| 1 | package rollout |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "crypto/rand" |
| 6 | "encoding/hex" |
| 7 | "errors" |
| 8 | "log" |
| 9 | "os" |
| 10 | "slices" |
| 11 | "strconv" |
| 12 | "strings" |
| 13 | "sync" |
| 14 | "time" |
| 15 | |
| 16 | "github.com/carlosframework/platform/internal/activator" |
| 17 | "github.com/carlosframework/platform/internal/appconfig" |
| 18 | "github.com/carlosframework/platform/internal/registry" |
| 19 | "github.com/carlosframework/platform/internal/release" |
| 20 | ) |
| 21 | |
| 22 | // Activator is the part of *activator.Service the scheduler drives. Swap is |
| 23 | // plan 3's, for instances; ReplaceWorker is Task 11's same-entry |
| 24 | // stop-then-start for workers and native sidecar members (spec §7.9), and |
| 25 | // PublishHeldServing releases a worker fallback's serving identity once |
| 26 | // done() has recorded the failure it falls back from (spec §8.1). |
| 27 | // EffectiveMode is the mode Swap will run an attempt of mode m on host in |
| 28 | // (an overlap on a foreign-layout row runs as a hold, spec §2.9), which the |
| 29 | // attempt's intent records before Swap runs (plan review finding 2). |
| 30 | type Activator interface { |
| 31 | Current(host string) (activator.Gen, bool) |
| 32 | Owns(host string) bool |
| 33 | Swap(ctx context.Context, a activator.Attempt, newest func(host string) uint64) activator.SwapResult |
| 34 | ReplaceWorker(ctx context.Context, a activator.Attempt, fallback *release.StagedIdentity, newest func(host string) uint64) activator.SwapResult |
| 35 | PublishHeldServing(host string) |
| 36 | EffectiveMode(host string, m activator.SwapMode) activator.SwapMode |
| 37 | } |
| 38 | |
| 39 | // Routes is the part of *registry.Registry the scheduler reads and writes. |
| 40 | type Routes interface { |
| 41 | Get(host string) (*registry.Route, error) |
| 42 | SetGoodBuild(host string, expect, next registry.GoodBuild) error |
| 43 | SetServingVersion(host, ver string) error |
| 44 | } |
| 45 | |
| 46 | // Deps is everything the manager needs from the box. Its funcs are always |
| 47 | // called without the manager's locks held. |
| 48 | type Deps struct { |
| 49 | Act Activator |
| 50 | Routes Routes |
| 51 | Book *FailureBook |
| 52 | OptDir string |
| 53 | DataDir string |
| 54 | // Box is this box's label (host.BoxLabel), named by the cohort-halted |
| 55 | // event (spec §9.4: app, box, build, reason). |
| 56 | Box string |
| 57 | // Demand is release.CanonicalDemand bound to the deploy store and this |
| 58 | // box's platform (spec §4.2). |
| 59 | Demand func(ctx context.Context, acct, app, version, stagedPath string) (demand, digest string, err error) |
| 60 | // Policy reads the app's default config bundle (spec §4.1). |
| 61 | Policy func(ctx context.Context, acct, app string) (*appconfig.Manifest, error) |
| 62 | // Pointer reads a channel's signed pointer, verified (release.Channel). |
| 63 | // It is the retry signal (spec rev 14 §8.4) and the PromotedAt a failure |
| 64 | // record keeps. Nil: records clear only by a newer installed build. It |
| 65 | // is the raw signed read, not the per-box digest ratchet: only its |
| 66 | // Version and PromotedAt are used, to clear failure records, and root's |
| 67 | // install, which the ratchet guards, decides which build is installed. |
| 68 | Pointer func(ctx context.Context, acct, app, channel string) (*release.ChannelPointer, error) |
| 69 | Events activator.EventSink |
| 70 | // Touch asks root for a stage pass (the stage relay, spec §7.7). |
| 71 | Touch func() |
| 72 | // Materialize installs and stamps a cold managed host's build (the exec |
| 73 | // backend's Materialize hook) so a wake is planned against it (Task 10, |
| 74 | // spec §6.4, §8.4). Nil: a cold managed host's wake is refused. |
| 75 | Materialize func(ctx context.Context, host string) error |
| 76 | // OnChange, when set, is told after every rollout transition (an |
| 77 | // attempt's outcome, a wake's event, a newly installed build, a retry), |
| 78 | // so the box republishes its status within seconds instead of at the |
| 79 | // next tick (plan 6's deploy watch). It is a nudge: it is called with no |
| 80 | // Manager or scheduler lock held, and it must not block. |
| 81 | OnChange func() |
| 82 | Now func() time.Time |
| 83 | Logf func(format string, args ...any) |
| 84 | } |
| 85 | |
| 86 | // pointerReadEvery bounds how often an app with a standing failure record |
| 87 | // has its channel pointers read for a retry. An app with no record costs no |
| 88 | // read at all. |
| 89 | const pointerReadEvery = 30 * time.Second |
| 90 | |
| 91 | // stagingDeferral is how long an attempt refused as a staging error waits: |
| 92 | // one stage-relay window, for root to restage (spec §4.2, §7.7). |
| 93 | const stagingDeferral = 15 * time.Second |
| 94 | |
| 95 | // abandonDeferral is how long a host waits after an attempt was abandoned |
| 96 | // while it was still the host's newest: a hold whose stop withdrew, a lost |
| 97 | // lease, a re-check that refused. Running the same attempt again at once |
| 98 | // would only repeat the stop cycle (plan 3's deferral to this plan). An |
| 99 | // attempt abandoned because a newer desire superseded it does not wait: the |
| 100 | // newer attempt goes as soon as it is admitted. |
| 101 | const abandonDeferral = 15 * time.Second |
| 102 | |
| 103 | // Manager is the box's rollout scheduler: one goroutine per (account, app), |
| 104 | // fed by the agent pass, the adopt pass, `carlos restart` and config sync. |
| 105 | // |
| 106 | // Lock order: goodMu, then activator.Service.mu, then Manager.mu, then |
| 107 | // scheduler.mu, then FailureBook.mu. The activator (Current, Owns, Swap, |
| 108 | // ReplaceWorker) is never called while Manager.mu or a scheduler.mu is held. |
| 109 | // goodMu comes first because served, publishServing and writeGood call |
| 110 | // Act.Current under it, and observeGood takes Manager.mu and then a |
| 111 | // scheduler.mu under it. So never call served, publishServing or writeGood |
| 112 | // (or OnGen, which reaches them) while holding activator.Service.mu: that |
| 113 | // caller would hold Service.mu and wait for goodMu while another holds goodMu |
| 114 | // and waits for Service.mu. |
| 115 | type Manager struct { |
| 116 | d Deps |
| 117 | ctx context.Context |
| 118 | policies *policyCache |
| 119 | |
| 120 | mu sync.Mutex |
| 121 | apps map[appKey]*scheduler |
| 122 | hostApp map[string]appKey |
| 123 | cfg map[string]*configTrack |
| 124 | retried map[string]string // host -> build a retry cleared: its next attempt's cause is retry |
| 125 | pointerRead map[appKey]time.Time |
| 126 | // nudging holds the hosts a Nudge is working on; true asks it to look |
| 127 | // once more for a nudge that arrived meanwhile. |
| 128 | nudging map[string]bool |
| 129 | // observing holds the hosts whose first wake is inside its 10 s |
| 130 | // observation, each with the token of the wake it observes: opened by |
| 131 | // that wake's healthy first-wake event, closed by an event carrying the |
| 132 | // same token (OnGen). Status reports them as settling. |
| 133 | observing map[string]uint64 |
| 134 | // planned holds each host whose planned first wake is in flight, and the |
| 135 | // build it boots: from its Plan to its outcome (an OnGen event carrying |
| 136 | // the plan's token), its next Plan, or the activator letting the host go |
| 137 | // (a Pass that finds Owns false). prune keeps those builds' failure |
| 138 | // records, because a wake's last selection check (planCheck) reads only |
| 139 | // the book: a record pruned after the plan, once a promotion moved every |
| 140 | // route off the build, would let the wake boot a build that failed on |
| 141 | // this box. plannedSeq tells one registration from the next for that |
| 142 | // Pass, which asks Owns off every lock. |
| 143 | planned map[string]plannedWake |
| 144 | plannedSeq uint64 |
| 145 | // wakeTok is each host's newest wake token: the token of the last plan |
| 146 | // made for it (final review F4). Tokens come from one counter, so they |
| 147 | // are unique and only grow. An event carrying any other token is an older |
| 148 | // wake's, and acts on nothing that belongs to the newer one. |
| 149 | wakeTok map[string]uint64 |
| 150 | tokenCount uint64 |
| 151 | // wakeAt holds, per host, the start of each planned wake that may still |
| 152 | // report (wakeStart, by token): its plan time, which a hold-ran marker |
| 153 | // the wake leaves carries as its At (ruling 205, Task 3 fix round 1), and |
| 154 | // the failure book's snapshot taken with it, which orders the wake's |
| 155 | // outcome against the book (Astra R3 N6, ruling 191). Kept per token, |
| 156 | // not only the newest plan's, because neither a marker nor a failure is |
| 157 | // token-scoped: a late event of an older wake still acts, as of its own |
| 158 | // start. An entry goes with its wake's last event, with the host |
| 159 | // (forget), or when the host has wakeStartsKept newer ones. |
| 160 | wakeAt map[string][]wakeStart |
| 161 | // execCount numbers the scheduled attempts' executions (nextExecID, |
| 162 | // dispatch.ExecID). |
| 163 | execCount uint64 |
| 164 | // inc is this manager's incarnation: 8 random bytes, hex, drawn once per |
| 165 | // process start (NewManager). Every execution identity carries it |
| 166 | // (nextExecID, wakeExecID), because the counters above start again at |
| 167 | // zero in a new process while the markers and intents they named |
| 168 | // survive in the book: without it the first wake after a restart could |
| 169 | // be "w1" again, and markLocked would take the new execution for the |
| 170 | // old one and keep its Seq and At, so a grant or a rollback that matched |
| 171 | // the old execution would clear the new one (final review round 1, |
| 172 | // finding 1, Astra). |
| 173 | inc string |
| 174 | |
| 175 | // ackSeq numbers the config acks the tracks hand back, under mu, in the |
| 176 | // critical section of the transition that releases each one: the |
| 177 | // writer's order (ackLocked, plan review #10). |
| 178 | ackSeq uint64 |
| 179 | // acks writes those acks, off the scheduler's slot and loop (Task 11). |
| 180 | acks *ackWriter |
| 181 | |
| 182 | // goodMu serialises every serving- and good-build write with its |
| 183 | // acknowledgement check (served, publishServing, writeGood), so an older |
| 184 | // generation's write can never land after a newer one's (spec §8.5). |
| 185 | goodMu sync.Mutex |
| 186 | |
| 187 | // The shutdown join (Close, spec §8.6). joinMu is a leaf: nothing else |
| 188 | // is taken while it is held, so it may be taken under any other lock |
| 189 | // (CoalesceRecycle takes it under activator.Service.mu and a |
| 190 | // scheduler.mu). inflight counts the attempts' executions, the Plan and |
| 191 | // OnGen calls under way, and the planned wakes in wakes; closed stops |
| 192 | // admission; joined, when set, is closed once inflight is zero. |
| 193 | joinMu sync.Mutex |
| 194 | closed bool |
| 195 | inflight int |
| 196 | joined chan struct{} |
| 197 | // wakes holds the token of each planned wake whose intent beginWake |
| 198 | // wrote and whose terminal event has not yet reached OnGen (Task 11 fix |
| 199 | // round 1). Each is one count in inflight, taken and given back exactly |
| 200 | // once whatever the order of plans and events. |
| 201 | wakes map[uint64]struct{} |
| 202 | } |
| 203 | |
| 204 | // NewManager builds the manager. Its schedulers run until ctx ends. |
| 205 | func NewManager(ctx context.Context, d Deps) *Manager { |
| 206 | if d.Now == nil { |
| 207 | d.Now = time.Now |
| 208 | } |
| 209 | if d.Logf == nil { |
| 210 | d.Logf = log.Printf |
| 211 | } |
| 212 | m := &Manager{d: d, ctx: ctx, inc: newIncarnation(), |
| 213 | apps: map[appKey]*scheduler{}, hostApp: map[string]appKey{}, cfg: map[string]*configTrack{}, |
| 214 | retried: map[string]string{}, pointerRead: map[appKey]time.Time{}, nudging: map[string]bool{}, observing: map[string]uint64{}, |
| 215 | planned: map[string]plannedWake{}, wakeTok: map[string]uint64{}, wakeAt: map[string][]wakeStart{}} |
| 216 | m.policies = &policyCache{load: d.Policy, ttl: 30 * time.Second, now: d.Now, m: map[appKey]*policyEntry{}} |
| 217 | m.acks = newAckWriter(ctx, d.Logf) |
| 218 | return m |
| 219 | } |
| 220 | |
| 221 | // Close is the scheduler's part of a graceful stop (spec §8.6: shutdown |
| 222 | // cancels and joins the whole operation set; Task 5 review). The agent calls |
| 223 | // it after the activator's Shutdown, which has stopped every entry and so |
| 224 | // cancelled any swap in flight. It stops admitting attempts, then waits, |
| 225 | // until ctx ends, for every attempt's execution to finish its done() (the |
| 226 | // outcome's book writes and the intent's End), for every Plan and OnGen |
| 227 | // call under way to return, and for every planned wake whose intent is on |
| 228 | // disk to deliver the terminal event that ends it, and then writes the acks |
| 229 | // still pending and stops the ack writer. A planned restart in the middle |
| 230 | // of a swap or a wake so leaves the outcome on disk and no intent behind, |
| 231 | // and the next agent reconciles nothing. A Close that runs out of ctx |
| 232 | // returns anyway: what was not finished is what OpenFailureBook reconciles. |
| 233 | // |
| 234 | // The planned wakes are counted here, not left to Shutdown (Task 11 fix |
| 235 | // round 1): Shutdown waits for each entry's e.ready, and wake() closes it |
| 236 | // before afterWake sends the terminal event, while the first-wake |
| 237 | // observer's cancellation is sent from its own goroutine. A wake already |
| 238 | // past doWake's context check when Shutdown began can also still plan and |
| 239 | // write its intent. Each of those reaches OnGen after Shutdown returns. |
| 240 | func (m *Manager) Close(ctx context.Context) { |
| 241 | m.joinMu.Lock() |
| 242 | m.closed = true |
| 243 | var wait chan struct{} |
| 244 | if m.inflight > 0 { |
| 245 | if m.joined == nil { |
| 246 | m.joined = make(chan struct{}) |
| 247 | } |
| 248 | wait = m.joined |
| 249 | } |
| 250 | n := m.inflight |
| 251 | m.joinMu.Unlock() |
| 252 | if wait != nil { |
| 253 | select { |
| 254 | case <-wait: |
| 255 | case <-ctx.Done(): |
| 256 | m.d.Logf("rollout: stopping with %d attempt(s), plan(s), wake(s) or wake event(s) still running: %v", n, ctx.Err()) |
| 257 | } |
| 258 | } |
| 259 | m.acks.close(ctx) |
| 260 | } |
| 261 | |
| 262 | // enter counts one operation Close joins. An admission (admit true) is |
| 263 | // refused once Close has begun; a Plan or OnGen call is always counted, so |
| 264 | // one that arrives while Close waits is waited for too. |
| 265 | func (m *Manager) enter(admit bool) bool { |
| 266 | m.joinMu.Lock() |
| 267 | defer m.joinMu.Unlock() |
| 268 | if admit && m.closed { |
| 269 | return false |
| 270 | } |
| 271 | m.inflight++ |
| 272 | return true |
| 273 | } |
| 274 | |
| 275 | // leave ends an operation enter counted. |
| 276 | func (m *Manager) leave() { |
| 277 | m.joinMu.Lock() |
| 278 | defer m.joinMu.Unlock() |
| 279 | m.leaveLocked() |
| 280 | } |
| 281 | |
| 282 | func (m *Manager) leaveLocked() { |
| 283 | m.inflight-- |
| 284 | if m.inflight == 0 && m.joined != nil { |
| 285 | close(m.joined) |
| 286 | m.joined = nil |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | // enterWake counts the planned wake with token tok, whose intent beginWake |
| 291 | // has just written, until leaveWake (its terminal event). Always counted, |
| 292 | // even once Close has begun: the intent is on disk, and only that event's |
| 293 | // OnGen ends it. A token already counted is not counted again. |
| 294 | func (m *Manager) enterWake(tok uint64) { |
| 295 | m.joinMu.Lock() |
| 296 | defer m.joinMu.Unlock() |
| 297 | if _, ok := m.wakes[tok]; ok { |
| 298 | return |
| 299 | } |
| 300 | if m.wakes == nil { |
| 301 | m.wakes = map[uint64]struct{}{} |
| 302 | } |
| 303 | m.wakes[tok] = struct{}{} |
| 304 | m.inflight++ |
| 305 | } |
| 306 | |
| 307 | // leaveWake gives back the count enterWake took for tok, once. Every |
| 308 | // terminal event of a planned wake ends here; one whose wake wrote no |
| 309 | // intent (a refused plan, a failed Begin, an ordinary wake), or one |
| 310 | // delivered again, finds nothing and changes nothing. |
| 311 | func (m *Manager) leaveWake(tok uint64) { |
| 312 | m.joinMu.Lock() |
| 313 | defer m.joinMu.Unlock() |
| 314 | if _, ok := m.wakes[tok]; !ok { |
| 315 | return |
| 316 | } |
| 317 | delete(m.wakes, tok) |
| 318 | m.leaveLocked() |
| 319 | } |
| 320 | |
| 321 | // closing reports whether Close has begun. |
| 322 | func (m *Manager) closing() bool { |
| 323 | m.joinMu.Lock() |
| 324 | defer m.joinMu.Unlock() |
| 325 | return m.closed |
| 326 | } |
| 327 | |
| 328 | // changed tells OnChange a transition happened. Every caller has released |
| 329 | // Manager.mu and any scheduler.mu first, so the callback may read Status. |
| 330 | func (m *Manager) changed() { |
| 331 | if m.d.OnChange != nil { |
| 332 | m.d.OnChange() |
| 333 | } |
| 334 | } |
| 335 | |
| 336 | func (m *Manager) scheduler(k appKey) *scheduler { |
| 337 | m.mu.Lock() |
| 338 | defer m.mu.Unlock() |
| 339 | s := m.apps[k] |
| 340 | if s == nil { |
| 341 | book := m.d.Book |
| 342 | s = &scheduler{m: m, key: k, kick: make(chan struct{}, 1), |
| 343 | q: newAppQueue(k.acct, k.app, func(build, host string, served bool) bool { |
| 344 | return book.Blocks(k.acct, k.app, build, host, served) |
| 345 | })} |
| 346 | m.apps[k] = s |
| 347 | go s.loop(m.ctx) |
| 348 | } |
| 349 | return s |
| 350 | } |
| 351 | |
| 352 | func (m *Manager) schedulerOf(host string) *scheduler { |
| 353 | m.mu.Lock() |
| 354 | defer m.mu.Unlock() |
| 355 | k, ok := m.hostApp[host] |
| 356 | if !ok { |
| 357 | return nil |
| 358 | } |
| 359 | return m.apps[k] |
| 360 | } |
| 361 | |
| 362 | // track is host's config track; the caller holds m.mu. |
| 363 | func (m *Manager) track(host string) *configTrack { |
| 364 | t := m.cfg[host] |
| 365 | if t == nil { |
| 366 | t = &configTrack{} |
| 367 | m.cfg[host] = t |
| 368 | } |
| 369 | return t |
| 370 | } |
| 371 | |
| 372 | // Pass is the agent pass of spec §6.1: each managed route's installed build, |
| 373 | // stamped by root, against the build its current generation serves. routes |
| 374 | // must be the box's whole list, because Pass also prunes failure records no |
| 375 | // route installs any more, and lifts the halts those records held. |
| 376 | func (m *Manager) Pass(routes []registry.Route) { |
| 377 | keys := m.observe(routes) |
| 378 | m.forget(routes) |
| 379 | m.applyPointers(routes, keys) |
| 380 | m.dropUnownedWakes() |
| 381 | m.prune(routes) |
| 382 | m.liftClearedHalts() |
| 383 | } |
| 384 | |
| 385 | // heldOnBox reports whether rt is a route this manager keeps state for: |
| 386 | // Managed, but with a passing status (draining, moving, landing) read as |
| 387 | // ready. A row passing through one of those for a Pass keeps its place in |
| 388 | // the queue, its pending restart or config, and its held ack; a row that is |
| 389 | // gone, or that stopped being an exec route this scheduler moves, does not. |
| 390 | func heldOnBox(rt registry.Route) bool { |
| 391 | return Restrictable(rt) |
| 392 | } |
| 393 | |
| 394 | // forget drops what the manager holds for hosts no longer on the box: a |
| 395 | // route removed or moved away, or one that now belongs to another app or |
| 396 | // stopped being managed. Without it such a host stayed in Status for the |
| 397 | // life of the process, possibly as queued, and could hold a deploy watch |
| 398 | // open (Task 13). A host with an attempt running is kept: the attempt's |
| 399 | // done() still needs its slot, and the next Pass after it finishes forgets |
| 400 | // it. Failure records and cohort halts are not per host and are not |
| 401 | // touched; prune and liftClearedHalts own those. |
| 402 | // |
| 403 | // Manager.mu, then each scheduler.mu: the documented order. Nothing here |
| 404 | // calls the activator. |
| 405 | func (m *Manager) forget(routes []registry.Route) { |
| 406 | present := make(map[string]appKey, len(routes)) |
| 407 | for _, rt := range routes { |
| 408 | if heldOnBox(rt) { |
| 409 | present[rt.Host] = appKey{release.RouteAccount(rt), rt.App} |
| 410 | } |
| 411 | } |
| 412 | m.mu.Lock() |
| 413 | running := map[string]bool{} |
| 414 | var gone []string |
| 415 | for k, s := range m.apps { |
| 416 | s.mu.Lock() |
| 417 | for host, h := range s.q.hosts { |
| 418 | if pk, ok := present[host]; ok && pk == k { |
| 419 | continue |
| 420 | } |
| 421 | if h.running != nil { |
| 422 | running[host] = true |
| 423 | continue |
| 424 | } |
| 425 | delete(s.q.hosts, host) |
| 426 | gone = append(gone, host) |
| 427 | } |
| 428 | s.mu.Unlock() |
| 429 | } |
| 430 | for _, host := range gone { |
| 431 | if _, ok := present[host]; ok || running[host] { |
| 432 | continue // moved to another app on this box, or still running there |
| 433 | } |
| 434 | delete(m.hostApp, host) |
| 435 | delete(m.cfg, host) |
| 436 | delete(m.retried, host) |
| 437 | delete(m.observing, host) |
| 438 | // A later event of the host's last wake then matches no token: the |
| 439 | // host is gone, and there is nothing of its wake left to act on. |
| 440 | delete(m.wakeTok, host) |
| 441 | delete(m.wakeAt, host) |
| 442 | } |
| 443 | m.mu.Unlock() |
| 444 | if len(gone) > 0 { |
| 445 | m.d.Logf("rollout: forgot %d host(s) no longer on this box: %v", len(gone), gone) |
| 446 | m.changed() // the hosts left the report |
| 447 | } |
| 448 | } |
| 449 | |
| 450 | // Nudge re-reads one host and wakes its scheduler. The activator sweep's |
| 451 | // drift arm calls it (spec §10, OnDrift) on a goroutine of its own, for every |
| 452 | // drifted entry on every sweep, with no dedupe of its own. So nudges for one |
| 453 | // host collapse here: while one is reading and observing, the others only |
| 454 | // ask it for one more look and return, and a slow read (the row, the |
| 455 | // policy) never collects a goroutine per sweep. A nudge for the build the |
| 456 | // host already wants changes nothing (wantBuild), so a queued or running |
| 457 | // attempt is neither queued again nor superseded. |
| 458 | func (m *Manager) Nudge(host string) { |
| 459 | m.mu.Lock() |
| 460 | if _, busy := m.nudging[host]; busy { |
| 461 | m.nudging[host] = true |
| 462 | m.mu.Unlock() |
| 463 | return |
| 464 | } |
| 465 | m.nudging[host] = false |
| 466 | m.mu.Unlock() |
| 467 | for { |
| 468 | if rt, err := m.d.Routes.Get(host); err == nil && rt != nil { |
| 469 | m.observe([]registry.Route{*rt}) |
| 470 | } |
| 471 | m.mu.Lock() |
| 472 | again := m.nudging[host] |
| 473 | if !again { |
| 474 | delete(m.nudging, host) |
| 475 | m.mu.Unlock() |
| 476 | return |
| 477 | } |
| 478 | m.nudging[host] = false |
| 479 | m.mu.Unlock() |
| 480 | } |
| 481 | } |
| 482 | |
| 483 | // Adopted replaces BounceAdopted as the adopt pass's restart callback when |
| 484 | // the flag is on: the pass has repointed the live symlink, and the scheduler |
| 485 | // takes the move from here, paced. It reports whether an awake instance will |
| 486 | // be attempted, so makeRestartRoute does not fall through to systemctl. |
| 487 | func (m *Manager) Adopted(host string) bool { |
| 488 | if _, awake := m.d.Act.Current(host); !awake { |
| 489 | return false |
| 490 | } |
| 491 | m.Nudge(host) |
| 492 | return true |
| 493 | } |
| 494 | |
| 495 | func (m *Manager) observe(routes []registry.Route) []appKey { |
| 496 | type seen struct { |
| 497 | rt registry.Route |
| 498 | serving string |
| 499 | known bool // false while waking or stopping: the slot is left alone |
| 500 | } |
| 501 | byApp := map[appKey][]seen{} |
| 502 | for _, rt := range routes { |
| 503 | if !Managed(rt) { |
| 504 | continue |
| 505 | } |
| 506 | s := seen{rt: rt, known: true} |
| 507 | if g, ok := m.d.Act.Current(rt.Host); ok && g.Build != "" { |
| 508 | s.serving = g.Build |
| 509 | } else if m.d.Act.Owns(rt.Host) { |
| 510 | s.known = false |
| 511 | } |
| 512 | k := appKey{release.RouteAccount(rt), rt.App} |
| 513 | byApp[k] = append(byApp[k], s) |
| 514 | } |
| 515 | now := m.d.Now() |
| 516 | keys := make([]appKey, 0, len(byApp)) |
| 517 | moved := false |
| 518 | for k, list := range byApp { |
| 519 | keys = append(keys, k) |
| 520 | policy := m.policies.get(m.ctx, k.acct, k.app) |
| 521 | s := m.scheduler(k) |
| 522 | retried := map[string]string{} |
| 523 | m.mu.Lock() |
| 524 | for _, x := range list { |
| 525 | m.hostApp[x.rt.Host] = k |
| 526 | if b, ok := m.retried[x.rt.Host]; ok { |
| 527 | retried[x.rt.Host] = b |
| 528 | } |
| 529 | } |
| 530 | m.mu.Unlock() |
| 531 | rearm := false |
| 532 | s.mu.Lock() |
| 533 | s.q.concurrency = policy.Concurrency |
| 534 | for _, x := range list { |
| 535 | if x.known { |
| 536 | // Serving state only from a stable entry. The desired build |
| 537 | // and its revision are updated either way (spec §6.3): a |
| 538 | // build root installs while a hold drains or a wake runs |
| 539 | // supersedes the attempt's, so its gates (pinCheck, the |
| 540 | // result check) refuse the older build. Admission still |
| 541 | // waits for a stable entry (admit's view leaves a waking or |
| 542 | // stopping host out). |
| 543 | s.q.observe(x.rt.Host, x.serving, x.rt.GoodVersion, now) |
| 544 | } |
| 545 | cause := activator.CauseRelease |
| 546 | if retried[x.rt.Host] == x.rt.Version { |
| 547 | cause = activator.CauseRetry |
| 548 | } |
| 549 | if s.q.wantBuild(x.rt.Host, cause, x.rt.Version, now) { |
| 550 | moved = true |
| 551 | // A marked host's installed build moved: it may be a rollback |
| 552 | // onto its good build, so the app's pointers are read at the |
| 553 | // next applyPointers rather than up to pointerReadEvery later |
| 554 | // (ruling 205). The book's lock is the last in the order. |
| 555 | if _, ok := holdRanFor(m.d.Book, k.acct, k.app, x.rt.Host); ok { |
| 556 | rearm = true |
| 557 | } |
| 558 | } |
| 559 | } |
| 560 | s.mu.Unlock() |
| 561 | if rearm { |
| 562 | m.pointerReadDue(k) |
| 563 | } |
| 564 | s.poke() |
| 565 | } |
| 566 | if moved { |
| 567 | // A sleeper has no attempt to finish, so the pass that first sees |
| 568 | // its new build is the news for the deploy watch. |
| 569 | m.changed() |
| 570 | } |
| 571 | return keys |
| 572 | } |
| 573 | |
| 574 | // pointerReadDue makes the next applyPointers read k's pointers at once |
| 575 | // rather than up to pointerReadEvery later (ruling 205). observe calls it |
| 576 | // when a marked host's installed build moves; done() and OnGen call it when |
| 577 | // they set a marker, because the rollback that superseded the hold build is |
| 578 | // signed and installed while that build runs, before any marker exists, so |
| 579 | // the pass that saw the install could not make the read due. The caller |
| 580 | // holds no lock (Manager.mu is taken here). |
| 581 | func (m *Manager) pointerReadDue(k appKey) { |
| 582 | m.mu.Lock() |
| 583 | delete(m.pointerRead, k) |
| 584 | m.mu.Unlock() |
| 585 | } |
| 586 | |
| 587 | // routeChannel is the channel a route follows: "" means stable, as |
| 588 | // release's own (unexported) routeChannel reads it. |
| 589 | func routeChannel(rt registry.Route) string { |
| 590 | if rt.Channel == "" { |
| 591 | return "stable" |
| 592 | } |
| 593 | return rt.Channel |
| 594 | } |
| 595 | |
| 596 | // applyPointers is the retry signal of spec rev 14 §8.4. For each app with a |
| 597 | // standing failure record, at most every pointerReadEvery, it reads the |
| 598 | // signed pointer of every channel the app's managed routes on this box |
| 599 | // follow, and clears a record when a pointer names its build with a newer |
| 600 | // PromotedAt (`carlos promote --retry`). The pointer the build failed under, |
| 601 | // read again, clears nothing. A second retry of the same build is the same |
| 602 | // rule again: the retry's own failure recorded the newer PromotedAt, so only |
| 603 | // a pointer signed after it clears the record. |
| 604 | // |
| 605 | // It is also the rollback signal for hold-ran markers (ruling 205, Astra |
| 606 | // R3). An app with a marked host whose route installs its good build is read |
| 607 | // too, on the same schedule, and that host's marker is cleared when its |
| 608 | // channel's pointer names the good build with a PromotedAt newer than the |
| 609 | // marker: `carlos rollback` re-signs the pointer, root's automatic sibling |
| 610 | // repoint does not, so only an explicit rollback lets the good build boot |
| 611 | // over the marker. The signature is the pointer read's to verify |
| 612 | // (release.Channel). The marker's At is the start of the execution that set |
| 613 | // it (an attempt's dispatch, a planned wake's plan), so the rollback that |
| 614 | // superseded a running hold build is newer than it. The read is made due at |
| 615 | // once when a marked host's installed build moves and when a marker is set |
| 616 | // (pointerReadDue), so a rollback is not left waiting out pointerReadEvery. |
| 617 | func (m *Manager) applyPointers(routes []registry.Route, keys []appKey) { |
| 618 | if m.d.Pointer == nil { |
| 619 | return |
| 620 | } |
| 621 | now := m.d.Now() |
| 622 | for _, k := range keys { |
| 623 | w := pointerWorkFor(m.d.Book, routes, k) |
| 624 | if w.empty() { |
| 625 | continue |
| 626 | } |
| 627 | m.mu.Lock() |
| 628 | due := now.Sub(m.pointerRead[k]) >= pointerReadEvery |
| 629 | if due { |
| 630 | m.pointerRead[k] = now |
| 631 | } |
| 632 | m.mu.Unlock() |
| 633 | if !due { |
| 634 | continue |
| 635 | } |
| 636 | // Each cleared record's build is requeued as a retry at once, before |
| 637 | // the app's next channel is read (Task 10 review minor 1). |
| 638 | _, unmarked := w.apply(m.ctx, m.d.Book, m.d.Pointer, routes, k, m.d.Logf, func(build string) { m.retryBuild(k, build, "") }) |
| 639 | if unmarked { |
| 640 | m.changed() |
| 641 | } |
| 642 | } |
| 643 | } |
| 644 | |
| 645 | // pointerWork is what one app's channel pointers could clear on this box |
| 646 | // (applyPointers): its failure records, and its marked hosts whose route |
| 647 | // installs the good build, by channel. An app with neither costs no read. |
| 648 | type pointerWork struct { |
| 649 | recs []FailureRecord |
| 650 | rollback map[string][]registry.Route |
| 651 | } |
| 652 | |
| 653 | func (w pointerWork) empty() bool { return len(w.recs) == 0 && len(w.rollback) == 0 } |
| 654 | |
| 655 | // pointerWorkFor is k's pointerWork against book and the box's routes. |
| 656 | func pointerWorkFor(book *FailureBook, routes []registry.Route, k appKey) pointerWork { |
| 657 | w := pointerWork{recs: book.ForApp(k.acct, k.app), rollback: map[string][]registry.Route{}} |
| 658 | for _, rt := range routes { |
| 659 | if !Managed(rt) || release.RouteAccount(rt) != k.acct || rt.App != k.app || rt.GoodVersion == "" || rt.Version != rt.GoodVersion { |
| 660 | continue |
| 661 | } |
| 662 | if _, ok := holdRanFor(book, k.acct, k.app, rt.Host); ok { |
| 663 | ch := routeChannel(rt) |
| 664 | w.rollback[ch] = append(w.rollback[ch], rt) |
| 665 | } |
| 666 | } |
| 667 | return w |
| 668 | } |
| 669 | |
| 670 | // apply reads the signed pointer of every channel k's managed routes follow |
| 671 | // and applies applyPointers' two rules to the book: a record whose build a |
| 672 | // pointer names with a newer PromotedAt is cleared (ClearOnPointer), and so |
| 673 | // is the marker of a host whose channel's pointer names its installed good |
| 674 | // build with a PromotedAt after the marker's At (ClearHoldRanOnPointer). A |
| 675 | // read error is logged and that channel skipped; each read is bounded by |
| 676 | // readBound (Task 11 rule 2). onCleared is called with each cleared record's |
| 677 | // build as soon as the book has cleared it, before the next record or |
| 678 | // channel, with no lock held: the armed pass requeues the build there, so |
| 679 | // the retry is not left waiting on the rest of the app's reads (Task 10 |
| 680 | // review minor 1). It reports how many records it cleared and whether it |
| 681 | // cleared a marker; what follows from a marker is the caller's. |
| 682 | func (w pointerWork) apply(ctx context.Context, book *FailureBook, read PointerRead, routes []registry.Route, k appKey, |
| 683 | logf func(format string, args ...any), onCleared func(build string)) (cleared int, unmarked bool) { |
| 684 | channels := map[string]bool{} |
| 685 | for _, rt := range routes { |
| 686 | if Managed(rt) && release.RouteAccount(rt) == k.acct && rt.App == k.app { |
| 687 | channels[routeChannel(rt)] = true |
| 688 | } |
| 689 | } |
| 690 | for ch := range channels { |
| 691 | rctx, cancel := context.WithTimeout(ctx, readBound) |
| 692 | ptr, err := read(rctx, k.acct, k.app, ch) |
| 693 | cancel() |
| 694 | if err != nil { |
| 695 | logf("rollout: %s/%s channel %s: %v (retry not checked this pass)", k.acct, k.app, ch, err) |
| 696 | continue |
| 697 | } |
| 698 | if ptr == nil { |
| 699 | continue |
| 700 | } |
| 701 | for _, r := range w.recs { |
| 702 | if ptr.Version != r.Build { |
| 703 | continue |
| 704 | } |
| 705 | ok, err := book.ClearOnPointer(k.acct, k.app, r.Build, ptr.PromotedAt) |
| 706 | if err != nil { |
| 707 | logf("rollout: %s/%s clear record for %s: %v (cleared in memory)", k.acct, k.app, r.Build, err) |
| 708 | } |
| 709 | if ok { |
| 710 | cleared++ |
| 711 | onCleared(r.Build) |
| 712 | } |
| 713 | } |
| 714 | for _, rt := range w.rollback[ch] { |
| 715 | if ptr.Version != rt.GoodVersion { |
| 716 | continue |
| 717 | } |
| 718 | ok, err := book.ClearHoldRanOnPointer(rt.Host, k.acct, k.app, ptr.PromotedAt) |
| 719 | if err != nil { |
| 720 | logf("rollout: %s: clear hold-ran marker on a rollback to %s: %v (cleared in memory)", rt.Host, rt.GoodVersion, err) |
| 721 | } |
| 722 | if ok { |
| 723 | logf("rollout: %s: a signed rollback to %s (pointer of %s) cleared its hold-ran marker", rt.Host, rt.GoodVersion, ptr.PromotedAt.UTC().Format(time.RFC3339)) |
| 724 | unmarked = true |
| 725 | } |
| 726 | } |
| 727 | } |
| 728 | return cleared, unmarked |
| 729 | } |
| 730 | |
| 731 | // PointerRead reads a channel's signed pointer, verified (release.Channel |
| 732 | // with the box's store and root key): Deps.Pointer's shape. |
| 733 | type PointerRead = func(ctx context.Context, acct, app, channel string) (*release.ChannelPointer, error) |
| 734 | |
| 735 | // PointerReadEvery is pointerReadEvery, for the pointer pass of a box with |
| 736 | // no Manager (ApplyPointers), which paces itself. |
| 737 | const PointerReadEvery = pointerReadEvery |
| 738 | |
| 739 | // ApplyPointers is applyPointers for a box with no Manager: one that |
| 740 | // declined CARLOS_SWAP=overlap but honours the records and hold-ran markers |
| 741 | // an armed run left (internal/host's failureHolds; Astra R2 ruling 194 |
| 742 | // "startup-only records", Task 10). Every app with a managed route in routes |
| 743 | // and something its pointers could clear is read now, under the same two |
| 744 | // rules on the same book; the caller paces the pass, at most every |
| 745 | // PointerReadEvery. There is no scheduler to requeue a cleared build, so a |
| 746 | // cleared record is only logged. It reports whether anything was cleared. |
| 747 | func ApplyPointers(ctx context.Context, book *FailureBook, read PointerRead, routes []registry.Route, logf func(format string, args ...any)) bool { |
| 748 | changed := false |
| 749 | seen := map[appKey]bool{} |
| 750 | for _, rt := range routes { |
| 751 | k := appKey{release.RouteAccount(rt), rt.App} |
| 752 | if !Managed(rt) || seen[k] { |
| 753 | continue |
| 754 | } |
| 755 | seen[k] = true |
| 756 | w := pointerWorkFor(book, routes, k) |
| 757 | if w.empty() { |
| 758 | continue |
| 759 | } |
| 760 | cleared, unmarked := w.apply(ctx, book, read, routes, k, logf, func(build string) { |
| 761 | logf("rollout: %s/%s: a signed pointer re-promoted %s, so its failure record is cleared", k.acct, k.app, build) |
| 762 | }) |
| 763 | changed = changed || cleared > 0 || unmarked |
| 764 | } |
| 765 | return changed |
| 766 | } |
| 767 | |
| 768 | // retryBuild follows a cleared record: the next attempt onto build carries |
| 769 | // cause retry and runs as a fresh canary. host == "" is an app-wide retry, |
| 770 | // which reopens the halted cohort: its next attempt is a canary whatever |
| 771 | // wantBuild reports, because unhalt starts the cohort again. A host names |
| 772 | // `carlos instances retry`, whose attempt is that host's own canary; |
| 773 | // RetryHost has already bound the host's retry flag, in the step that |
| 774 | // cleared the book, and the flag is not set again here, where a forced |
| 775 | // attempt may already have started and spent it. |
| 776 | func (m *Manager) retryBuild(k appKey, build, host string) { |
| 777 | s := m.scheduler(k) |
| 778 | now := m.d.Now() |
| 779 | s.mu.Lock() |
| 780 | var hosts []string |
| 781 | for name, h := range s.q.hosts { |
| 782 | if h.wantBuild == build && (host == "" || name == host) { |
| 783 | hosts = append(hosts, name) |
| 784 | s.q.wantBuild(name, activator.CauseRetry, build, now) |
| 785 | } |
| 786 | } |
| 787 | if host == "" { |
| 788 | s.q.unhalt(build) |
| 789 | } |
| 790 | s.mu.Unlock() |
| 791 | m.mu.Lock() |
| 792 | for _, name := range hosts { |
| 793 | m.retried[name] = build |
| 794 | } |
| 795 | m.mu.Unlock() |
| 796 | m.d.Logf("rollout: %s/%s build %s retried (%v)", k.acct, k.app, build, hosts) |
| 797 | s.poke() |
| 798 | m.changed() |
| 799 | } |
| 800 | |
| 801 | // RetryHost is the box half of `carlos instances retry <host>` (spec §8.4, |
| 802 | // §11): a grant the console signed (release.HostRetryGrant), which the |
| 803 | // caller has verified, naming the build whose restriction of host it lifts |
| 804 | // and that restriction's epoch and book as the box last reported them |
| 805 | // (plan review findings 5 and 13; final review round 1, finding 3). The |
| 806 | // book clears build's record for host, when one holds it, and host's |
| 807 | // hold-ran marker, when the marker names build, and only while the app's |
| 808 | // restriction generation is still epoch, the book owes no save, and the |
| 809 | // grant names this box's book (RetryHostAt; the crash-epoch hard limit in |
| 810 | // FailureBook's doc): a grant from before any later change of the app's |
| 811 | // restrictions, one replayed after it was spent, or one for another book's |
| 812 | // restriction at the same value, clears nothing. True means either was |
| 813 | // cleared. |
| 814 | // |
| 815 | // When the record was cleared and host's slot wants build, its next attempt |
| 816 | // is a canary. The book entry is cleared and the host's retry flag bound in |
| 817 | // one hold of the scheduler's lock (scheduler.mu before FailureBook.mu), the |
| 818 | // book first. Admission cannot run between the two, so the adoption onto |
| 819 | // the cleared build is the attempt the flag lets through, as a canary, and |
| 820 | // a restart or config attempt cannot start in between and spend the flag. |
| 821 | // The book also notes the retry's place in its sequence there (ruling 191): |
| 822 | // an outcome of an execution that began before it is stale, and takes the |
| 823 | // retry back no more than it halts the cohort (failCohort). No write is |
| 824 | // added under the scheduler's lock: the book saves as it did before. |
| 825 | // |
| 826 | // A grant naming another build than the slot's (the marker's build B while |
| 827 | // the slot wants A) clears B's restriction and leaves A's attempt alone; a |
| 828 | // cleared record of B still binds the host's canary entitlement for B, |
| 829 | // which the adoption onto B spends when the slot comes to want it, through |
| 830 | // B's halt if one stands (final review round 1, finding 5). |
| 831 | func (m *Manager) RetryHost(host, build, book string, epoch uint64) bool { |
| 832 | m.mu.Lock() |
| 833 | k, ok := m.hostApp[host] |
| 834 | m.mu.Unlock() |
| 835 | if !ok || build == "" { |
| 836 | return false |
| 837 | } |
| 838 | s := m.scheduler(k) |
| 839 | s.mu.Lock() |
| 840 | want := "" |
| 841 | if h := s.q.hosts[host]; h != nil { |
| 842 | want = h.wantBuild |
| 843 | } |
| 844 | cleared, unmarked, cur, err := m.d.Book.RetryHostAt(k.acct, k.app, build, host, book, epoch) |
| 845 | if cleared { |
| 846 | // The canary entitlement is build's whatever the slot wants now |
| 847 | // (final review round 1, finding 5): bound here, in the step that |
| 848 | // cleared the book, and spent only by the adoption onto build. |
| 849 | s.q.retryHost(host, build) |
| 850 | } |
| 851 | requeue := cleared && want == build |
| 852 | if requeue { |
| 853 | s.q.wantBuild(host, activator.CauseRetry, build, m.d.Now()) |
| 854 | } |
| 855 | s.mu.Unlock() |
| 856 | if err != nil && (cleared || unmarked) { |
| 857 | m.d.Logf("rollout: retry %s on %s: %v (cleared in memory)", build, host, err) |
| 858 | } else if err != nil { |
| 859 | // The book owed a save and it failed again: nothing is cleared until |
| 860 | // one lands (the crash-epoch hard limit, rule 2). |
| 861 | m.d.Logf("rollout: retry %s on %s: the failure book cannot be saved (%v), so nothing was cleared", build, host, err) |
| 862 | } else if !cleared && !unmarked && cur != epoch { |
| 863 | m.d.Logf("rollout: retry %s on %s: the grant is for epoch %d, the restriction's is %d, so nothing was cleared", build, host, epoch, cur) |
| 864 | } else if !cleared && !unmarked && book != m.d.Book.ID() { |
| 865 | m.d.Logf("rollout: retry %s on %s: the grant is for book %q, not this box's, so nothing was cleared", build, host, book) |
| 866 | } |
| 867 | if requeue { |
| 868 | m.retryBuild(k, build, host) |
| 869 | } else if cleared || unmarked { |
| 870 | m.changed() |
| 871 | } |
| 872 | return cleared || unmarked |
| 873 | } |
| 874 | |
| 875 | // failedUnder is the PromotedAt a failure record keeps (spec rev 14 §8.4): |
| 876 | // the pointer that named the build when it failed. If that pointer cannot be |
| 877 | // read, or already names another build, it is zero, and the failure's own |
| 878 | // time stands in when the record is written: only a pointer signed after the |
| 879 | // failure can then clear the record, which is the safe direction. |
| 880 | // |
| 881 | // ctx is the caller's: the wake's in planRow, the attempt's in execute |
| 882 | // (Task 11 rule 2, Astra R1 Flag triage T10). A cancelled wake no longer |
| 883 | // waits on the store for its pointer, and no read waits past readBound. |
| 884 | func (m *Manager) failedUnder(ctx context.Context, rt registry.Route, build string) time.Time { |
| 885 | if m.d.Pointer != nil { |
| 886 | ctx, cancel := context.WithTimeout(ctx, readBound) |
| 887 | defer cancel() |
| 888 | ptr, err := m.d.Pointer(ctx, release.RouteAccount(rt), rt.App, routeChannel(rt)) |
| 889 | if err == nil && ptr != nil && ptr.Version == build { |
| 890 | return ptr.PromotedAt |
| 891 | } |
| 892 | } |
| 893 | return time.Time{} |
| 894 | } |
| 895 | |
| 896 | // plannedWake is a planned first wake in flight (Manager.planned). token is |
| 897 | // its plan's (WakePlan.Token): only an event carrying it ends the entry. |
| 898 | type plannedWake struct { |
| 899 | acct, app, build string |
| 900 | seq uint64 |
| 901 | token uint64 |
| 902 | } |
| 903 | |
| 904 | // newWakeToken starts a plan for host: a fresh token, recorded as the |
| 905 | // host's newest. It is taken before the plan registers anything (Plan), so |
| 906 | // the plan's registration and every event of its wake carry it. |
| 907 | // It also returns the plan's time, the start of the wake's execution |
| 908 | // (noteWakeStart). |
| 909 | func (m *Manager) newWakeToken(host string) (uint64, time.Time) { |
| 910 | now := m.d.Now() |
| 911 | m.mu.Lock() |
| 912 | defer m.mu.Unlock() |
| 913 | m.tokenCount++ |
| 914 | m.wakeTok[host] = m.tokenCount |
| 915 | return m.tokenCount, now |
| 916 | } |
| 917 | |
| 918 | // newIncarnation draws a manager's incarnation (Manager.inc). crypto/rand |
| 919 | // does not fail on the platforms the agent runs on; if it ever did, the |
| 920 | // clock's nanoseconds still tell one process start from the next. |
| 921 | func newIncarnation() string { |
| 922 | var b [8]byte |
| 923 | if _, err := rand.Read(b[:]); err != nil { |
| 924 | return strconv.FormatInt(time.Now().UnixNano(), 16) |
| 925 | } |
| 926 | return hex.EncodeToString(b[:]) |
| 927 | } |
| 928 | |
| 929 | // nextExecID is a scheduled attempt's execution identity (dispatch.ExecID): |
| 930 | // "a<incarnation>-<n>", n from this manager's counter. The "a" is what |
| 931 | // Begin's wake rule reads (wakeIntent). |
| 932 | func (m *Manager) nextExecID() string { |
| 933 | m.mu.Lock() |
| 934 | defer m.mu.Unlock() |
| 935 | m.execCount++ |
| 936 | return "a" + m.inc + "-" + strconv.FormatUint(m.execCount, 10) |
| 937 | } |
| 938 | |
| 939 | // wakeExecID is a planned wake's execution identity (HoldRan.Exec): |
| 940 | // "w<incarnation>-<token>". Zero, an event of no known plan, has none. The |
| 941 | // incarnation is fixed for the manager's life, so every event of one wake |
| 942 | // names the identity its plan's intent was written with. |
| 943 | func (m *Manager) wakeExecID(tok uint64) string { |
| 944 | if tok == 0 { |
| 945 | return "" |
| 946 | } |
| 947 | return "w" + m.inc + "-" + strconv.FormatUint(tok, 10) |
| 948 | } |
| 949 | |
| 950 | // wakeStart is a planned wake's start, by its token (Manager.wakeAt): its |
| 951 | // plan time (the plan's start) and the failure book's snapshot taken at the |
| 952 | // plan's end, after the planner's book reads (Plan). |
| 953 | type wakeStart struct { |
| 954 | token uint64 |
| 955 | at time.Time |
| 956 | snap uint64 |
| 957 | } |
| 958 | |
| 959 | // wakeStartsKept bounds Manager.wakeAt per host. Every planned wake ends |
| 960 | // with a terminal event (Task 5), which drops its entry; a plan the |
| 961 | // activator never ran (one Plan returned to a caller that then gave up) |
| 962 | // leaves its entry until this many newer plans push it out; a live wake is |
| 963 | // never that far behind its host's newest plan, because the activator waits |
| 964 | // for a planned wake's outcome before it plans again (outcomeFenceWait). |
| 965 | const wakeStartsKept = 4 |
| 966 | |
| 967 | // noteWakeStart keeps the start of a planned wake that boots a build (tok): |
| 968 | // a first wake's end can leave a hold-ran marker or a failure (OnGen), and |
| 969 | // any wake's healthy event writes the good build, which clears only markers |
| 970 | // older than its snapshot (Astra R3 N6). Any mode is kept: the activator |
| 971 | // reports the mode the wake ran in, which the plan does not fix. |
| 972 | func (m *Manager) noteWakeStart(host string, tok uint64, at time.Time, snap uint64, plan activator.WakePlan) { |
| 973 | if tok == 0 || plan.Refuse != "" || plan.Boot.StagedPath == "" { |
| 974 | return |
| 975 | } |
| 976 | m.mu.Lock() |
| 977 | defer m.mu.Unlock() |
| 978 | starts := append(m.wakeAt[host], wakeStart{token: tok, at: at, snap: snap}) |
| 979 | if len(starts) > wakeStartsKept { |
| 980 | starts = slices.Delete(starts, 0, len(starts)-wakeStartsKept) |
| 981 | } |
| 982 | m.wakeAt[host] = starts |
| 983 | } |
| 984 | |
| 985 | // wakeStarted is the start of the wake whose token is tok, and drops its |
| 986 | // entry when ended (the wake's last event). A zero result means the start |
| 987 | // is not known (a token pushed out by newer plans, or a host forgotten): the |
| 988 | // book then stamps a marker with its own clock, later than the real start, |
| 989 | // so only a pointer signed later still lifts it; the good write clears no |
| 990 | // marker; and a failure is never taken as stale. Each is the safe |
| 991 | // direction. |
| 992 | func (m *Manager) wakeStarted(host string, tok uint64, ended bool) wakeStart { |
| 993 | m.mu.Lock() |
| 994 | defer m.mu.Unlock() |
| 995 | if tok == 0 { |
| 996 | return wakeStart{} |
| 997 | } |
| 998 | starts := m.wakeAt[host] |
| 999 | i := slices.IndexFunc(starts, func(w wakeStart) bool { return w.token == tok }) |
| 1000 | if i < 0 { |
| 1001 | return wakeStart{} |
| 1002 | } |
| 1003 | w := starts[i] |
| 1004 | if ended { |
| 1005 | if starts = slices.Delete(starts, i, i+1); len(starts) == 0 { |
| 1006 | delete(m.wakeAt, host) |
| 1007 | } else { |
| 1008 | m.wakeAt[host] = starts |
| 1009 | } |
| 1010 | } |
| 1011 | return w |
| 1012 | } |
| 1013 | |
| 1014 | // notePlanned records what host's newest plan boots: an entry for a first |
| 1015 | // wake, none for anything else. Either way it replaces the host's previous |
| 1016 | // plan, whose wake is over or is being planned again. A plan that is no |
| 1017 | // longer the host's newest (tok) changes nothing: the newer plan owns the |
| 1018 | // entry. |
| 1019 | func (m *Manager) notePlanned(rt registry.Route, plan activator.WakePlan, tok uint64) { |
| 1020 | m.mu.Lock() |
| 1021 | defer m.mu.Unlock() |
| 1022 | if m.wakeTok[rt.Host] != tok { |
| 1023 | return |
| 1024 | } |
| 1025 | if !plan.FirstWake || plan.Boot.Version == "" || plan.Refuse != "" { |
| 1026 | delete(m.planned, rt.Host) |
| 1027 | return |
| 1028 | } |
| 1029 | m.plannedSeq++ |
| 1030 | m.planned[rt.Host] = plannedWake{acct: release.RouteAccount(rt), app: rt.App, build: plan.Boot.Version, |
| 1031 | seq: m.plannedSeq, token: tok} |
| 1032 | } |
| 1033 | |
| 1034 | // notePlanning registers rt's installed build as host's in-flight planned |
| 1035 | // build while its plan (tok) is still being made (Plan, planRow), so a Pass |
| 1036 | // during the planner's slow reads keeps the build's failure record. A |
| 1037 | // plan's end (notePlanned) replaces or clears the entry, and a newer seq |
| 1038 | // keeps a Pass that snapshotted the old one (dropUnownedWakes) from |
| 1039 | // dropping it. |
| 1040 | func (m *Manager) notePlanning(rt registry.Route, tok uint64) { |
| 1041 | m.mu.Lock() |
| 1042 | defer m.mu.Unlock() |
| 1043 | if m.wakeTok[rt.Host] != tok { |
| 1044 | return |
| 1045 | } |
| 1046 | if rt.Version == "" { |
| 1047 | delete(m.planned, rt.Host) |
| 1048 | return |
| 1049 | } |
| 1050 | m.plannedSeq++ |
| 1051 | m.planned[rt.Host] = plannedWake{acct: release.RouteAccount(rt), app: rt.App, build: rt.Version, |
| 1052 | seq: m.plannedSeq, token: tok} |
| 1053 | } |
| 1054 | |
| 1055 | // dropUnownedWakes forgets the planned wakes of hosts the activator no |
| 1056 | // longer owns: a planned wake of an older activator, or one whose terminal |
| 1057 | // event is still on its way, leaves its entry behind, and only the entry's |
| 1058 | // own end may let the record go. Owns is asked with no lock held (the lock |
| 1059 | // order on Manager), and an entry is dropped only if no newer plan replaced |
| 1060 | // it meanwhile. |
| 1061 | // |
| 1062 | // It never ends a wake's intent (plan review #1): a failed wake's entry goes |
| 1063 | // before its GenFailed is delivered (activator's wake, then afterWake), so |
| 1064 | // Owns false is no proof the outcome is written, and ending the intent there |
| 1065 | // could lose a failure across a failed save and a restart. The intent ends |
| 1066 | // on the wake's own terminal event (OnGen); every planned wake sends one. |
| 1067 | func (m *Manager) dropUnownedWakes() { |
| 1068 | m.mu.Lock() |
| 1069 | seen := make(map[string]uint64, len(m.planned)) |
| 1070 | for host, w := range m.planned { |
| 1071 | seen[host] = w.seq |
| 1072 | } |
| 1073 | m.mu.Unlock() |
| 1074 | gone := map[string]uint64{} |
| 1075 | for host, seq := range seen { |
| 1076 | if !m.d.Act.Owns(host) { |
| 1077 | gone[host] = seq |
| 1078 | } |
| 1079 | } |
| 1080 | if len(gone) == 0 { |
| 1081 | return |
| 1082 | } |
| 1083 | m.mu.Lock() |
| 1084 | for host, seq := range gone { |
| 1085 | if w, ok := m.planned[host]; ok && w.seq == seq { |
| 1086 | delete(m.planned, host) |
| 1087 | } |
| 1088 | } |
| 1089 | m.mu.Unlock() |
| 1090 | } |
| 1091 | |
| 1092 | // prune drops failure records for builds no route installs. A held-down |
| 1093 | // instance's route still names its failed build, so its record stays (spec |
| 1094 | // §8.4). Every route counts, whatever its status or shape: a row passing |
| 1095 | // through "draining", "moving" or "landing" for one Pass, or one the |
| 1096 | // scheduler does not manage, still names its build, and dropping the record |
| 1097 | // then would lift the halt and let a later wake boot the failed hold build |
| 1098 | // against its database. Pruning errs towards keeping. An empty route list is |
| 1099 | // never authority to drop anything: registry.Open creates an empty database |
| 1100 | // on a wrong path. A build a planned first wake still carries counts as |
| 1101 | // installed (Manager.planned): the wake was planned before the install moved |
| 1102 | // off it, and its exec may yet boot it. The installed callback runs under the |
| 1103 | // book's lock and reads only this map, built first. |
| 1104 | // |
| 1105 | // The pass also retries a save the book owes (the crash-epoch hard limit, |
| 1106 | // rule 6; SaveIfDirty): while one is owed status publishes no generation |
| 1107 | // and no retry grant clears anything, so a box whose disk recovered must |
| 1108 | // not wait for another change, pointer or grant to save. Prune retries it |
| 1109 | // when it runs; a pass with no routes still calls SaveIfDirty, and so does |
| 1110 | // one whose Prune landed, for a save owed since. |
| 1111 | func (m *Manager) prune(routes []registry.Route) { |
| 1112 | if err := m.pruneBook(routes); err != nil { |
| 1113 | m.d.Logf("rollout: prune failure records: %v", err) |
| 1114 | } else if err := m.d.Book.SaveIfDirty(); err != nil { |
| 1115 | m.d.Logf("rollout: save the failure book: %v", err) |
| 1116 | } |
| 1117 | } |
| 1118 | |
| 1119 | // pruneBook is prune's Prune call, nil when there are no routes to judge by. |
| 1120 | func (m *Manager) pruneBook(routes []registry.Route) error { |
| 1121 | if len(routes) == 0 { |
| 1122 | return nil |
| 1123 | } |
| 1124 | installed := map[[3]string]bool{} |
| 1125 | for _, rt := range routes { |
| 1126 | if rt.Version != "" { |
| 1127 | installed[[3]string{release.RouteAccount(rt), rt.App, rt.Version}] = true |
| 1128 | } |
| 1129 | } |
| 1130 | m.mu.Lock() |
| 1131 | for _, w := range m.planned { |
| 1132 | installed[[3]string{w.acct, w.app, w.build}] = true |
| 1133 | } |
| 1134 | m.mu.Unlock() |
| 1135 | return m.d.Book.Prune(func(acct, app, build string) bool { return installed[[3]string{acct, app, build}] }) |
| 1136 | } |
| 1137 | |
| 1138 | // liftClearedHalts reopens every halted adoption cohort whose failure record |
| 1139 | // no longer stands (spec §8.2, §8.4: a newer build, a retry or an install by |
| 1140 | // hand lifts the hold). A halt lives in memory beside its record, so a record |
| 1141 | // pruned after an install by hand would otherwise leave the build refused for |
| 1142 | // good if it were installed again. failCohort writes the record and the halt |
| 1143 | // under the same scheduler lock this reads them under. In the same hold it |
| 1144 | // drops every retry entitlement the book no longer lists (final review round |
| 1145 | // 2, finding 5): RetryHost binds one in the step that lists the host, so an |
| 1146 | // entitlement and its listing go together. |
| 1147 | func (m *Manager) liftClearedHalts() { |
| 1148 | m.mu.Lock() |
| 1149 | scheds := make([]*scheduler, 0, len(m.apps)) |
| 1150 | for _, s := range m.apps { |
| 1151 | scheds = append(scheds, s) |
| 1152 | } |
| 1153 | m.mu.Unlock() |
| 1154 | for _, s := range scheds { |
| 1155 | var lifted []string |
| 1156 | s.mu.Lock() |
| 1157 | for key, c := range s.q.cohorts { |
| 1158 | build, adopt := strings.CutPrefix(key, "adopt:") |
| 1159 | if !adopt || !c.halted { |
| 1160 | continue |
| 1161 | } |
| 1162 | if _, held := m.d.Book.Lookup(s.key.acct, s.key.app, build); !held { |
| 1163 | s.q.unhalt(build) |
| 1164 | lifted = append(lifted, build) |
| 1165 | } |
| 1166 | } |
| 1167 | s.q.dropRetries(func(build, host string) bool { |
| 1168 | return m.d.Book.Retried(s.key.acct, s.key.app, build, host) |
| 1169 | }) |
| 1170 | s.mu.Unlock() |
| 1171 | if len(lifted) > 0 { |
| 1172 | m.d.Logf("rollout: %s/%s: halt lifted for %v, their failure records are gone", s.key.acct, s.key.app, lifted) |
| 1173 | s.poke() |
| 1174 | m.changed() |
| 1175 | } |
| 1176 | } |
| 1177 | } |
| 1178 | |
| 1179 | // Bootstrap is spec §8.5's first start: every awake instance records the |
| 1180 | // build it serves as good. Nothing is inferred from Version or PrevVersion. |
| 1181 | // No execution of this agent started the generation, so the write has no |
| 1182 | // book snapshot and clears no hold-ran marker (Astra R3 N6). |
| 1183 | func (m *Manager) Bootstrap(routes []registry.Route) { |
| 1184 | for _, rt := range routes { |
| 1185 | if !Managed(rt) || rt.GoodVersion != "" { |
| 1186 | continue |
| 1187 | } |
| 1188 | if g, ok := m.d.Act.Current(rt.Host); ok && g.Build != "" { |
| 1189 | m.writeGood(rt.Host, g, 0) |
| 1190 | } |
| 1191 | } |
| 1192 | } |
| 1193 | |
| 1194 | // Restart is a `carlos restart` request for host (spec §10). id is the |
| 1195 | // request's identity: the restart generation's etag, or a narrowed request's |
| 1196 | // id. A sleeping instance is never woken. One that is waking, stopping or |
| 1197 | // inside its first-wake observation keeps the request, which runs once it |
| 1198 | // is stable (spec §6.2). |
| 1199 | func (m *Manager) Restart(host, id string) bool { |
| 1200 | s := m.schedulerOf(host) |
| 1201 | if s == nil { |
| 1202 | return false |
| 1203 | } |
| 1204 | if !m.d.Act.Owns(host) { |
| 1205 | return false |
| 1206 | } |
| 1207 | s.mu.Lock() |
| 1208 | changed := s.q.wantRestart(host, id, m.d.Now()) |
| 1209 | s.mu.Unlock() |
| 1210 | if changed { |
| 1211 | s.poke() |
| 1212 | m.changed() |
| 1213 | } |
| 1214 | return changed |
| 1215 | } |
| 1216 | |
| 1217 | // ConfigDelivered is config sync's report that host's config file now holds |
| 1218 | // revision rev (spec §8.1). An instance with a process (awake, waking or |
| 1219 | // stopping) gets a pending config attempt; if its wake turns out to have read |
| 1220 | // rev, the wake's activation clears the attempt before it runs. A sleeping |
| 1221 | // one needs none, because its next wake starts with the file: rev counts as |
| 1222 | // what that wake will run, and config sync's ack goes out at once with |
| 1223 | // today's hibernating acknowledgement (configTrack.deferAck). A host this |
| 1224 | // manager does not manage is not tracked, so its ack goes out as before. |
| 1225 | func (m *Manager) ConfigDelivered(host, rev string) { |
| 1226 | s := m.schedulerOf(host) |
| 1227 | if s == nil { |
| 1228 | return |
| 1229 | } |
| 1230 | // Asked before m.mu: the activator is never called under the manager's |
| 1231 | // lock (the lock order on Manager). |
| 1232 | awake := m.d.Act.Owns(host) |
| 1233 | m.mu.Lock() |
| 1234 | t := m.track(host) |
| 1235 | t.delivered, t.fails, t.lastErr = rev, 0, "" |
| 1236 | t.sleeping = !awake |
| 1237 | if !awake { |
| 1238 | t.active = rev |
| 1239 | } |
| 1240 | m.mu.Unlock() |
| 1241 | if !awake { |
| 1242 | // A sleeper needs no attempt, including one still pending for an |
| 1243 | // older revision it was awake for: its next wake starts with this |
| 1244 | // file. Left in place, that attempt would restart a generation |
| 1245 | // already running rev and then mark the older revision active. |
| 1246 | s.mu.Lock() |
| 1247 | s.q.dropConfig(host) |
| 1248 | s.mu.Unlock() |
| 1249 | return |
| 1250 | } |
| 1251 | s.mu.Lock() |
| 1252 | changed := s.q.wantConfig(host, rev, m.d.Now()) |
| 1253 | s.mu.Unlock() |
| 1254 | if changed { |
| 1255 | s.poke() |
| 1256 | m.changed() // the revision is delivered and not yet active, which Status shows |
| 1257 | } |
| 1258 | } |
| 1259 | |
| 1260 | // DeferAck wraps config sync's ack writer: a revision is acknowledged only |
| 1261 | // once it is active (spec §8.1). Every ack, one written at once included, |
| 1262 | // goes through the manager's writer (Manager.acks), numbered here under m.mu |
| 1263 | // with the transitions that release held ones: written in line, an ack at |
| 1264 | // delivery could land before an older held ack the writer still has, and be |
| 1265 | // overwritten by it (plan review #10). write is handed a context bounded by |
| 1266 | // ackWriteBound. |
| 1267 | func (m *Manager) DeferAck(write func(context.Context, appconfig.ConfigAck)) func(appconfig.ConfigAck) { |
| 1268 | return func(a appconfig.ConfigAck) { |
| 1269 | m.mu.Lock() |
| 1270 | now := m.ackLocked(m.track(a.Host).deferAck(a, write)) |
| 1271 | m.mu.Unlock() |
| 1272 | m.submitAck(a.Host, now) |
| 1273 | } |
| 1274 | } |
| 1275 | |
| 1276 | // ackLocked numbers an ack a config track handed back (nil when none). The |
| 1277 | // caller holds m.mu and is in the critical section of the transition that |
| 1278 | // released it, so the numbers follow the order of the transitions, which |
| 1279 | // the writer keeps per host (plan review #10). |
| 1280 | func (m *Manager) ackLocked(write func(context.Context)) ackJob { |
| 1281 | if write == nil { |
| 1282 | return ackJob{} |
| 1283 | } |
| 1284 | m.ackSeq++ |
| 1285 | return ackJob{seq: m.ackSeq, write: write} |
| 1286 | } |
| 1287 | |
| 1288 | // submitAck hands an ack to the writer; it never blocks. |
| 1289 | func (m *Manager) submitAck(host string, j ackJob) { |
| 1290 | if j.write != nil { |
| 1291 | m.acks.submit(host, j) |
| 1292 | } |
| 1293 | } |
| 1294 | |
| 1295 | // resolve turns a dispatch into the Attempt Swap runs: its artifact and mode. |
| 1296 | // Recovery onto the serving build uses the generation's captured identity, |
| 1297 | // and a move back onto the good build uses the retained one; neither is |
| 1298 | // checked against the canonical manifest (spec §4.2). An adoption checks the |
| 1299 | // staged file exists and then the canonical demand. A missing staged file, |
| 1300 | // or a staged file that is not the release's canonical artifact |
| 1301 | // (release.ErrStagedDigestMismatch), is a staging refusal, never a build |
| 1302 | // failure. Any other trouble reading the demand runs the attempt in hold: |
| 1303 | // missing or unverifiable policy data never permits an overlap. |
| 1304 | func (m *Manager) resolve(ctx context.Context, d dispatch, rt registry.Route) (activator.Attempt, error) { |
| 1305 | acct, app := release.RouteAccount(rt), rt.App |
| 1306 | policy := m.policies.get(ctx, acct, app) |
| 1307 | a := activator.Attempt{Host: d.ID.Host, Cause: d.ID.Cause, Build: d.ID.Build, ConfigRev: d.ID.ConfigRev, |
| 1308 | RestartID: d.ID.RestartID, SchedRev: d.SchedRev, Canary: d.Canary} |
| 1309 | if g, ok := m.d.Act.Current(rt.Host); ok && g.Build == d.ID.Build { |
| 1310 | a.StagedPath, a.Digest, a.Demand = g.StagedPath, g.Digest, g.Demand |
| 1311 | a.Mode = RecoveryMode(policy, g.Demand) |
| 1312 | return a, nil |
| 1313 | } |
| 1314 | path := release.VersionPath(m.d.OptDir, acct, d.ID.Build, app) |
| 1315 | if d.ID.Build == rt.GoodVersion && rt.GoodDigest != "" { |
| 1316 | a.StagedPath, a.Digest, a.Demand = path, rt.GoodDigest, rt.GoodDemand |
| 1317 | a.Mode = RecoveryMode(policy, rt.GoodDemand) |
| 1318 | return a, nil |
| 1319 | } |
| 1320 | demand, digest, err := m.stagedDemand(ctx, rt, d.ID.Build, path) |
| 1321 | if err != nil { |
| 1322 | return a, err |
| 1323 | } |
| 1324 | a.StagedPath, a.Digest, a.Demand = path, digest, demand |
| 1325 | a.Mode = AttemptMode(policy, demand) |
| 1326 | return a, nil |
| 1327 | } |
| 1328 | |
| 1329 | // stagedDemand is the canonical demand and digest of build's staged file at |
| 1330 | // path, for an adoption attempt (resolve) and a first wake (Plan) alike |
| 1331 | // (spec §4.2). A missing staged file, or one that is not the release's |
| 1332 | // canonical artifact (release.ErrStagedDigestMismatch), is an error: a |
| 1333 | // staging refusal, never a build failure. Any other trouble reading the |
| 1334 | // demand is hold, on the staged bytes hashed here. |
| 1335 | func (m *Manager) stagedDemand(ctx context.Context, rt registry.Route, build, path string) (demand, digest string, err error) { |
| 1336 | if _, err := os.Stat(path); err != nil { |
| 1337 | return "", "", err |
| 1338 | } |
| 1339 | // Bounded (Task 11 rule 2): a demand read that times out is a read |
| 1340 | // error, so the attempt or the wake runs in hold, as below. |
| 1341 | rctx, cancel := context.WithTimeout(ctx, readBound) |
| 1342 | demand, digest, err = m.d.Demand(rctx, release.RouteAccount(rt), rt.App, build, path) |
| 1343 | cancel() |
| 1344 | if errors.Is(err, release.ErrStagedDigestMismatch) { |
| 1345 | return "", "", err |
| 1346 | } |
| 1347 | if err != nil { |
| 1348 | // The canonical manifest could not be read or compared: hold, on |
| 1349 | // the bytes hashed below. |
| 1350 | m.d.Logf("rollout: %s: canonical demand of %s: %v (runs in hold)", rt.Host, build, err) |
| 1351 | demand, digest = release.DemandHold, "" |
| 1352 | } |
| 1353 | if digest == "" { |
| 1354 | // Nothing compared the staged bytes with the manifest (plan 4: |
| 1355 | // DemandHold with no digest), so the demand read describes nothing |
| 1356 | // verified: hold. The build still runs only on the bytes hashed |
| 1357 | // here, because plan 3's verifyStaged refuses an empty digest. |
| 1358 | demand = release.DemandHold |
| 1359 | if digest, err = release.FileSHA256(path); err != nil { |
| 1360 | return "", "", err |
| 1361 | } |
| 1362 | } |
| 1363 | return demand, digest, nil |
| 1364 | } |
| 1365 | |
| 1366 | // served records a generation that became current after its checks: the |
| 1367 | // serving build, and the good build (spec §8.5). Both only while g is the |
| 1368 | // generation the activator serves (acknowledged). snap is the book's |
| 1369 | // snapshot from the start of the execution that produced g (writeGoodLocked). |
| 1370 | func (m *Manager) served(host string, g activator.Gen, snap uint64) { |
| 1371 | if g.Build == "" { |
| 1372 | return |
| 1373 | } |
| 1374 | m.goodMu.Lock() |
| 1375 | defer m.goodMu.Unlock() |
| 1376 | if !m.acknowledged(host, g) { |
| 1377 | m.d.Logf("rollout: %s: generation %d (%s) is no longer served; not recording it", host, g.N, g.Build) |
| 1378 | return |
| 1379 | } |
| 1380 | m.setServingLocked(host, g) |
| 1381 | m.writeGoodLocked(host, g, snap) |
| 1382 | } |
| 1383 | |
| 1384 | // servedOneShot records the good build of a replacement whose build exited |
| 1385 | // cleanly as a one-shot (spec §6.5, §8.5; Astra R3 flag triage T9). Nothing |
| 1386 | // of it keeps running, so the activator has no current generation for the |
| 1387 | // host to acknowledge, and served would drop the write. It skips only the |
| 1388 | // acknowledged check: done() calls it before the attempt's slot is freed, |
| 1389 | // so no newer attempt for the host can have been admitted, let alone |
| 1390 | // completed, and the result is the host's newest outcome. Everything else |
| 1391 | // is served's: goodMu, the generation's identity, and the book snapshot of |
| 1392 | // the execution's start (snap), so only a hold-ran marker older than the |
| 1393 | // attempt is cleared. A one-shot has no serving build to publish, and a |
| 1394 | // worker's is the activator's alone (setServingLocked). |
| 1395 | func (m *Manager) servedOneShot(host string, g activator.Gen, snap uint64) { |
| 1396 | if g.Build == "" { |
| 1397 | return |
| 1398 | } |
| 1399 | m.goodMu.Lock() |
| 1400 | defer m.goodMu.Unlock() |
| 1401 | m.writeGoodLocked(host, g, snap) |
| 1402 | } |
| 1403 | |
| 1404 | // setServingLocked writes an instance's serving build. A worker's is the |
| 1405 | // activator's alone (Task 11, publishWorkerServing, ordered by incarnation), |
| 1406 | // so two writers never race on one row. The caller holds m.goodMu. |
| 1407 | func (m *Manager) setServingLocked(host string, g activator.Gen) { |
| 1408 | if g.Worker { |
| 1409 | return |
| 1410 | } |
| 1411 | if err := m.d.Routes.SetServingVersion(host, g.Build); err != nil { |
| 1412 | m.d.Logf("rollout: %s: record serving build %s: %v", host, g.Build, err) |
| 1413 | } |
| 1414 | } |
| 1415 | |
| 1416 | // publishServing records the serving build alone (a healthy first wake, a |
| 1417 | // fallback), on the same terms as served. |
| 1418 | func (m *Manager) publishServing(host string, g activator.Gen) { |
| 1419 | if g.Build == "" { |
| 1420 | return |
| 1421 | } |
| 1422 | m.goodMu.Lock() |
| 1423 | defer m.goodMu.Unlock() |
| 1424 | if !m.acknowledged(host, g) { |
| 1425 | return |
| 1426 | } |
| 1427 | m.setServingLocked(host, g) |
| 1428 | } |
| 1429 | |
| 1430 | // acknowledged reports whether g is the generation the activator serves |
| 1431 | // for host right now (spec §8.5: the good build is written expecting the |
| 1432 | // current one). A late completion or event of an older generation is not, |
| 1433 | // so it can never overwrite what a newer one wrote. Generation numbers |
| 1434 | // decide when both sides have one; a worker's (no number) is compared by |
| 1435 | // its staged artifact. |
| 1436 | func (m *Manager) acknowledged(host string, g activator.Gen) bool { |
| 1437 | cur, ok := m.d.Act.Current(host) |
| 1438 | if !ok { |
| 1439 | return false |
| 1440 | } |
| 1441 | if cur.N != 0 && g.N != 0 { |
| 1442 | return cur.N == g.N |
| 1443 | } |
| 1444 | return cur.StagedPath == g.StagedPath && cur.Digest == g.Digest |
| 1445 | } |
| 1446 | |
| 1447 | // writeGood is writeGoodLocked for a caller outside served: Bootstrap. |
| 1448 | // Same terms: g must be acknowledged. |
| 1449 | func (m *Manager) writeGood(host string, g activator.Gen, snap uint64) { |
| 1450 | m.goodMu.Lock() |
| 1451 | defer m.goodMu.Unlock() |
| 1452 | if !m.acknowledged(host, g) { |
| 1453 | return |
| 1454 | } |
| 1455 | m.writeGoodLocked(host, g, snap) |
| 1456 | } |
| 1457 | |
| 1458 | // writeGoodLocked replaces host's good build with g's identity, |
| 1459 | // conditionally on the row still holding what was just read. goodMu |
| 1460 | // serialises every writer and each has just proven g current, so a retry |
| 1461 | // only covers a write from outside this manager; a second loss is logged |
| 1462 | // and the next serve tries again. The caller holds m.goodMu. |
| 1463 | // |
| 1464 | // A good build recorded (written, or found already written) clears the |
| 1465 | // host's hold-ran marker (final review wave 2, N2): the instance passed on |
| 1466 | // g, so g is what a later wake may fall back to. Only a marker set before |
| 1467 | // g's execution began (snap, Astra R3 N6): the write can be delayed after |
| 1468 | // its acknowledgement check, and a hold build that ran on the host |
| 1469 | // meanwhile (an attempt done() records without goodMu) left a marker the |
| 1470 | // instance has not passed over. |
| 1471 | func (m *Manager) writeGoodLocked(host string, g activator.Gen, snap uint64) { |
| 1472 | if g.Build == "" || g.Digest == "" { |
| 1473 | return |
| 1474 | } |
| 1475 | next := registry.GoodBuild{Version: g.Build, Digest: g.Digest, Demand: g.Demand} |
| 1476 | for range 2 { |
| 1477 | cur, err := m.d.Routes.Get(host) |
| 1478 | if err != nil || cur == nil { |
| 1479 | return |
| 1480 | } |
| 1481 | if cur.Good() == next { |
| 1482 | m.goodRecorded(host, next.Version, snap) |
| 1483 | return |
| 1484 | } |
| 1485 | err = m.d.Routes.SetGoodBuild(host, cur.Good(), next) |
| 1486 | if err == nil { |
| 1487 | m.goodRecorded(host, next.Version, snap) |
| 1488 | return |
| 1489 | } |
| 1490 | if !errors.Is(err, registry.ErrGoodBuildChanged) { |
| 1491 | m.d.Logf("rollout: %s: record good build %s: %v", host, g.Build, err) |
| 1492 | return |
| 1493 | } |
| 1494 | } |
| 1495 | m.d.Logf("rollout: %s: good build moved twice under this write, leaving it", host) |
| 1496 | } |
| 1497 | |
| 1498 | // goodRecorded follows a good-build write that landed (writeGoodLocked): |
| 1499 | // a hold-ran marker older than the write's execution (snap) goes, and the |
| 1500 | // queue learns the build. |
| 1501 | func (m *Manager) goodRecorded(host, ver string, snap uint64) { |
| 1502 | if _, err := m.d.Book.ClearHoldRanBefore(host, snap); err != nil { |
| 1503 | m.d.Logf("rollout: %s: clear hold-ran marker after good build %s: %v (cleared in memory)", host, ver, err) |
| 1504 | } |
| 1505 | m.observeGood(host, ver) |
| 1506 | } |
| 1507 | |
| 1508 | func (m *Manager) observeGood(host, ver string) { |
| 1509 | s := m.schedulerOf(host) |
| 1510 | if s == nil { |
| 1511 | return |
| 1512 | } |
| 1513 | s.mu.Lock() |
| 1514 | if h := s.q.hosts[host]; h != nil { |
| 1515 | h.good = ver |
| 1516 | } |
| 1517 | s.mu.Unlock() |
| 1518 | } |
| 1519 | |
| 1520 | func (m *Manager) configActivated(host, rev string) { |
| 1521 | if rev == "" { |
| 1522 | return |
| 1523 | } |
| 1524 | m.mu.Lock() |
| 1525 | write, _ := m.track(host).activate(rev) |
| 1526 | ack := m.ackLocked(write) |
| 1527 | m.mu.Unlock() |
| 1528 | m.activationDone(host, rev, ack) |
| 1529 | } |
| 1530 | |
| 1531 | // activationDone is the rest of an activation of rev, after the config |
| 1532 | // track has recorded it under m.mu: the queue's pending config attempt for |
| 1533 | // rev is cleared, and the held ack (ack, empty when none), numbered in that |
| 1534 | // critical section, is handed to the writer. Called with no lock held. The |
| 1535 | // ack is not written here (Task 11 rule 3, Astra R1 Flag triage T8): done() |
| 1536 | // calls this before the attempt's slot frees, and configSlept from the |
| 1537 | // scheduler loop, and neither may wait on the store. |
| 1538 | func (m *Manager) activationDone(host, rev string, ack ackJob) { |
| 1539 | if s := m.schedulerOf(host); s != nil { |
| 1540 | s.mu.Lock() |
| 1541 | s.q.configActive(host, rev) |
| 1542 | s.mu.Unlock() |
| 1543 | } |
| 1544 | m.submitAck(host, ack) |
| 1545 | } |
| 1546 | |
| 1547 | // configSlept is configActivated for an instance seen asleep with rev |
| 1548 | // pending (spec §8.1): its next wake starts with the delivered file, so rev |
| 1549 | // is what that wake runs and no config attempt is left to make. A held ack |
| 1550 | // goes out as a sleeper's does at delivery, marked Hibernating. rev was read |
| 1551 | // before the activator reported the instance asleep, so the file already |
| 1552 | // held it when the instance stopped: this can only under-claim. |
| 1553 | func (m *Manager) configSlept(host, rev string) { |
| 1554 | m.mu.Lock() |
| 1555 | t := m.track(host) |
| 1556 | if t.ack != nil && t.ackRev == rev { |
| 1557 | t.ack.Hibernating, t.ack.Restarted = true, false |
| 1558 | } |
| 1559 | write, _ := t.activate(rev) |
| 1560 | ack := m.ackLocked(write) |
| 1561 | m.mu.Unlock() |
| 1562 | m.activationDone(host, rev, ack) |
| 1563 | } |
| 1564 | |
| 1565 | // configFailure counts a failed attempt that carried config revision rev |
| 1566 | // and returns its backoff. Only a failure of the delivered revision counts: |
| 1567 | // one of an older revision says nothing about the newer one's attempts. The |
| 1568 | // caller sets the queue's backoff (appQueue.configFailed) in the step that |
| 1569 | // frees the attempt's slot. |
| 1570 | func (m *Manager) configFailure(host, rev, reason string) (time.Duration, bool) { |
| 1571 | m.mu.Lock() |
| 1572 | defer m.mu.Unlock() |
| 1573 | t := m.track(host) |
| 1574 | if rev == "" || rev != t.delivered { |
| 1575 | return 0, false |
| 1576 | } |
| 1577 | t.fails++ |
| 1578 | t.lastErr = reason |
| 1579 | return configBackoff(t.fails), true |
| 1580 | } |
| 1581 | |
| 1582 | // cohortFailed runs the cohort-failure transition for s's app. A failure of |
| 1583 | // an execution that began before a retry of its build runs none of it |
| 1584 | // (PreRetry, ruling 191); a hold-down it asked for is a hold-ran marker |
| 1585 | // instead (FailureBook.Record), whose rollback read is then due. |
| 1586 | func (m *Manager) cohortFailed(s *scheduler, rt registry.Route, rec FailureRecord) { |
| 1587 | s.mu.Lock() |
| 1588 | res, superseded, err := failCohort(m.d.Book, s.q, rec) |
| 1589 | s.mu.Unlock() |
| 1590 | if err != nil { |
| 1591 | m.d.Logf("rollout: write failure record %s/%s %s: %v (held in memory)", rec.Account, rec.App, rec.Build, err) |
| 1592 | } |
| 1593 | if res == PreRetry { |
| 1594 | m.d.Logf("rollout: %s/%s build %s failed on %s, but in an execution that began before a retry of the build: the retry stands", rec.Account, rec.App, rec.Build, rec.Host) |
| 1595 | if slices.Contains(rec.HeldDown, rec.Host) { |
| 1596 | m.pointerReadDue(s.key) |
| 1597 | } |
| 1598 | s.poke() |
| 1599 | return |
| 1600 | } |
| 1601 | if len(superseded) > 0 { |
| 1602 | m.d.Logf("rollout: %s/%s build %s failed: superseding attempts past spawn on %v", rec.Account, rec.App, rec.Build, superseded) |
| 1603 | } |
| 1604 | if res == Recorded { |
| 1605 | m.d.Logf("rollout: %s/%s build %s failed on %s: %s", rec.Account, rec.App, rec.Build, rec.Host, rec.Reason) |
| 1606 | m.emit(rt, EvCohortHalted, "warn", textCohortHalted(rec.Build, m.d.Box, reasonWord(rec.Reason))) |
| 1607 | } |
| 1608 | s.poke() |
| 1609 | } |
| 1610 | |
| 1611 | func (m *Manager) stagingRefused(rt registry.Route, build string, err error) { |
| 1612 | m.d.Logf("rollout: %s: build %s refused as a staging error: %v", rt.Host, build, err) |
| 1613 | m.emit(rt, EvStagingRefused, "warn", textStagingRefused(build)) |
| 1614 | if m.d.Touch != nil { |
| 1615 | m.d.Touch() |
| 1616 | } |
| 1617 | } |
| 1618 | |
| 1619 | // emit logs an event and publishes it to the member's log. text is built |
| 1620 | // from build names and reason words only; RedactPaths stays as a second |
| 1621 | // guard against a box path in a build name. |
| 1622 | func (m *Manager) emit(rt registry.Route, code EventCode, level, text string) { |
| 1623 | m.d.Logf("rollout: %s %s: %s", code, rt.Host, text) |
| 1624 | if m.d.Events != nil { |
| 1625 | m.d.Events.Event(release.RouteAccount(rt), rt.App, rt.Host, level, activator.RedactPaths(text)) |
| 1626 | } |
| 1627 | } |
| 1628 | |
amadan