduler's own goroutine |
| 352 | return |
| 353 | } |
| 354 | if once.CompareAndSwap(false, true) { // the names are read; now a host joins |
| 355 | s.mu.Lock() |
| 356 | s.q.hosts["late"] = &hostSlot{wantBuild: "v1", good: "v1", serving: "v1"} |
| 357 | s.mu.Unlock() |
| 358 | } |
| 359 | }) |
| 360 | defer h.probe.onAsk(nil) |
| 361 | r := h.m.Status() |
| 362 | if late := r.Hosts["late"]; !late.Transient || late.Serving != "" || late.Awake { |
| 363 | t.Fatalf("late = %+v, want transient for the report that never asked", late) |
| 364 | } |
| 365 | if a := r.Hosts["a"]; a.Transient || a.Serving != "v1" { |
| 366 | t.Fatalf("a = %+v, want awake on v1", a) |
| 367 | } |
| 368 | } |
| 369 | |
amadan