| 1 | package main |
| 2 | |
| 3 | // Unit tests for the activator state machine (activator.go) with a fake |
| 4 | // backend — wake coalescing, idle→sleep, lease-lost stand-down, contention |
| 5 | // 503s. No AWS, no processes: the live end-to-end test (the wake/restore/ |
| 6 | // sleep money path against real S3 + a real spawned instance) is |
| 7 | // TestActivatorLiveHibernation in activator_live_test.go. |
| 8 | |
| 9 | import ( |
| 10 | "context" |
| 11 | "errors" |
| 12 | "fmt" |
| 13 | "io" |
| 14 | "net" |
| 15 | "net/http" |
| 16 | "net/http/httptest" |
| 17 | "os" |
| 18 | "path/filepath" |
| 19 | "sync" |
| 20 | "sync/atomic" |
| 21 | "testing" |
| 22 | "time" |
| 23 | |
| 24 | "lchat/internal/bot" |
| 25 | ) |
| 26 | |
| 27 | // --- fakes ------------------------------------------------------------------- |
| 28 | |
| 29 | type fakeRep struct { |
| 30 | mu sync.Mutex |
| 31 | err error |
| 32 | closed bool |
| 33 | } |
| 34 | |
| 35 | func (r *fakeRep) Err() error { |
| 36 | r.mu.Lock() |
| 37 | defer r.mu.Unlock() |
| 38 | return r.err |
| 39 | } |
| 40 | func (r *fakeRep) setErr(err error) { |
| 41 | r.mu.Lock() |
| 42 | defer r.mu.Unlock() |
| 43 | r.err = err |
| 44 | } |
| 45 | func (r *fakeRep) Close(context.Context) error { |
| 46 | r.mu.Lock() |
| 47 | defer r.mu.Unlock() |
| 48 | r.closed = true |
| 49 | return nil |
| 50 | } |
| 51 | func (r *fakeRep) isClosed() bool { |
| 52 | r.mu.Lock() |
| 53 | defer r.mu.Unlock() |
| 54 | return r.closed |
| 55 | } |
| 56 | |
| 57 | type fakeProc struct { |
| 58 | mu sync.Mutex |
| 59 | stopped bool |
| 60 | dead bool |
| 61 | // stopGate, when non-nil, blocks stop() until closed — holds an entry in |
| 62 | // the stopping state so a test can poke at it. |
| 63 | stopGate chan struct{} |
| 64 | // label + rec feed the backend's order log ("instance" / "sidecar"), so |
| 65 | // the co-spawn/co-stop tests can assert interleavings. |
| 66 | label string |
| 67 | rec *fakeBackend |
| 68 | } |
| 69 | |
| 70 | func (p *fakeProc) stop(context.Context) error { |
| 71 | p.mu.Lock() |
| 72 | gate := p.stopGate |
| 73 | p.mu.Unlock() |
| 74 | if gate != nil { |
| 75 | <-gate |
| 76 | } |
| 77 | p.mu.Lock() |
| 78 | p.stopped = true |
| 79 | p.mu.Unlock() |
| 80 | if p.rec != nil { |
| 81 | p.rec.record("stop:" + p.label) |
| 82 | } |
| 83 | return nil |
| 84 | } |
| 85 | func (p *fakeProc) exited() bool { |
| 86 | p.mu.Lock() |
| 87 | defer p.mu.Unlock() |
| 88 | return p.dead |
| 89 | } |
| 90 | func (p *fakeProc) isStopped() bool { |
| 91 | p.mu.Lock() |
| 92 | defer p.mu.Unlock() |
| 93 | return p.stopped |
| 94 | } |
| 95 | |
| 96 | type fakeLease struct { |
| 97 | rep *fakeRep |
| 98 | b *fakeBackend // records the sync interval the wake threads through |
| 99 | } |
| 100 | |
| 101 | func (l *fakeLease) replicate(_ context.Context, _ string, syncInterval time.Duration) (actRep, error) { |
| 102 | if l.b != nil { |
| 103 | l.b.mu.Lock() |
| 104 | l.b.syncIntervals = append(l.b.syncIntervals, syncInterval) |
| 105 | l.b.mu.Unlock() |
| 106 | } |
| 107 | return l.rep, nil |
| 108 | } |
| 109 | func (l *fakeLease) release(context.Context) error { return nil } |
| 110 | |
| 111 | type fakeBackend struct { |
| 112 | mu sync.Mutex |
| 113 | acquires int |
| 114 | restores int |
| 115 | starts int |
| 116 | acquireErr error |
| 117 | startGate chan struct{} // when non-nil, start() blocks until closed |
| 118 | reps []*fakeRep |
| 119 | procs []*fakeProc |
| 120 | // syncIntervals records, per wake, the replication-class tick doWake |
| 121 | // threads into replicate() — the ladder's seam assertion. |
| 122 | syncIntervals []time.Duration |
| 123 | dueVal int64 // what the instance reports at the sleep-time next-due probe |
| 124 | dueErr error // when non-nil, the probe fails (dead socket, missing token, …) |
| 125 | |
| 126 | // execVerFn models the live/<host> target the child execs at spawn — start() |
| 127 | // returns it as the spawnVersion, exactly as s3Backend reads the symlink just |
| 128 | // before Start(). Nil → "" (unknown), which disables convergence, so tests |
| 129 | // that don't care are unaffected. Point it at the same source as the sweep's |
| 130 | // liveVer seam to model a symlink faithfully (both read the current target). |
| 131 | execVerFn func() string |
| 132 | |
| 133 | // Bot sidecar co-spawn. sidecarStaged is hasSidecar's answer (false by |
| 134 | // default, so every pre-co-spawn test doubles as the "host without a |
| 135 | // sidecar spawns none" case); sidecarErr injects spawn failures. |
| 136 | sidecarStaged bool |
| 137 | sidecarErr error |
| 138 | sidecarStarts int |
| 139 | sidecars []*fakeProc |
| 140 | // order interleaves backend calls and proc stops (start:instance, |
| 141 | // start:sidecar, nextDue, stop:sidecar, stop:instance) for the ordering |
| 142 | // assertions. |
| 143 | order []string |
| 144 | } |
| 145 | |
| 146 | func (b *fakeBackend) record(ev string) { |
| 147 | b.mu.Lock() |
| 148 | b.order = append(b.order, ev) |
| 149 | b.mu.Unlock() |
| 150 | } |
| 151 | |
| 152 | func (b *fakeBackend) events() []string { |
| 153 | b.mu.Lock() |
| 154 | defer b.mu.Unlock() |
| 155 | return append([]string(nil), b.order...) |
| 156 | } |
| 157 | |
| 158 | func (b *fakeBackend) acquire(context.Context, string) (actLease, error) { |
| 159 | b.mu.Lock() |
| 160 | b.acquires++ |
| 161 | err := b.acquireErr |
| 162 | rep := &fakeRep{} |
| 163 | b.reps = append(b.reps, rep) |
| 164 | b.mu.Unlock() |
| 165 | if err != nil { |
| 166 | return nil, err |
| 167 | } |
| 168 | return &fakeLease{rep: rep, b: b}, nil |
| 169 | } |
| 170 | |
| 171 | func (b *fakeBackend) restore(context.Context, string, string) error { |
| 172 | b.mu.Lock() |
| 173 | defer b.mu.Unlock() |
| 174 | b.restores++ |
| 175 | return nil |
| 176 | } |
| 177 | |
| 178 | func (b *fakeBackend) start(context.Context, Route) (actProc, string, error) { |
| 179 | b.mu.Lock() |
| 180 | b.starts++ |
| 181 | b.order = append(b.order, "start:instance") |
| 182 | gate := b.startGate |
| 183 | fn := b.execVerFn |
| 184 | proc := &fakeProc{label: "instance", rec: b} |
| 185 | b.procs = append(b.procs, proc) |
| 186 | b.mu.Unlock() |
| 187 | if gate != nil { |
| 188 | <-gate |
| 189 | } |
| 190 | ver := "" |
| 191 | if fn != nil { |
| 192 | ver = fn() |
| 193 | } |
| 194 | return proc, ver, nil |
| 195 | } |
| 196 | |
| 197 | func (b *fakeBackend) nextDue(context.Context, Route) (int64, error) { |
| 198 | b.mu.Lock() |
| 199 | defer b.mu.Unlock() |
| 200 | b.order = append(b.order, "nextDue") |
| 201 | if b.dueErr != nil { |
| 202 | return 0, b.dueErr |
| 203 | } |
| 204 | return b.dueVal, nil |
| 205 | } |
| 206 | |
| 207 | func (b *fakeBackend) hasSidecar(string) bool { |
| 208 | b.mu.Lock() |
| 209 | defer b.mu.Unlock() |
| 210 | return b.sidecarStaged |
| 211 | } |
| 212 | |
| 213 | func (b *fakeBackend) startSidecar(context.Context, Route) (actProc, error) { |
| 214 | b.mu.Lock() |
| 215 | defer b.mu.Unlock() |
| 216 | b.sidecarStarts++ |
| 217 | b.order = append(b.order, "start:sidecar") |
| 218 | if b.sidecarErr != nil { |
| 219 | return nil, b.sidecarErr |
| 220 | } |
| 221 | p := &fakeProc{label: "sidecar", rec: b} |
| 222 | b.sidecars = append(b.sidecars, p) |
| 223 | return p, nil |
| 224 | } |
| 225 | |
| 226 | func (b *fakeBackend) sidecarStartCount() int { |
| 227 | b.mu.Lock() |
| 228 | defer b.mu.Unlock() |
| 229 | return b.sidecarStarts |
| 230 | } |
| 231 | |
| 232 | func (b *fakeBackend) lastSidecar() *fakeProc { |
| 233 | b.mu.Lock() |
| 234 | defer b.mu.Unlock() |
| 235 | return b.sidecars[len(b.sidecars)-1] |
| 236 | } |
| 237 | |
| 238 | func (b *fakeBackend) setSidecarErr(err error) { |
| 239 | b.mu.Lock() |
| 240 | defer b.mu.Unlock() |
| 241 | b.sidecarErr = err |
| 242 | } |
| 243 | |
| 244 | func (b *fakeBackend) setDue(v int64) { |
| 245 | b.mu.Lock() |
| 246 | defer b.mu.Unlock() |
| 247 | b.dueVal = v |
| 248 | } |
| 249 | |
| 250 | func (b *fakeBackend) counts() (acquires, restores, starts int) { |
| 251 | b.mu.Lock() |
| 252 | defer b.mu.Unlock() |
| 253 | return b.acquires, b.restores, b.starts |
| 254 | } |
| 255 | |
| 256 | func (b *fakeBackend) lastProc() *fakeProc { |
| 257 | b.mu.Lock() |
| 258 | defer b.mu.Unlock() |
| 259 | return b.procs[len(b.procs)-1] |
| 260 | } |
| 261 | |
| 262 | func (b *fakeBackend) lastRep() *fakeRep { |
| 263 | b.mu.Lock() |
| 264 | defer b.mu.Unlock() |
| 265 | return b.reps[len(b.reps)-1] |
| 266 | } |
| 267 | |
| 268 | // fakeClock is the activator's `now` seam. |
| 269 | type fakeClock struct { |
| 270 | mu sync.Mutex |
| 271 | t time.Time |
| 272 | } |
| 273 | |
| 274 | func (c *fakeClock) now() time.Time { |
| 275 | c.mu.Lock() |
| 276 | defer c.mu.Unlock() |
| 277 | return c.t |
| 278 | } |
| 279 | func (c *fakeClock) advance(d time.Duration) { |
| 280 | c.mu.Lock() |
| 281 | defer c.mu.Unlock() |
| 282 | c.t = c.t.Add(d) |
| 283 | } |
| 284 | |
| 285 | // testActivator wires the fakes to a REAL (temp-dir) registry as the due store, |
| 286 | // so the due-time tests exercise the actual routes.next_due SQL, not a fake of it. |
| 287 | func testActivator(t *testing.T, idle time.Duration) (*activatorSvc, *fakeBackend, *fakeClock, *Registry) { |
| 288 | t.Helper() |
| 289 | b := &fakeBackend{} |
| 290 | clk := &fakeClock{t: time.Unix(1_700_000_000, 0)} |
| 291 | reg, err := openRegistry(filepath.Join(t.TempDir(), "registry.db")) |
| 292 | if err != nil { |
| 293 | t.Fatalf("open registry: %v", err) |
| 294 | } |
| 295 | a := newActivator(hibernateConfig{Bucket: "test-bucket", Idle: idle}, b, reg) |
| 296 | a.now = clk.now |
| 297 | return a, b, clk, reg |
| 298 | } |
| 299 | |
| 300 | func testRoute(host string) Route { |
| 301 | // Hibernate is set because these routes model activator-managed hosts — |
| 302 | // the per-host gate (dueInstances, the router's upstream flag) only acts |
| 303 | // on flagged instances. |
| 304 | return Route{Host: host, Kind: "instance", Socket: "/tmp/nope.sock", |
| 305 | DBPath: filepath.Join("/tmp", host+".db"), Hibernate: true} |
| 306 | } |
| 307 | |
| 308 | // (waitFor — poll until a condition holds — is shared from bot_run_test.go; |
| 309 | // the stop path runs on its own goroutine, so tests observe its completion |
| 310 | // rather than assuming it.) |
| 311 | |
| 312 | func (a *activatorSvc) entryCount() int { |
| 313 | a.mu.Lock() |
| 314 | defer a.mu.Unlock() |
| 315 | return len(a.entries) |
| 316 | } |
| 317 | |
| 318 | // --- tests --------------------------------------------------------------------- |
| 319 | |
| 320 | // Many concurrent requests during a wake coalesce into exactly one wake: one |
| 321 | // lease claim, one restore, one spawn. |
| 322 | func TestActivatorWakeCoalesces(t *testing.T) { |
| 323 | a, b, _, _ := testActivator(t, time.Minute) |
| 324 | b.startGate = make(chan struct{}) |
| 325 | |
| 326 | const n = 12 |
| 327 | var ok atomic.Int32 |
| 328 | var wg sync.WaitGroup |
| 329 | for i := 0; i < n; i++ { |
| 330 | wg.Add(1) |
| 331 | go func() { |
| 332 | defer wg.Done() |
| 333 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 334 | if err != nil { |
| 335 | t.Errorf("beginRequest: %v", err) |
| 336 | return |
| 337 | } |
| 338 | ok.Add(1) |
| 339 | release() |
| 340 | }() |
| 341 | } |
| 342 | // Let the requests pile up on the gated wake, then open it. |
| 343 | waitFor(t, "the wake to begin", func() bool { _, _, s := b.counts(); return s == 1 }) |
| 344 | time.Sleep(50 * time.Millisecond) |
| 345 | close(b.startGate) |
| 346 | wg.Wait() |
| 347 | |
| 348 | if got := ok.Load(); got != n { |
| 349 | t.Fatalf("%d/%d requests succeeded", got, n) |
| 350 | } |
| 351 | acquires, restores, starts := b.counts() |
| 352 | if acquires != 1 || restores != 1 || starts != 1 { |
| 353 | t.Fatalf("wake did not coalesce: acquires=%d restores=%d starts=%d (want 1 each)", acquires, restores, starts) |
| 354 | } |
| 355 | } |
| 356 | |
| 357 | // The wake threads the route's replication class into replicate() as the |
| 358 | // litestream sync interval — the per-plan durability ladder's seam. Unset |
| 359 | // (every pre-ladder route) must thread 0, i.e. litestream's default: a deploy |
| 360 | // of the ladder changes no host's durability by itself. |
| 361 | func TestActivatorWakeThreadsReplicationClass(t *testing.T) { |
| 362 | a, b, _, _ := testActivator(t, time.Minute) |
| 363 | |
| 364 | for i, tc := range []struct { |
| 365 | class string |
| 366 | want time.Duration |
| 367 | }{ |
| 368 | {"", 0}, |
| 369 | {replicationStream, 0}, |
| 370 | {replicationBatched, 5 * time.Minute}, |
| 371 | {replicationDaily, 24 * time.Hour}, |
| 372 | } { |
| 373 | rt := testRoute(fmt.Sprintf("ladder%d.example.com", i)) |
| 374 | rt.Replication = tc.class |
| 375 | release, err := a.beginRequest(context.Background(), rt) |
| 376 | if err != nil { |
| 377 | t.Fatalf("beginRequest(%q): %v", tc.class, err) |
| 378 | } |
| 379 | release() |
| 380 | b.mu.Lock() |
| 381 | got := b.syncIntervals[len(b.syncIntervals)-1] |
| 382 | b.mu.Unlock() |
| 383 | if got != tc.want { |
| 384 | t.Errorf("class %q threaded sync interval %v, want %v", tc.class, got, tc.want) |
| 385 | } |
| 386 | } |
| 387 | } |
| 388 | |
| 389 | // After the idle window with nothing in flight, the sweeper hibernates the |
| 390 | // instance: process stopped, replication closed (final checkpoint + lease |
| 391 | // release), entry forgotten; the next request wakes it afresh. |
| 392 | func TestActivatorIdleSleep(t *testing.T) { |
| 393 | a, b, clk, _ := testActivator(t, 10*time.Second) |
| 394 | |
| 395 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 396 | if err != nil { |
| 397 | t.Fatalf("beginRequest: %v", err) |
| 398 | } |
| 399 | release() |
| 400 | |
| 401 | // Not yet idle: nothing happens. |
| 402 | clk.advance(9 * time.Second) |
| 403 | a.sweepOnce() |
| 404 | if a.entryCount() != 1 { |
| 405 | t.Fatal("instance hibernated before its idle window elapsed") |
| 406 | } |
| 407 | |
| 408 | clk.advance(2 * time.Second) |
| 409 | a.sweepOnce() |
| 410 | waitFor(t, "hibernation", func() bool { return a.entryCount() == 0 }) |
| 411 | if !b.lastProc().isStopped() { |
| 412 | t.Fatal("idle instance's process was not stopped") |
| 413 | } |
| 414 | if !b.lastRep().isClosed() { |
| 415 | t.Fatal("idle instance's replication was not closed (no final checkpoint/lease release)") |
| 416 | } |
| 417 | |
| 418 | // A new request wakes it again — a second full wake. |
| 419 | release, err = a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 420 | if err != nil { |
| 421 | t.Fatalf("beginRequest after sleep: %v", err) |
| 422 | } |
| 423 | release() |
| 424 | if _, _, starts := b.counts(); starts != 2 { |
| 425 | t.Fatalf("starts=%d after sleep+rewake, want 2", starts) |
| 426 | } |
| 427 | } |
| 428 | |
| 429 | // An in-flight request (an open WebSocket, in production) holds the instance |
| 430 | // awake past the idle window; only after it finishes does the idle clock run. |
| 431 | func TestActivatorInflightKeepsAwake(t *testing.T) { |
| 432 | a, b, clk, _ := testActivator(t, 10*time.Second) |
| 433 | |
| 434 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 435 | if err != nil { |
| 436 | t.Fatalf("beginRequest: %v", err) |
| 437 | } |
| 438 | |
| 439 | clk.advance(time.Hour) |
| 440 | a.sweepOnce() |
| 441 | if a.entryCount() != 1 || b.lastProc().isStopped() { |
| 442 | t.Fatal("instance with an in-flight request was hibernated") |
| 443 | } |
| 444 | |
| 445 | release() // stamps lastActive = now |
| 446 | a.sweepOnce() |
| 447 | if a.entryCount() != 1 { |
| 448 | t.Fatal("instance hibernated immediately after its last request finished") |
| 449 | } |
| 450 | clk.advance(11 * time.Second) |
| 451 | a.sweepOnce() |
| 452 | waitFor(t, "hibernation", func() bool { return a.entryCount() == 0 }) |
| 453 | } |
| 454 | |
| 455 | // A ship repoints live/<host> under an awake child. A running process never |
| 456 | // re-reads its symlink, so idle-hibernate is the only thing that converges an |
| 457 | // instance — and an always-awake space (every bot-sidecar host) never idles. |
| 458 | // The sweep closes that gap: it notices the drift and, after a grace window and |
| 459 | // rate-limited to one per tick, gracefully cycles the child so its next wake |
| 460 | // execs the new binary. See docs/awake-instance-convergence.md. |
| 461 | func TestActivatorConvergesOnSymlinkRepoint(t *testing.T) { |
| 462 | a, b, clk, _ := testActivator(t, time.Hour) // idle far away — isolate the convergence path |
| 463 | var ver atomic.Value // the live/<host> symlink target; swapped to model a ship |
| 464 | ver.Store("v1") |
| 465 | // Model one symlink: the child execs whatever it points at now (execVerFn, |
| 466 | // frozen into spawnVersion at wake), and the sweep reads the current target |
| 467 | // (liveVer). A "ship" swaps ver; the running child keeps its frozen version. |
| 468 | a.liveVer = func(string) (string, error) { return ver.Load().(string), nil } |
| 469 | b.execVerFn = func() string { return ver.Load().(string) } |
| 470 | |
| 471 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 472 | if err != nil { |
| 473 | t.Fatalf("beginRequest: %v", err) |
| 474 | } |
| 475 | release() |
| 476 | |
| 477 | // Symlink unchanged: never cycles, however long it stays awake. |
| 478 | clk.advance(3 * convergeGrace) |
| 479 | a.sweepOnce() |
| 480 | if a.entryCount() != 1 { |
| 481 | t.Fatal("cycled an instance whose symlink never moved") |
| 482 | } |
| 483 | |
| 484 | // Ship: repoint the symlink. The first sweep after only starts the grace |
| 485 | // clock — it must not cycle yet, so a move-then-rollback never drops a live |
| 486 | // connection. |
| 487 | ver.Store("v2") |
| 488 | a.sweepOnce() |
| 489 | if a.entryCount() != 1 { |
| 490 | t.Fatal("cycled inside the grace window — a transient repoint would drop a live connection") |
| 491 | } |
| 492 | |
| 493 | // Past the grace: gracefully cycle (stop → final checkpoint → lease release → |
| 494 | // asleep). The reconnecting client/sidecar would wake it onto v2. |
| 495 | clk.advance(convergeGrace + time.Second) |
| 496 | a.sweepOnce() |
| 497 | waitFor(t, "convergence cycle", func() bool { return a.entryCount() == 0 }) |
| 498 | if !b.lastProc().isStopped() { |
| 499 | t.Fatal("stale instance's process was not stopped") |
| 500 | } |
| 501 | if !b.lastRep().isClosed() { |
| 502 | t.Fatal("convergence cycle skipped the final checkpoint / lease release") |
| 503 | } |
| 504 | |
| 505 | // Re-wake against the now-current symlink: spawnVersion tracks v2, so it |
| 506 | // never cycles again — no thrash. |
| 507 | release, err = a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 508 | if err != nil { |
| 509 | t.Fatalf("re-wake after convergence: %v", err) |
| 510 | } |
| 511 | release() |
| 512 | clk.advance(3 * convergeGrace) |
| 513 | a.sweepOnce() |
| 514 | if a.entryCount() != 1 { |
| 515 | t.Fatal("cycled again after already converging onto the current build") |
| 516 | } |
| 517 | } |
| 518 | |
| 519 | // Two awake children go stale at once (a fleet-wide ship). One sweep cycles at |
| 520 | // most one — a staggered convergence, never a simultaneous blip of every space. |
| 521 | func TestActivatorConvergenceRateLimitsPerSweep(t *testing.T) { |
| 522 | a, b, clk, _ := testActivator(t, time.Hour) |
| 523 | var ver atomic.Value |
| 524 | ver.Store("v1") |
| 525 | a.liveVer = func(string) (string, error) { return ver.Load().(string), nil } |
| 526 | b.execVerFn = func() string { return ver.Load().(string) } |
| 527 | |
| 528 | for _, h := range []string{"one.example.com", "two.example.com"} { |
| 529 | release, err := a.beginRequest(context.Background(), testRoute(h)) |
| 530 | if err != nil { |
| 531 | t.Fatalf("beginRequest %s: %v", h, err) |
| 532 | } |
| 533 | release() |
| 534 | } |
| 535 | if a.entryCount() != 2 { |
| 536 | t.Fatalf("want 2 awake instances, got %d", a.entryCount()) |
| 537 | } |
| 538 | |
| 539 | ver.Store("v2") // ship under both |
| 540 | a.sweepOnce() // starts the grace clock for both |
| 541 | clk.advance(convergeGrace + time.Second) |
| 542 | |
| 543 | a.sweepOnce() // both are past grace, but only one may cycle this tick |
| 544 | waitFor(t, "first convergence cycle", func() bool { return a.entryCount() == 1 }) |
| 545 | // A moment later, the sweep takes the second. |
| 546 | a.sweepOnce() |
| 547 | waitFor(t, "second convergence cycle", func() bool { return a.entryCount() == 0 }) |
| 548 | } |
| 549 | |
| 550 | // The feature's whole premise: converge a space that NEVER reaches inflight==0 |
| 551 | // (an open WebSocket, or a bot sidecar's persistent connection), which |
| 552 | // idle-hibernate can never touch. Unlike idle-hibernate, convergence fires with |
| 553 | // a request in flight — the graceful stop drains it exactly as a restart would. |
| 554 | func TestActivatorConvergesWithRequestInFlight(t *testing.T) { |
| 555 | a, b, clk, _ := testActivator(t, time.Hour) |
| 556 | var ver atomic.Value |
| 557 | ver.Store("v1") |
| 558 | a.liveVer = func(string) (string, error) { return ver.Load().(string), nil } |
| 559 | b.execVerFn = func() string { return ver.Load().(string) } |
| 560 | |
| 561 | // Hold the request open for the whole test — inflight stays 1, so the idle |
| 562 | // path (which needs inflight==0) can never converge this host. |
| 563 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 564 | if err != nil { |
| 565 | t.Fatalf("beginRequest: %v", err) |
| 566 | } |
| 567 | defer release() |
| 568 | |
| 569 | ver.Store("v2") // ship under the busy space |
| 570 | a.sweepOnce() // grace starts; must not cycle a busy space mid-grace |
| 571 | if a.entryCount() != 1 { |
| 572 | t.Fatal("cycled inside the grace window") |
| 573 | } |
| 574 | clk.advance(convergeGrace + time.Second) |
| 575 | a.sweepOnce() |
| 576 | // Cycles despite the in-flight request — idle-hibernate never would. |
| 577 | waitFor(t, "busy-space convergence", func() bool { return a.entryCount() == 0 }) |
| 578 | if !b.lastProc().isStopped() { |
| 579 | t.Fatal("busy stale instance's process was not stopped") |
| 580 | } |
| 581 | } |
| 582 | |
| 583 | // convergeDecision's state machine, exhaustively: the grace clock starts on |
| 584 | // first drift, survives a rollback correctly, and a read error never clears |
| 585 | // accrued grace (it means "unknown", not "converged"). |
| 586 | func TestConvergeDecision(t *testing.T) { |
| 587 | base := time.Unix(1_700_000_000, 0) |
| 588 | set := base.Add(-40 * time.Second) // a grace clock that started 40s ago |
| 589 | errBoom := errors.New("readlink boom") |
| 590 | cases := []struct { |
| 591 | name string |
| 592 | cur string |
| 593 | readErr error |
| 594 | spawnVer string |
| 595 | staleSince time.Time |
| 596 | wantCycle bool |
| 597 | wantStale time.Time |
| 598 | }{ |
| 599 | {"matches — clears grace", "v1", nil, "v1", set, false, time.Time{}}, |
| 600 | {"empty cur — no data, clears", "", nil, "v1", set, false, time.Time{}}, |
| 601 | {"empty spawnVer — unknown, clears", "v2", nil, "", set, false, time.Time{}}, |
| 602 | {"first drift — starts grace, no cycle", "v2", nil, "v1", time.Time{}, false, base}, |
| 603 | {"drift within grace — holds, no cycle", "v2", nil, "v1", base.Add(-30 * time.Second), false, base.Add(-30 * time.Second)}, |
| 604 | {"drift past grace — cycles", "v2", nil, "v1", base.Add(-convergeGrace - time.Second), true, base.Add(-convergeGrace - time.Second)}, |
| 605 | {"read error — preserves accrued grace", "", errBoom, "v1", set, false, set}, |
| 606 | {"read error before any drift — stays zero", "", errBoom, "v1", time.Time{}, false, time.Time{}}, |
| 607 | {"rollback to spawn — clears grace", "v1", nil, "v1", base.Add(-convergeGrace - time.Second), false, time.Time{}}, |
| 608 | } |
| 609 | for _, c := range cases { |
| 610 | t.Run(c.name, func(t *testing.T) { |
| 611 | cycle, next := convergeDecision(c.cur, c.readErr, c.spawnVer, c.staleSince, base, convergeGrace) |
| 612 | if cycle != c.wantCycle { |
| 613 | t.Errorf("cycle = %v, want %v", cycle, c.wantCycle) |
| 614 | } |
| 615 | if !next.Equal(c.wantStale) { |
| 616 | t.Errorf("nextStale = %v, want %v", next, c.wantStale) |
| 617 | } |
| 618 | }) |
| 619 | } |
| 620 | } |
| 621 | |
| 622 | // A lost lease (failed renewal) stands the instance down immediately — no |
| 623 | // waiting for idle: someone else may own the S3 prefix now. |
| 624 | func TestActivatorLeaseLostStandsDown(t *testing.T) { |
| 625 | a, b, _, _ := testActivator(t, time.Hour) |
| 626 | |
| 627 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 628 | if err != nil { |
| 629 | t.Fatalf("beginRequest: %v", err) |
| 630 | } |
| 631 | release() |
| 632 | |
| 633 | b.lastRep().setErr(errors.New("lease renewal failed")) |
| 634 | a.sweepOnce() // idle window is an hour away; the lost lease alone must trigger the stop |
| 635 | waitFor(t, "stand-down", func() bool { return a.entryCount() == 0 }) |
| 636 | if !b.lastProc().isStopped() { |
| 637 | t.Fatal("lease-lost instance's process was not stopped") |
| 638 | } |
| 639 | if !b.lastRep().isClosed() { |
| 640 | t.Fatal("lease-lost instance's replication was not closed") |
| 641 | } |
| 642 | } |
| 643 | |
| 644 | // A dead child (crashed instance) is recycled by the next request: the stale |
| 645 | // entry is stopped and a fresh wake happens in the same beginRequest call. |
| 646 | func TestActivatorDeadChildRewakesOnRequest(t *testing.T) { |
| 647 | a, b, _, _ := testActivator(t, time.Hour) |
| 648 | |
| 649 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 650 | if err != nil { |
| 651 | t.Fatalf("beginRequest: %v", err) |
| 652 | } |
| 653 | release() |
| 654 | |
| 655 | p := b.lastProc() |
| 656 | p.mu.Lock() |
| 657 | p.dead = true |
| 658 | p.mu.Unlock() |
| 659 | |
| 660 | release, err = a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 661 | if err != nil { |
| 662 | t.Fatalf("beginRequest after crash: %v", err) |
| 663 | } |
| 664 | release() |
| 665 | if _, _, starts := b.counts(); starts != 2 { |
| 666 | t.Fatalf("starts=%d after crash+request, want 2 (recycle + rewake)", starts) |
| 667 | } |
| 668 | } |
| 669 | |
| 670 | // A request that arrives mid-stop waits for the stop, then wakes fresh. |
| 671 | func TestActivatorRequestDuringStopRewakes(t *testing.T) { |
| 672 | a, b, clk, _ := testActivator(t, 10*time.Second) |
| 673 | |
| 674 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 675 | if err != nil { |
| 676 | t.Fatalf("beginRequest: %v", err) |
| 677 | } |
| 678 | release() |
| 679 | |
| 680 | // Hold the stop open so the entry sits in stopping. |
| 681 | gate := make(chan struct{}) |
| 682 | p := b.lastProc() |
| 683 | p.mu.Lock() |
| 684 | p.stopGate = gate |
| 685 | p.mu.Unlock() |
| 686 | |
| 687 | clk.advance(time.Minute) |
| 688 | a.sweepOnce() |
| 689 | |
| 690 | got := make(chan error, 1) |
| 691 | go func() { |
| 692 | release, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 693 | if err == nil { |
| 694 | release() |
| 695 | } |
| 696 | got <- err |
| 697 | }() |
| 698 | time.Sleep(50 * time.Millisecond) // the request should now be parked on e.stopped |
| 699 | close(gate) |
| 700 | if err := <-got; err != nil { |
| 701 | t.Fatalf("request during stop: %v", err) |
| 702 | } |
| 703 | if _, _, starts := b.counts(); starts != 2 { |
| 704 | t.Fatalf("starts=%d, want 2 (sleep then rewake)", starts) |
| 705 | } |
| 706 | } |
| 707 | |
| 708 | // A wake that exceeds the bound turns queued requests away with errWakeTimeout |
| 709 | // (503 at the router) rather than piling connections up behind it. |
| 710 | func TestActivatorWakeTimeout(t *testing.T) { |
| 711 | a, b, _, _ := testActivator(t, time.Minute) |
| 712 | old := wakeWaitBound |
| 713 | wakeWaitBound = 100 * time.Millisecond |
| 714 | t.Cleanup(func() { wakeWaitBound = old }) |
| 715 | |
| 716 | gate := make(chan struct{}) |
| 717 | b.startGate = gate |
| 718 | t.Cleanup(func() { close(gate) }) // let the stuck wake finish at test end |
| 719 | |
| 720 | _, err := a.beginRequest(context.Background(), testRoute("one.example.com")) |
| 721 | if !errors.Is(err, errWakeTimeout) { |
| 722 | t.Fatalf("got %v, want errWakeTimeout", err) |
| 723 | } |
| 724 | } |
| 725 | |
| 726 | // Lease contention surfaces through the router as 503 + Retry-After, and a |
| 727 | // failed wake leaves no entry behind (the next request retries the claim). |
| 728 | func TestActivatorLeaseContention503(t *testing.T) { |
| 729 | a, b, _, _ := testActivator(t, time.Minute) |
| 730 | b.acquireErr = &errLeaseHeld{holder: "old-incarnation", retryAfter: 7 * time.Second} |
| 731 | |
| 732 | reg, err := openRegistry(filepath.Join(t.TempDir(), "registry.db")) |
| 733 | if err != nil { |
| 734 | t.Fatalf("open registry: %v", err) |
| 735 | } |
| 736 | if err := reg.upsert(Route{Host: "held.example.com", Kind: "instance", Socket: "/tmp/held.sock", Hibernate: true}); err != nil { |
| 737 | t.Fatalf("upsert: %v", err) |
| 738 | } |
| 739 | rt, err := newRouter(reg, a) |
| 740 | if err != nil { |
| 741 | t.Fatalf("new router: %v", err) |
| 742 | } |
| 743 | front := httptest.NewServer(rt) |
| 744 | t.Cleanup(front.Close) |
| 745 | |
| 746 | req, _ := http.NewRequest("GET", front.URL+"/", nil) |
| 747 | req.Host = "held.example.com" |
| 748 | resp, err := http.DefaultClient.Do(req) |
| 749 | if err != nil { |
| 750 | t.Fatalf("request: %v", err) |
| 751 | } |
| 752 | defer resp.Body.Close() |
| 753 | if resp.StatusCode != http.StatusServiceUnavailable { |
| 754 | t.Fatalf("status %d, want 503", resp.StatusCode) |
| 755 | } |
| 756 | if ra := resp.Header.Get("Retry-After"); ra != "8" { // 7s truncated + 1 |
| 757 | t.Fatalf("Retry-After %q, want 8", ra) |
| 758 | } |
| 759 | if a.entryCount() != 0 { |
| 760 | t.Fatal("failed wake left an entry behind") |
| 761 | } |
| 762 | } |
| 763 | |
| 764 | // Per-host gating (#550 Part 4 v2 opt-in): with the activator CONFIGURED, only |
| 765 | // a hibernate-flagged instance goes through the wake path; an unflagged one is |
| 766 | // proxied straight, exactly as if there were no activator at all — the |
| 767 | // deploy-safety property that lets the bucket env ship fleet-wide before any |
| 768 | // host is flagged. |
| 769 | func TestRouterPerHostHibernateGating(t *testing.T) { |
| 770 | a, b, _, reg := testActivator(t, time.Minute) |
| 771 | |
| 772 | // Two live backends playing the instances. The flagged host's listener |
| 773 | // stands in for the child a wake would have spawned (the fake backend's |
| 774 | // start() spawns nothing). |
| 775 | bind := func(name string) string { |
| 776 | sock := tmpSock(t) |
| 777 | ln, err := net.Listen("unix", sock) |
| 778 | if err != nil { |
| 779 | t.Fatalf("listen unix: %v", err) |
| 780 | } |
| 781 | srv := &http.Server{Handler: http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { |
| 782 | io.WriteString(w, name) |
| 783 | })} |
| 784 | go srv.Serve(ln) |
| 785 | t.Cleanup(func() { srv.Close() }) |
| 786 | return sock |
| 787 | } |
| 788 | if err := reg.upsert(Route{Host: "hib.example.com", Kind: "instance", Socket: bind("hib"), Hibernate: true}); err != nil { |
| 789 | t.Fatal(err) |
| 790 | } |
| 791 | if err := reg.upsert(Route{Host: "plain.example.com", Kind: "instance", Socket: bind("plain")}); err != nil { |
| 792 | t.Fatal(err) |
| 793 | } |
| 794 | |
| 795 | rt, err := newRouter(reg, a) |
| 796 | if err != nil { |
| 797 | t.Fatalf("new router: %v", err) |
| 798 | } |
| 799 | front := httptest.NewServer(rt) |
| 800 | t.Cleanup(front.Close) |
| 801 | |
| 802 | // The unflagged host: proxied straight, the activator never touched. |
| 803 | if body, code := proxyGet(t, front.URL, "plain.example.com"); code != 200 || body != "plain" { |
| 804 | t.Fatalf("unflagged host: got %d %q, want 200 plain", code, body) |
| 805 | } |
| 806 | if acquires, _, starts := b.counts(); acquires != 0 || starts != 0 { |
| 807 | t.Fatalf("unflagged instance touched the activator (acquires=%d starts=%d)", acquires, starts) |
| 808 | } |
| 809 | |
| 810 | // The flagged host: served through a full wake — lease, restore, spawn. |
| 811 | if body, code := proxyGet(t, front.URL, "hib.example.com"); code != 200 || body != "hib" { |
| 812 | t.Fatalf("flagged host: got %d %q, want 200 hib", code, body) |
| 813 | } |
| 814 | if acquires, restores, starts := b.counts(); acquires != 1 || restores != 1 || starts != 1 { |
| 815 | t.Fatalf("flagged host skipped the wake path: acquires=%d restores=%d starts=%d (want 1 each)", acquires, restores, starts) |
| 816 | } |
| 817 | } |
| 818 | |
| 819 | // A reload that finds an awake host no longer hibernate-flagged retires its |
| 820 | // child — stopped, final-checkpointed (replication closed), entry forgotten — |
| 821 | // the router-process half of `lchat hibernate --off`. And once the flag is |
| 822 | // off, a recorded next_due no longer wakes it: the due index only acts on |
| 823 | // hibernating hosts, so a systemd-managed instance can never be spawned |
| 824 | // against its own unit. |
| 825 | func TestActivatorReloadRetiresUnflaggedHost(t *testing.T) { |
| 826 | a, b, clk, reg := testActivator(t, time.Hour) |
| 827 | route := testRoute("flip.example.com") |
| 828 | if err := reg.upsert(route); err != nil { |
| 829 | t.Fatal(err) |
| 830 | } |
| 831 | rt, err := newRouter(reg, a) |
| 832 | if err != nil { |
| 833 | t.Fatalf("new router: %v", err) |
| 834 | } |
| 835 | |
| 836 | release, err := a.beginRequest(context.Background(), route) |
| 837 | if err != nil { |
| 838 | t.Fatalf("beginRequest: %v", err) |
| 839 | } |
| 840 | release() |
| 841 | // The retire's sleep re-records next_due AFTER the flag is cleared; an |
| 842 | // already-past value must stay inert under the flag filter. |
| 843 | b.setDue(clk.now().Add(-5 * time.Second).Unix()) |
| 844 | |
| 845 | if err := reg.setHibernate("flip.example.com", false); err != nil { |
| 846 | t.Fatal(err) |
| 847 | } |
| 848 | if err := rt.reload(); err != nil { |
| 849 | t.Fatalf("reload: %v", err) |
| 850 | } |
| 851 | waitFor(t, "retire", func() bool { return a.entryCount() == 0 }) |
| 852 | if !b.lastProc().isStopped() { |
| 853 | t.Fatal("retired child was not stopped") |
| 854 | } |
| 855 | if !b.lastRep().isClosed() { |
| 856 | t.Fatal("retired child's replication was not closed (no final checkpoint/lease release)") |
| 857 | } |
| 858 | |
| 859 | // The due sweeper must not resurrect a host that's back under systemd. |
| 860 | a.sweepOnce() |
| 861 | if _, _, starts := b.counts(); starts != 1 { |
| 862 | t.Fatalf("starts=%d after retire, want 1 (no due wake for a systemd-managed host)", starts) |
| 863 | } |
| 864 | } |
| 865 | |
| 866 | // A retire that lands mid-wake can't stop the child yet — it finishes the job |
| 867 | // once the wake resolves, so a flip-off during the small wake window still |
| 868 | // releases cleanly. |
| 869 | func TestActivatorRetireDuringWake(t *testing.T) { |
| 870 | a, b, clk, reg := testActivator(t, time.Hour) |
| 871 | route := testRoute("mid.example.com") |
| 872 | if err := reg.upsert(route); err != nil { |
| 873 | t.Fatal(err) |
| 874 | } |
| 875 | rt, err := newRouter(reg, a) |
| 876 | if err != nil { |
| 877 | t.Fatalf("new router: %v", err) |
| 878 | } |
| 879 | |
| 880 | // Start a request-less (due-time) wake and hold it in the waking state. |
| 881 | b.startGate = make(chan struct{}) |
| 882 | if err := reg.setNextDue("mid.example.com", clk.now().Add(-time.Second).Unix()); err != nil { |
| 883 | t.Fatal(err) |
| 884 | } |
| 885 | a.sweepOnce() |
| 886 | waitFor(t, "the wake to begin", func() bool { _, _, s := b.counts(); return s == 1 }) |
| 887 | |
| 888 | // Flip off mid-wake: the reload's retire must defer, not deadlock. |
| 889 | if err := reg.setHibernate("mid.example.com", false); err != nil { |
| 890 | t.Fatal(err) |
| 891 | } |
| 892 | if err := rt.reload(); err != nil { |
| 893 | t.Fatalf("reload: %v", err) |
| 894 | } |
| 895 | close(b.startGate) |
| 896 | |
| 897 | waitFor(t, "post-wake retire", func() bool { return a.entryCount() == 0 }) |
| 898 | if !b.lastProc().isStopped() { |
| 899 | t.Fatal("child woken mid-retire was not stopped") |
| 900 | } |
| 901 | if !b.lastRep().isClosed() { |
| 902 | t.Fatal("child woken mid-retire left replication open (lease never released)") |
| 903 | } |
| 904 | } |
| 905 | |
| 906 | // With hibernation on, sibling services (kind "service") are untouched: the |
| 907 | // router proxies them directly, no wake, no activity tracking. |
| 908 | func TestActivatorServiceRoutesBypass(t *testing.T) { |
| 909 | sock := tmpSock(t) |
| 910 | ln, err := net.Listen("unix", sock) |
| 911 | if err != nil { |
| 912 | t.Fatalf("listen unix: %v", err) |
| 913 | } |
| 914 | backend := &http.Server{Handler: http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { |
| 915 | io.WriteString(w, "service ok") |
| 916 | })} |
| 917 | go backend.Serve(ln) |
| 918 | t.Cleanup(func() { backend.Close() }) |
| 919 | |
| 920 | a, b, _, _ := testActivator(t, time.Minute) |
| 921 | reg, err := openRegistry(filepath.Join(t.TempDir(), "registry.db")) |
| 922 | if err != nil { |
| 923 | t.Fatalf("open registry: %v", err) |
| 924 | } |
| 925 | if err := reg.upsert(Route{Host: "svc.example.com", Kind: "service", Socket: sock}); err != nil { |
| 926 | t.Fatalf("upsert: %v", err) |
| 927 | } |
| 928 | rt, err := newRouter(reg, a) |
| 929 | if err != nil { |
| 930 | t.Fatalf("new router: %v", err) |
| 931 | } |
| 932 | front := httptest.NewServer(rt) |
| 933 | t.Cleanup(front.Close) |
| 934 | |
| 935 | req, _ := http.NewRequest("GET", front.URL+"/", nil) |
| 936 | req.Host = "svc.example.com" |
| 937 | resp, err := http.DefaultClient.Do(req) |
| 938 | if err != nil { |
| 939 | t.Fatalf("request: %v", err) |
| 940 | } |
| 941 | body, _ := io.ReadAll(resp.Body) |
| 942 | resp.Body.Close() |
| 943 | if resp.StatusCode != 200 || string(body) != "service ok" { |
| 944 | t.Fatalf("got %d %q, want 200 service ok", resp.StatusCode, body) |
| 945 | } |
| 946 | if acquires, _, _ := b.counts(); acquires != 0 { |
| 947 | t.Fatal("a service route touched the activator") |
| 948 | } |
| 949 | } |
| 950 | |
| 951 | // shutdown checkpoints and stops every awake child before the router exits. |
| 952 | func TestActivatorShutdownStopsChildren(t *testing.T) { |
| 953 | a, b, _, _ := testActivator(t, time.Hour) |
| 954 | for _, h := range []string{"one.example.com", "two.example.com"} { |
| 955 | release, err := a.beginRequest(context.Background(), testRoute(h)) |
| 956 | if err != nil { |
| 957 | t.Fatalf("beginRequest %s: %v", h, err) |
| 958 | } |
| 959 | release() |
| 960 | } |
| 961 | |
| 962 | ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second) |
| 963 | defer cancel() |
| 964 | a.shutdown(ctx) |
| 965 | |
| 966 | if a.entryCount() != 0 { |
| 967 | t.Fatal("shutdown left entries behind") |
| 968 | } |
| 969 | b.mu.Lock() |
| 970 | defer b.mu.Unlock() |
| 971 | for i, p := range b.procs { |
| 972 | if !p.stopped { |
| 973 | t.Fatalf("proc %d not stopped on shutdown", i) |
| 974 | } |
| 975 | } |
| 976 | for i, r := range b.reps { |
| 977 | if !r.closed { |
| 978 | t.Fatalf("replication %d not closed on shutdown", i) |
| 979 | } |
| 980 | } |
| 981 | } |
| 982 | |
| 983 | // --- due-time index -------------------------------------------------------------- |
| 984 | |
| 985 | // entryAwake reports whether host has a live entry in the awake state — how the |
| 986 | // due tests observe an async due-time wake resolving. |
| 987 | func (a *activatorSvc) entryAwake(host string) bool { |
| 988 | a.mu.Lock() |
| 989 | defer a.mu.Unlock() |
| 990 | e := a.entries[host] |
| 991 | return e != nil && e.state == actAwake |
| 992 | } |
| 993 | |
| 994 | // The full self-maintaining cycle: sleep records the instance's next due time |
| 995 | // in the registry; the sweeper wakes the host (no request) when it arrives; the |
| 996 | // following sleep re-records — and once nothing is scheduled, no more due wakes. |
| 997 | func TestActivatorDueTimeWakeCycle(t *testing.T) { |
| 998 | a, b, clk, reg := testActivator(t, 10*time.Second) |
| 999 | route := testRoute("due.example.com") |
| 1000 | if err := reg.upsert(route); err != nil { |
| 1001 | t.Fatalf("upsert: %v", err) |
| 1002 | } |
| 1003 | |
| 1004 | // Wake by request; the instance has a delivery scheduled 30s out. |
| 1005 | dueAt := clk.now().Add(30 * time.Second).Unix() |
| 1006 | b.setDue(dueAt) |
| 1007 | release, err := a.beginRequest(context.Background(), route) |
| 1008 | if err != nil { |
| 1009 | t.Fatalf("beginRequest: %v", err) |
| 1010 | } |
| 1011 | release() |
| 1012 | |
| 1013 | // Idle out: the sleep probes the instance and records its due time. |
| 1014 | clk.advance(11 * time.Second) |
| 1015 | a.sweepOnce() |
| 1016 | waitFor(t, "hibernation", func() bool { return a.entryCount() == 0 }) |
| 1017 | if rt, err := reg.get(route.Host); err != nil || rt == nil || rt.NextDue != dueAt { |
| 1018 | t.Fatalf("registry next_due = %+v, %v; want %d", rt, err, dueAt) |
| 1019 | } |
| 1020 | |
| 1021 | // Before the due time: sweeps leave it asleep. |
| 1022 | a.sweepOnce() |
| 1023 | if _, _, starts := b.counts(); starts != 1 { |
| 1024 | t.Fatalf("woke before its due time (starts=%d)", starts) |
| 1025 | } |
| 1026 | |
| 1027 | // At the due time: the sweeper wakes it with no client request. |
| 1028 | clk.advance(20 * time.Second) // now = wake+31s ≥ due |
| 1029 | a.sweepOnce() |
| 1030 | waitFor(t, "due-time wake", func() bool { return a.entryAwake(route.Host) }) |
| 1031 | if _, _, starts := b.counts(); starts != 2 { |
| 1032 | t.Fatalf("starts=%d after due wake, want 2", starts) |
| 1033 | } |
| 1034 | |
| 1035 | // The woken instance stays up for the normal idle window, then sleeps and |
| 1036 | // re-records — nothing scheduled now, so the index clears and stays quiet. |
| 1037 | b.setDue(0) |
| 1038 | clk.advance(11 * time.Second) |
| 1039 | a.sweepOnce() |
| 1040 | waitFor(t, "re-hibernation", func() bool { return a.entryCount() == 0 }) |
| 1041 | if rt, _ := reg.get(route.Host); rt.NextDue != 0 { |
| 1042 | t.Fatalf("next_due = %d after empty-schedule sleep, want 0", rt.NextDue) |
| 1043 | } |
| 1044 | clk.advance(time.Hour) |
| 1045 | a.sweepOnce() |
| 1046 | if _, _, starts := b.counts(); starts != 2 { |
| 1047 | t.Fatalf("starts=%d with nothing scheduled, want 2 (no phantom due wake)", starts) |
| 1048 | } |
| 1049 | } |
| 1050 | |
| 1051 | // A due time already in the past at sleep time (a delivery that raced the stop) |
| 1052 | // is recorded as-is: the sweeper wakes the host on its next tick — once — and |
| 1053 | // there is no ping-pong, because a woken instance can't sleep again until a |
| 1054 | // full idle window passes. |
| 1055 | func TestActivatorDuePastWakesOnceNoPingPong(t *testing.T) { |
| 1056 | a, b, clk, reg := testActivator(t, 10*time.Second) |
| 1057 | route := testRoute("late.example.com") |
| 1058 | if err := reg.upsert(route); err != nil { |
| 1059 | t.Fatalf("upsert: %v", err) |
| 1060 | } |
| 1061 | b.setDue(clk.now().Add(-5 * time.Second).Unix()) // overdue before it even sleeps |
| 1062 | |
| 1063 | release, err := a.beginRequest(context.Background(), route) |
| 1064 | if err != nil { |
| 1065 | t.Fatalf("beginRequest: %v", err) |
| 1066 | } |
| 1067 | release() |
| 1068 | clk.advance(11 * time.Second) |
| 1069 | a.sweepOnce() |
| 1070 | waitFor(t, "hibernation", func() bool { return a.entryCount() == 0 }) |
| 1071 | |
| 1072 | // Next tick: woken for the overdue delivery. |
| 1073 | a.sweepOnce() |
| 1074 | waitFor(t, "overdue wake", func() bool { return a.entryAwake(route.Host) }) |
| 1075 | |
| 1076 | // No thrash: with the wake fresh (clock unmoved), repeated sweeps neither |
| 1077 | // re-sleep it nor double-wake it, even though next_due is still in the past. |
| 1078 | for i := 0; i < 5; i++ { |
| 1079 | a.sweepOnce() |
| 1080 | } |
| 1081 | if a.entryCount() != 1 || !a.entryAwake(route.Host) { |
| 1082 | t.Fatal("due-woken instance did not stay awake through its idle window") |
| 1083 | } |
| 1084 | if _, _, starts := b.counts(); starts != 2 { |
| 1085 | t.Fatalf("starts=%d, want 2 (request wake + one due wake)", starts) |
| 1086 | } |
| 1087 | } |
| 1088 | |
| 1089 | // A failed next-due probe (dead socket, missing admin token, …) records 0 and |
| 1090 | // the sleep proceeds — the instance degrades to "sleeps until a request", |
| 1091 | // never a crash and never a stuck stop. |
| 1092 | func TestActivatorDueProbeFailureRecordsZero(t *testing.T) { |
| 1093 | a, b, clk, reg := testActivator(t, 10*time.Second) |
| 1094 | route := testRoute("deaf.example.com") |
| 1095 | if err := reg.upsert(route); err != nil { |
| 1096 | t.Fatalf("upsert: %v", err) |
| 1097 | } |
| 1098 | b.dueErr = errors.New("dial unix: no such file") |
| 1099 | |
| 1100 | release, err := a.beginRequest(context.Background(), route) |
| 1101 | if err != nil { |
| 1102 | t.Fatalf("beginRequest: %v", err) |
| 1103 | } |
| 1104 | release() |
| 1105 | clk.advance(11 * time.Second) |
| 1106 | a.sweepOnce() |
| 1107 | waitFor(t, "hibernation despite probe failure", func() bool { return a.entryCount() == 0 }) |
| 1108 | |
| 1109 | if rt, _ := reg.get(route.Host); rt.NextDue != 0 { |
| 1110 | t.Fatalf("next_due = %d after failed probe, want 0", rt.NextDue) |
| 1111 | } |
| 1112 | clk.advance(time.Hour) |
| 1113 | a.sweepOnce() |
| 1114 | if _, _, starts := b.counts(); starts != 1 { |
| 1115 | t.Fatalf("starts=%d, want 1 (a failed probe must not invent a wake)", starts) |
| 1116 | } |
| 1117 | } |
| 1118 | |
| 1119 | // The index lives in the registry, so a due time recorded before a router |
| 1120 | // restart still fires: a brand-new activator over the same registry — no |
| 1121 | // in-memory state carried over — wakes the host when its time arrives. |
| 1122 | func TestActivatorDueSurvivesRestart(t *testing.T) { |
| 1123 | a1, b1, clk, reg := testActivator(t, 10*time.Second) |
| 1124 | route := testRoute("reborn.example.com") |
| 1125 | if err := reg.upsert(route); err != nil { |
| 1126 | t.Fatalf("upsert: %v", err) |
| 1127 | } |
| 1128 | b1.setDue(clk.now().Add(30 * time.Second).Unix()) |
| 1129 | |
| 1130 | release, err := a1.beginRequest(context.Background(), route) |
| 1131 | if err != nil { |
| 1132 | t.Fatalf("beginRequest: %v", err) |
| 1133 | } |
| 1134 | release() |
| 1135 | clk.advance(11 * time.Second) |
| 1136 | a1.sweepOnce() |
| 1137 | waitFor(t, "hibernation", func() bool { return a1.entryCount() == 0 }) |
| 1138 | |
| 1139 | // "Restart": a fresh activator, same registry. |
| 1140 | b2 := &fakeBackend{} |
| 1141 | a2 := newActivator(hibernateConfig{Bucket: "test-bucket", Idle: 10 * time.Second}, b2, reg) |
| 1142 | a2.now = clk.now |
| 1143 | |
| 1144 | clk.advance(25 * time.Second) // past the due time |
| 1145 | a2.sweepOnce() |
| 1146 | waitFor(t, "post-restart due wake", func() bool { return a2.entryAwake(route.Host) }) |
| 1147 | if _, _, starts := b2.counts(); starts != 1 { |
| 1148 | t.Fatalf("starts=%d on the new activator, want 1", starts) |
| 1149 | } |
| 1150 | } |
| 1151 | |
| 1152 | // --- bot sidecar co-spawn ---------------------------------------------------- |
| 1153 | |
| 1154 | // indexOf: position of ev in the backend's order log (-1 = absent). |
| 1155 | func indexOf(events []string, ev string) int { |
| 1156 | for i, e := range events { |
| 1157 | if e == ev { |
| 1158 | return i |
| 1159 | } |
| 1160 | } |
| 1161 | return -1 |
| 1162 | } |
| 1163 | |
| 1164 | // A host with a sidecar staged gets it co-spawned on wake, strictly AFTER the |
| 1165 | // instance process is up — the sidecar dials the instance, so spawning it |
| 1166 | // first would just burn its reconnect backoff. (Every other test in this file |
| 1167 | // runs with sidecarStaged=false and asserts nothing sidecar-ish happens, which |
| 1168 | // is the "host without a sidecar behaves exactly as today" half.) |
| 1169 | func TestActivatorCoSpawnsSidecarAfterInstance(t *testing.T) { |
| 1170 | a, b, _, _ := testActivator(t, time.Minute) |
| 1171 | b.sidecarStaged = true |
| 1172 | |
| 1173 | release, err := a.beginRequest(context.Background(), testRoute("bots.example.com")) |
| 1174 | if err != nil { |
| 1175 | t.Fatalf("beginRequest: %v", err) |
| 1176 | } |
| 1177 | release() |
| 1178 | |
| 1179 | if got := b.sidecarStartCount(); got != 1 { |
| 1180 | t.Fatalf("sidecar starts = %d, want 1", got) |
| 1181 | } |
| 1182 | ev := b.events() |
| 1183 | if i, s := indexOf(ev, "start:instance"), indexOf(ev, "start:sidecar"); i == -1 || s == -1 || s < i { |
| 1184 | t.Fatalf("sidecar did not start strictly after the instance: %v", ev) |
| 1185 | } |
| 1186 | } |
| 1187 | |
| 1188 | // A sidecar that fails to spawn must not fail the wake — members > bots: the |
| 1189 | // space serves without its bots, and the sweeper keeps retrying with backoff |
| 1190 | // until a spawn sticks. |
| 1191 | func TestActivatorSidecarFailureDoesNotFailWake(t *testing.T) { |
| 1192 | a, b, clk, _ := testActivator(t, time.Hour) |
| 1193 | b.sidecarStaged = true |
| 1194 | b.setSidecarErr(errors.New("exec: live symlink gone")) |
| 1195 | route := testRoute("botless.example.com") |
| 1196 | |
| 1197 | release, err := a.beginRequest(context.Background(), route) |
| 1198 | if err != nil { |
| 1199 | t.Fatalf("beginRequest with a failing sidecar: %v", err) |
| 1200 | } |
| 1201 | release() |
| 1202 | if !a.entryAwake(route.Host) { |
| 1203 | t.Fatal("instance not awake after a sidecar-only failure") |
| 1204 | } |
| 1205 | if got := b.sidecarStartCount(); got != 1 { |
| 1206 | t.Fatalf("sidecar attempts = %d, want 1 (the wake's)", got) |
| 1207 | } |
| 1208 | |
| 1209 | // The sweeper retries: attempt 2 immediately (first retry), then gated by |
| 1210 | // the doubling backoff; once the fault clears, a retry lands the sidecar. |
| 1211 | a.sweepOnce() |
| 1212 | if got := b.sidecarStartCount(); got != 2 { |
| 1213 | t.Fatalf("sidecar attempts after first sweep = %d, want 2", got) |
| 1214 | } |
| 1215 | a.sweepOnce() // still inside the 2s gate — no attempt |
| 1216 | if got := b.sidecarStartCount(); got != 2 { |
| 1217 | t.Fatalf("backoff gate ignored: attempts = %d, want 2", got) |
| 1218 | } |
| 1219 | b.setSidecarErr(nil) |
| 1220 | clk.advance(3 * time.Second) |
| 1221 | a.sweepOnce() |
| 1222 | if got := b.sidecarStartCount(); got != 3 { |
| 1223 | t.Fatalf("sidecar attempts after fault cleared = %d, want 3", got) |
| 1224 | } |
| 1225 | if b.lastSidecar().isStopped() || b.lastSidecar().exited() { |
| 1226 | t.Fatal("recovered sidecar isn't running") |
| 1227 | } |
| 1228 | } |
| 1229 | |
| 1230 | // Co-stop ordering at sleep: the sidecar is stopped FIRST — before the |
| 1231 | // next-due probe (so bots can't shift the schedule being recorded) and before |
| 1232 | // the instance's own stop (so the sidecar never outlives its instance). |
| 1233 | func TestActivatorCoStopOrdering(t *testing.T) { |
| 1234 | a, b, clk, _ := testActivator(t, 10*time.Second) |
| 1235 | b.sidecarStaged = true |
| 1236 | |
| 1237 | release, err := a.beginRequest(context.Background(), testRoute("bots.example.com")) |
| 1238 | if err != nil { |
| 1239 | t.Fatalf("beginRequest: %v", err) |
| 1240 | } |
| 1241 | release() |
| 1242 | |
| 1243 | clk.advance(11 * time.Second) |
| 1244 | a.sweepOnce() |
| 1245 | waitFor(t, "hibernation", func() bool { return a.entryCount() == 0 }) |
| 1246 | |
| 1247 | if !b.lastSidecar().isStopped() { |
| 1248 | t.Fatal("sidecar not stopped at sleep") |
| 1249 | } |
| 1250 | ev := b.events() |
| 1251 | sc, due, inst := indexOf(ev, "stop:sidecar"), indexOf(ev, "nextDue"), indexOf(ev, "stop:instance") |
| 1252 | if sc == -1 || due == -1 || inst == -1 || !(sc < due && due < inst) { |
| 1253 | t.Fatalf("co-stop order wrong: %v (want stop:sidecar < nextDue < stop:instance)", ev) |
| 1254 | } |
| 1255 | } |
| 1256 | |
| 1257 | // Orphan prevention: whatever takes the instance down — here a crashed child, |
| 1258 | // but a lost lease and a retire ride the identical stop path — takes the |
| 1259 | // sidecar with it. An orphaned sidecar would reconnect-loop against the router |
| 1260 | // forever (503s at best; a wake source if the marker rule ever broke). |
| 1261 | func TestActivatorInstanceDeathStopsSidecar(t *testing.T) { |
| 1262 | a, b, _, _ := testActivator(t, time.Hour) |
| 1263 | b.sidecarStaged = true |
| 1264 | |
| 1265 | release, err := a.beginRequest(context.Background(), testRoute("bots.example.com")) |
| 1266 | if err != nil { |
| 1267 | t.Fatalf("beginRequest: %v", err) |
| 1268 | } |
| 1269 | release() |
| 1270 | |
| 1271 | p := b.lastProc() |
| 1272 | p.mu.Lock() |
| 1273 | p.dead = true |
| 1274 | p.mu.Unlock() |
| 1275 | a.sweepOnce() |
| 1276 | waitFor(t, "stand-down", func() bool { return a.entryCount() == 0 }) |
| 1277 | if !b.lastSidecar().isStopped() { |
| 1278 | t.Fatal("instance died but its sidecar was left running (orphan)") |
| 1279 | } |
| 1280 | } |
| 1281 | |
| 1282 | // Sidecar death while the instance stays awake: restarted with doubling, |
| 1283 | // capped backoff — Restart=always semantics, sweep-driven — and a run that |
| 1284 | // held for sidecarHealthyRun earns a fresh backoff. |
| 1285 | func TestActivatorSidecarRestartBackoff(t *testing.T) { |
| 1286 | a, b, clk, _ := testActivator(t, time.Hour) |
| 1287 | b.sidecarStaged = true |
| 1288 | |
| 1289 | release, err := a.beginRequest(context.Background(), testRoute("bots.example.com")) |
| 1290 | if err != nil { |
| 1291 | t.Fatalf("beginRequest: %v", err) |
| 1292 | } |
| 1293 | release() |
| 1294 | kill := func() { |
| 1295 | sc := b.lastSidecar() |
| 1296 | sc.mu.Lock() |
| 1297 | sc.dead = true |
| 1298 | sc.mu.Unlock() |
| 1299 | } |
| 1300 | |
| 1301 | // T0: first crash — restarted on the next sweep (attempt 2), gate 2s. |
| 1302 | kill() |
| 1303 | a.sweepOnce() |
| 1304 | if got := b.sidecarStartCount(); got != 2 { |
| 1305 | t.Fatalf("first restart: attempts = %d, want 2", got) |
| 1306 | } |
| 1307 | // Crash-loop: still inside the 2s gate — no restart. |
| 1308 | kill() |
| 1309 | a.sweepOnce() |
| 1310 | if got := b.sidecarStartCount(); got != 2 { |
| 1311 | t.Fatalf("gate ignored: attempts = %d, want 2", got) |
| 1312 | } |
| 1313 | // T3: past the 2s gate — attempt 3, gate doubles to 4s. |
| 1314 | clk.advance(3 * time.Second) |
| 1315 | a.sweepOnce() |
| 1316 | if got := b.sidecarStartCount(); got != 3 { |
| 1317 | t.Fatalf("second restart: attempts = %d, want 3", got) |
| 1318 | } |
| 1319 | // T6: 3s later is inside the 4s gate — the doubling is real. |
| 1320 | kill() |
| 1321 | clk.advance(3 * time.Second) |
| 1322 | a.sweepOnce() |
| 1323 | if got := b.sidecarStartCount(); got != 3 { |
| 1324 | t.Fatalf("doubled gate ignored: attempts = %d, want 3", got) |
| 1325 | } |
| 1326 | // T8: past it — attempt 4 (gate now 8s). |
| 1327 | clk.advance(2 * time.Second) |
| 1328 | a.sweepOnce() |
| 1329 | if got := b.sidecarStartCount(); got != 4 { |
| 1330 | t.Fatalf("third restart: attempts = %d, want 4", got) |
| 1331 | } |
| 1332 | // A healthy run (≥ sidecarHealthyRun) resets the ladder: the next crash |
| 1333 | // restarts immediately and its gate is back to the 2s base… |
| 1334 | clk.advance(2 * time.Minute) |
| 1335 | kill() |
| 1336 | a.sweepOnce() |
| 1337 | if got := b.sidecarStartCount(); got != 5 { |
| 1338 | t.Fatalf("post-healthy restart: attempts = %d, want 5", got) |
| 1339 | } |
| 1340 | // …which a 3s advance clears (16s — the unreset ladder — would not). |
| 1341 | kill() |
| 1342 | clk.advance(3 * time.Second) |
| 1343 | a.sweepOnce() |
| 1344 | if got := b.sidecarStartCount(); got != 6 { |
| 1345 | t.Fatalf("backoff not reset after a healthy run: attempts = %d, want 6", got) |
| 1346 | } |
| 1347 | } |
| 1348 | |
| 1349 | // The activity question (#550 Part 4 v2): the co-spawned sidecar dials its |
| 1350 | // instance's public URL through this very router (the box pins its hosted |
| 1351 | // names to loopback), so without the marker its keepalives would be member |
| 1352 | // activity and the pair could never sleep. Full front-door check: marked |
| 1353 | // requests are served while awake — with the marker stripped before the |
| 1354 | // backend — but stamp nothing (the host hibernates on schedule beneath them), |
| 1355 | // and once asleep they get 503 and wake nothing. An outsider spoofing the |
| 1356 | // marker gets exactly that same treatment: it can't wake the host and can't |
| 1357 | // suppress anyone else's stamps (stamps are per-request), so the marker is |
| 1358 | // not a DoS lever in either direction. |
| 1359 | func TestActivatorSidecarRequestsInvisibleToActivity(t *testing.T) { |
| 1360 | a, b, clk, reg := testActivator(t, 10*time.Second) |
| 1361 | |
| 1362 | // A live backend listener stands in for the child a wake would have |
| 1363 | // spawned (the fake backend's start() spawns nothing), recording whether |
| 1364 | // the sidecar marker ever leaks through the proxy. |
| 1365 | sock := tmpSock(t) |
| 1366 | ln, err := net.Listen("unix", sock) |
| 1367 | if err != nil { |
| 1368 | t.Fatalf("listen unix: %v", err) |
| 1369 | } |
| 1370 | var leaked atomic.Bool |
| 1371 | srv := &http.Server{Handler: http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { |
| 1372 | if r.Header.Get(bot.SidecarHeader) != "" { |
| 1373 | leaked.Store(true) |
| 1374 | } |
| 1375 | io.WriteString(w, "ok") |
| 1376 | })} |
| 1377 | go srv.Serve(ln) |
| 1378 | t.Cleanup(func() { srv.Close() }) |
| 1379 | |
| 1380 | const host = "bots.example.com" |
| 1381 | if err := reg.upsert(Route{Host: host, Kind: "instance", Socket: sock, Hibernate: true}); err != nil { |
| 1382 | t.Fatal(err) |
| 1383 | } |
| 1384 | rt, err := newRouter(reg, a) |
| 1385 | if err != nil { |
| 1386 | t.Fatalf("new router: %v", err) |
| 1387 | } |
| 1388 | front := httptest.NewServer(rt) |
| 1389 | t.Cleanup(front.Close) |
| 1390 | |
| 1391 | markedGet := func() int { |
| 1392 | req, _ := http.NewRequest("GET", front.URL+"/", nil) |
| 1393 | req.Host = host |
| 1394 | req.Header.Set(bot.SidecarHeader, "1") |
| 1395 | resp, err := http.DefaultClient.Do(req) |
| 1396 | if err != nil { |
| 1397 | t.Fatalf("marked request: %v", err) |
| 1398 | } |
| 1399 | defer resp.Body.Close() |
| 1400 | io.Copy(io.Discard, resp.Body) |
| 1401 | return resp.StatusCode |
| 1402 | } |
| 1403 | |
| 1404 | // A member request wakes the host. |
| 1405 | if body, code := proxyGet(t, front.URL, host); code != 200 || body != "ok" { |
| 1406 | t.Fatalf("member wake: got %d %q", code, body) |
| 1407 | } |
| 1408 | // Marked requests are served while awake… |
| 1409 | if code := markedGet(); code != 200 { |
| 1410 | t.Fatalf("marked request while awake: %d, want 200", code) |
| 1411 | } |
| 1412 | // …but keep stamping nothing: sidecar traffic riding through the whole |
| 1413 | // idle window must not delay the sleep. |
| 1414 | clk.advance(6 * time.Second) |
| 1415 | if code := markedGet(); code != 200 { |
| 1416 | t.Fatalf("marked request mid-window: %d, want 200", code) |
| 1417 | } |
| 1418 | clk.advance(5 * time.Second) // 11s since the member request |
| 1419 | a.sweepOnce() |
| 1420 | waitFor(t, "hibernation beneath sidecar traffic", func() bool { return a.entryCount() == 0 }) |
| 1421 | |
| 1422 | // Asleep: marked requests are refused and wake NOTHING — this is also the |
| 1423 | // outsider-spoof case, and it's harmless. |
| 1424 | if code := markedGet(); code != 503 { |
| 1425 | t.Fatalf("marked request while asleep: %d, want 503", code) |
| 1426 | } |
| 1427 | if a.entryCount() != 0 { |
| 1428 | t.Fatal("a marked request created an entry (a wake) for a sleeping host") |
| 1429 | } |
| 1430 | if _, _, starts := b.counts(); starts != 1 { |
| 1431 | t.Fatalf("starts = %d after marked requests, want 1 (no sidecar-triggered wake)", starts) |
| 1432 | } |
| 1433 | // An unmarked request still wakes it, and the marker never reached the |
| 1434 | // backend on any request. |
| 1435 | if _, code := proxyGet(t, front.URL, host); code != 200 { |
| 1436 | t.Fatalf("member re-wake: %d, want 200", code) |
| 1437 | } |
| 1438 | if _, _, starts := b.counts(); starts != 2 { |
| 1439 | t.Fatalf("starts = %d after member re-wake, want 2", starts) |
| 1440 | } |
| 1441 | if leaked.Load() { |
| 1442 | t.Fatal("the sidecar marker leaked through the proxy to the backend") |
| 1443 | } |
| 1444 | } |
| 1445 | |
| 1446 | func TestHibernateConfigFromEnv(t *testing.T) { |
| 1447 | t.Setenv("LCHAT_HIBERNATE_BUCKET", "") |
| 1448 | if cfg, err := hibernateConfigFromEnv(); err != nil || cfg != nil { |
| 1449 | t.Fatalf("unset bucket: got %+v, %v; want nil, nil (feature off)", cfg, err) |
| 1450 | } |
| 1451 | |
| 1452 | t.Setenv("LCHAT_HIBERNATE_BUCKET", "b") |
| 1453 | t.Setenv("LCHAT_HIBERNATE_PREFIX", "/instances/") |
| 1454 | t.Setenv("LCHAT_HIBERNATE_IDLE", "90s") |
| 1455 | cfg, err := hibernateConfigFromEnv() |
| 1456 | if err != nil { |
| 1457 | t.Fatalf("config: %v", err) |
| 1458 | } |
| 1459 | if cfg.Bucket != "b" || cfg.Prefix != "instances" || cfg.Idle != 90*time.Second { |
| 1460 | t.Fatalf("got %+v", cfg) |
| 1461 | } |
| 1462 | if got := cfg.hibernatePrefix("x.example.com"); got != "instances/x.example.com" { |
| 1463 | t.Fatalf("prefix: %q", got) |
| 1464 | } |
| 1465 | |
| 1466 | t.Setenv("LCHAT_HIBERNATE_IDLE", "") |
| 1467 | cfg, err = hibernateConfigFromEnv() |
| 1468 | if err != nil || cfg.Idle != 15*time.Minute { |
| 1469 | t.Fatalf("default idle: %+v, %v", cfg, err) |
| 1470 | } |
| 1471 | |
| 1472 | t.Setenv("LCHAT_HIBERNATE_IDLE", "banana") |
| 1473 | if _, err := hibernateConfigFromEnv(); err == nil { |
| 1474 | t.Fatal("bad idle duration should error, not default silently") |
| 1475 | } |
| 1476 | } |
| 1477 | |
| 1478 | // --- ensureLiveLink (the controld-mediated fix for mesh-replicated routes) --- |
| 1479 | |
| 1480 | // ensureLiveLinkTestEnv redirects optDir/selfPath into temp dirs, same seam |
| 1481 | // controld_test.go's controlTestEnv uses, so ensureSelfStaged/pointLive run |
| 1482 | // for real against throwaway paths. |
| 1483 | func ensureLiveLinkTestEnv(t *testing.T) { |
| 1484 | t.Helper() |
| 1485 | oldOpt, oldSelf := optDir, selfPath |
| 1486 | t.Cleanup(func() { optDir, selfPath = oldOpt, oldSelf }) |
| 1487 | optDir = t.TempDir() |
| 1488 | stub := filepath.Join(t.TempDir(), "stub") |
| 1489 | if err := os.WriteFile(stub, []byte("#!/bin/sh\n"), 0o755); err != nil { |
| 1490 | t.Fatal(err) |
| 1491 | } |
| 1492 | selfPath = func() (string, error) { return stub, nil } |
| 1493 | } |
| 1494 | |
| 1495 | // Already-materialized: the common case for every existing box A instance — |
| 1496 | // neither the direct path nor materialize is ever invoked. |
| 1497 | func TestEnsureLiveLinkAlreadyMaterialized(t *testing.T) { |
| 1498 | ensureLiveLinkTestEnv(t) |
| 1499 | ver, err := ensureSelfStaged() |
| 1500 | if err != nil { |
| 1501 | t.Fatal(err) |
| 1502 | } |
| 1503 | if err := pointLive("existing.example.com", ver); err != nil { |
| 1504 | t.Fatal(err) |
| 1505 | } |
| 1506 | called := false |
| 1507 | materialize := func(context.Context, string) error { called = true; return nil } |
| 1508 | if err := ensureLiveLink(context.Background(), "existing.example.com", materialize); err != nil { |
| 1509 | t.Fatalf("ensureLiveLink: %v", err) |
| 1510 | } |
| 1511 | if called { |
| 1512 | t.Error("materialize was called for an already-materialized host") |
| 1513 | } |
| 1514 | } |
| 1515 | |
| 1516 | // No materialize hook configured (no control token — the common case for any |
| 1517 | // box that hasn't opted in): falls back to the direct calls, today's |
| 1518 | // long-standing behavior. |
| 1519 | func TestEnsureLiveLinkFallsBackWithoutMaterialize(t *testing.T) { |
| 1520 | ensureLiveLinkTestEnv(t) |
| 1521 | if err := ensureLiveLink(context.Background(), "fresh.example.com", nil); err != nil { |
| 1522 | t.Fatalf("ensureLiveLink: %v", err) |
| 1523 | } |
| 1524 | if _, err := os.Stat(liveLink("fresh.example.com")); err != nil { |
| 1525 | t.Fatalf("live symlink was not created: %v", err) |
| 1526 | } |
| 1527 | } |
| 1528 | |
| 1529 | // Missing symlink + materialize configured: routed through materialize, not |
| 1530 | // the direct (sandboxed-and-failing, in production) calls. |
| 1531 | func TestEnsureLiveLinkUsesMaterializeWhenConfigured(t *testing.T) { |
| 1532 | ensureLiveLinkTestEnv(t) |
| 1533 | var gotHost string |
| 1534 | materialize := func(_ context.Context, host string) error { |
| 1535 | gotHost = host |
| 1536 | // The real controld handler stages + points live; a fake only needs |
| 1537 | // to prove IT was the one asked, and that its result is honored. |
| 1538 | ver, err := ensureSelfStaged() |
| 1539 | if err != nil { |
| 1540 | return err |
| 1541 | } |
| 1542 | return pointLive(host, ver) |
| 1543 | } |
| 1544 | if err := ensureLiveLink(context.Background(), "seeded.11ch.at", materialize); err != nil { |
| 1545 | t.Fatalf("ensureLiveLink: %v", err) |
| 1546 | } |
| 1547 | if gotHost != "seeded.11ch.at" { |
| 1548 | t.Fatalf("materialize called with %q, want seeded.11ch.at", gotHost) |
| 1549 | } |
| 1550 | if _, err := os.Stat(liveLink("seeded.11ch.at")); err != nil { |
| 1551 | t.Fatalf("live symlink was not created via materialize: %v", err) |
| 1552 | } |
| 1553 | } |
| 1554 | |
| 1555 | // A materialize failure (e.g. controld unreachable) must propagate — the |
| 1556 | // exact live symptom found 2026-07-13 (a 503 "instance failed to wake"), |
| 1557 | // never silently swallowed into "looks fine, wakes anyway." |
| 1558 | func TestEnsureLiveLinkPropagatesMaterializeError(t *testing.T) { |
| 1559 | ensureLiveLinkTestEnv(t) |
| 1560 | wantErr := fmt.Errorf("controld materialize seeded.11ch.at: HTTP 502") |
| 1561 | materialize := func(context.Context, string) error { return wantErr } |
| 1562 | err := ensureLiveLink(context.Background(), "seeded.11ch.at", materialize) |
| 1563 | if err == nil { |
| 1564 | t.Fatal("expected an error, got nil") |
| 1565 | } |
| 1566 | if _, statErr := os.Stat(liveLink("seeded.11ch.at")); statErr == nil { |
| 1567 | t.Fatal("live symlink should not exist after a failed materialize") |
| 1568 | } |
| 1569 | } |
| 1570 | |
amadan