| 1 | // The box half of a sidecar wake request: when a sidecar's wake object |
| 2 | // (release.SidecarWakeKey) MOVES, wake exactly that sidecar. Pull-based like |
| 3 | // the restart signal it is modelled on (restart.go) — no push channel, no |
| 4 | // box access, and the box need not even be reachable from anywhere for this |
| 5 | // to work. |
| 6 | // |
| 7 | // It exists because a sidecar has no socket and nothing dials it: an |
| 8 | // instance wakes on its next inbound request, but a hibernating sidecar has |
| 9 | // no "next request" to wake on, so waking one from outside the box needs an |
| 10 | // explicit signal a pull-based box can notice on its own. |
| 11 | // |
| 12 | // Split into the same two shapes restart.go uses for exactly the same |
| 13 | // reason: sidecarWakeChanges is the cheap, synchronous etag comparison |
| 14 | // (restartGenPass's counterpart), while dispatchSidecarWakes hands the |
| 15 | // WakeWorker fan-out to restart.go's shared dispatchBoundedWakes — which |
| 16 | // can block up to wakeWaitBound (30s, activator.go) per sidecar waiting |
| 17 | // out the settle window, and runs the WHOLE bounded fan-out inside one |
| 18 | // goroutine it launches itself, so the caller returns immediately |
| 19 | // regardless of how many sidecars are queued (review round 2: an app |
| 20 | // shaped like Eleven's, a sidecar per enrolled member, is routinely more |
| 21 | // than workerBootConcurrency). sidecarWakeLeg is the two glued together and |
| 22 | // wired into fastConvergeLoop beside the restart leg (agent.go's |
| 23 | // fastDeps.wakeLeg), so it rides the same 2s serial-driven convergence |
| 24 | // `carlos restart` already gets. |
| 25 | package host |
| 26 | |
| 27 | import ( |
| 28 | "context" |
| 29 | "encoding/json" |
| 30 | "errors" |
| 31 | "fmt" |
| 32 | "log" |
| 33 | "os" |
| 34 | |
| 35 | "github.com/carlosframework/platform/internal/registry" |
| 36 | "github.com/carlosframework/platform/internal/release" |
| 37 | ) |
| 38 | |
| 39 | // sidecarsServed maps every sidecar route this box serves to the (account, |
| 40 | // app, name) triple its wake object lives at — name comes from Slug, the |
| 41 | // field a sidecar route carries its short name on (registry.Route's own |
| 42 | // doc comment; provision's reconciler stamps it from the sidecar record). |
| 43 | // |
| 44 | // Gated on rt.IsSidecar() explicitly, not just a non-empty Slug: Slug is |
| 45 | // NOT sidecar-exclusive (it is also an ordinary app's root-domain short |
| 46 | // name), so an instance route that happens to carry the same Slug as a real |
| 47 | // sidecar must never be mistaken for that sidecar. |
| 48 | func sidecarsServed(routes []registry.Route) map[[3]string]registry.Route { |
| 49 | out := make(map[[3]string]registry.Route) |
| 50 | for _, rt := range routes { |
| 51 | if !rt.IsSidecar() || rt.Slug == "" { |
| 52 | continue |
| 53 | } |
| 54 | out[[3]string{release.RouteAccount(rt), rt.App, rt.Slug}] = rt |
| 55 | } |
| 56 | return out |
| 57 | } |
| 58 | |
| 59 | // sidecarWakeMemory is the wake leg's memory across ticks: the etag last |
| 60 | // seen for each served sidecar, plus the one bit that says whether the |
| 61 | // arming BASELINE has been established. |
| 62 | // |
| 63 | // The two live together deliberately. "First sight" means two completely |
| 64 | // different things either side of that baseline — history a booting box must |
| 65 | // not replay, versus a route that has just appeared under a running agent — |
| 66 | // and a bare map cannot tell them apart, which is exactly how the |
| 67 | // create-then-wake flow came to swallow its own first wake. |
| 68 | type sidecarWakeMemory struct { |
| 69 | claimsEnabled bool |
| 70 | identities map[string]release.SidecarWake |
| 71 | etags map[[3]string]string |
| 72 | // seeded is set by the first completed pass (the arming one in |
| 73 | // production, sidecarWakeSeed; the first leg pass when arming's route |
| 74 | // list could not be read). Before it, first sight is the baseline; after |
| 75 | // it, first sight is a NEW route. |
| 76 | seeded bool |
| 77 | } |
| 78 | |
| 79 | func newSidecarWakeMemory(claimsEnabled bool) *sidecarWakeMemory { |
| 80 | return &sidecarWakeMemory{claimsEnabled: claimsEnabled, identities: map[string]release.SidecarWake{}, etags: map[[3]string]string{}} |
| 81 | } |
| 82 | |
| 83 | // sidecarWakeChanges is sidecar-wake's counterpart to restartGenPass: read |
| 84 | // the wake-object etag of every sidecar this box serves and report which |
| 85 | // ones MOVED, as their routes. It performs no wake itself — see |
| 86 | // dispatchSidecarWakes for the half that can block. |
| 87 | // |
| 88 | // - First sight BEFORE the baseline exists records silently and is never |
| 89 | // reported as moved. A box (or an agent process) that has just started |
| 90 | // must not wake every sidecar it serves because it has never seen these |
| 91 | // etags before — that is the seed-don't-replay rule the serial watcher |
| 92 | // already applies (agent.go's fastTick). Without it, an agent restart |
| 93 | // would replay every wake this app's sidecars have ever received — a |
| 94 | // thundering herd on boot, and for a hibernating fleet a real cost, not |
| 95 | // a cosmetic one. sidecarWakeSeed runs one such silent pass at arming so |
| 96 | // the FIRST request after a boot is acted on rather than swallowed. |
| 97 | // - First sight AFTER it, with a wake object already present, is reported |
| 98 | // as MOVED. Such a route did not exist when this agent last looked, so |
| 99 | // it has just been minted — and a wake object standing at a route this |
| 100 | // very pass minted is a wake somebody already asked for and the CLI |
| 101 | // already reported as accepted. That is the create-then-wake flow, and |
| 102 | // it is the feature's PRIMARY one: an operator wake is a hibernating |
| 103 | // sidecar's only first-run mechanism (the timer needs a next-due that |
| 104 | // only a running child can set), and `carlos sidecars create` followed |
| 105 | // by `carlos sidecars wake` routinely lands both inside a single tick — |
| 106 | // the tick that mints the route would otherwise seed the wake's etag and |
| 107 | // silently drop the wake. A spurious extra wake is idempotent and cheap; |
| 108 | // a lost one is a bot that never runs. (A route that vanished and came |
| 109 | // back is a first sight too, and is woken once for the same reason: it |
| 110 | // has just been re-minted, and its wake object outlived it.) |
| 111 | // - An unchanged etag does nothing. |
| 112 | // - Absent that stays absent does nothing, forever: a sidecar nobody has |
| 113 | // ever woken this way has no object and never will until someone asks. |
| 114 | // A brand-new route with no wake object is therefore silent too — only a |
| 115 | // PRESENT object makes a first sight a wake. |
| 116 | // - Any movement counts: the etag changed, or absent -> present (the |
| 117 | // sidecar's first ever wake request). |
| 118 | // - A wake object naming a sidecar this box does not serve is never even |
| 119 | // read: this function only ever builds keys from sidecarsServed, which |
| 120 | // comes from THIS box's own route table. Such an object, however it |
| 121 | // moves in the bucket, is simply never asked about. |
| 122 | // - A read error for one sidecar leaves its cache entry untouched and the |
| 123 | // rest of the pass keeps going; the error comes back for the caller to |
| 124 | // aggregate into its own log line, same discipline as restartGenPass. |
| 125 | // After the baseline that means its next sighting is a first sight, so |
| 126 | // it may be woken once — the cheap direction of an unknown. |
| 127 | // |
| 128 | // The memory is mutated in place and is expected to live for the process |
| 129 | // lifetime, exactly like restartGenPass's cache. |
| 130 | func sidecarWakeChanges(ctx context.Context, routes []registry.Route, store release.Store, mem *sidecarWakeMemory) (moved []registry.Route, readErrs []error) { |
| 131 | served := sidecarsServed(routes) |
| 132 | live := make(map[[3]string]bool, len(served)) |
| 133 | baseline := !mem.seeded |
| 134 | for k, rt := range served { |
| 135 | live[k] = true |
| 136 | raw, etag, err := store.Get(ctx, release.SidecarWakeKey(k[0], k[1], k[2])) |
| 137 | switch { |
| 138 | case errors.Is(err, os.ErrNotExist): |
| 139 | etag = genAbsent |
| 140 | case err != nil: |
| 141 | readErrs = append(readErrs, fmt.Errorf("%s/%s/%s: %w", k[0], k[1], k[2], err)) |
| 142 | continue // cache entry untouched: unknown this tick, not changed |
| 143 | } |
| 144 | if mem.claimsEnabled { |
| 145 | var signal release.SidecarWake |
| 146 | if etag == genAbsent || json.Unmarshal(raw, &signal) != nil { |
| 147 | continue |
| 148 | } |
| 149 | if err := verifySidecarWake(ctx, store, rt, signal); err != nil { |
| 150 | readErrs = append(readErrs, err) |
| 151 | continue |
| 152 | } |
| 153 | if mem.identities == nil { |
| 154 | mem.identities = map[string]release.SidecarWake{} |
| 155 | } |
| 156 | mem.identities[rt.Host] = signal |
| 157 | etag = signal.AppID + "/" + signal.Incarnation + "/" + signal.RequestID |
| 158 | } |
| 159 | prev, seen := mem.etags[k] |
| 160 | mem.etags[k] = etag |
| 161 | switch { |
| 162 | case seen: |
| 163 | if prev != etag { |
| 164 | moved = append(moved, rt) |
| 165 | } |
| 166 | case baseline: |
| 167 | // The agent-restart protection: history, recorded silently. |
| 168 | case etag != genAbsent: |
| 169 | // A route minted since the last pass, with a wake already |
| 170 | // standing at it — the create-then-wake flow. |
| 171 | moved = append(moved, rt) |
| 172 | } |
| 173 | } |
| 174 | for k := range mem.etags { |
| 175 | if !live[k] { |
| 176 | delete(mem.etags, k) |
| 177 | } |
| 178 | } |
| 179 | // One completed pass IS the baseline, whichever caller ran it: arming's |
| 180 | // silent seed in production, or the first leg pass on a box whose arming |
| 181 | // route list could not be read. |
| 182 | mem.seeded = true |
| 183 | return moved, readErrs |
| 184 | } |
| 185 | |
| 186 | // dispatchSidecarWakes wakes every route in moved through the shared |
| 187 | // dispatchBoundedWakes helper (restart.go) — the SAME idiom Task 8's |
| 188 | // runWorkerBootPass and cycleWorkers' own wake half use, deliberately not a |
| 189 | // third shape. WakeWorker can block up to wakeWaitBound (30s) per sidecar |
| 190 | // waiting out the settle window, and this function's only caller |
| 191 | // (sidecarWakeLeg) runs synchronously on the box's single per-tick |
| 192 | // convergence goroutine (fastTick) — dispatchBoundedWakes' own doc comment |
| 193 | // covers why the WHOLE bounded fan-out, not just each individual wake, |
| 194 | // must be off that goroutine (review round 2: an app shaped like Eleven's, |
| 195 | // a sidecar per enrolled member, routinely queues more than |
| 196 | // workerBootConcurrency sidecars behind one wake burst). Safe because |
| 197 | // WakeWorker is idempotent and a wake request's contract is "wake the |
| 198 | // process", not "return once it's up" — a sidecar that comes up a moment |
| 199 | // after this tick has still been woken. A wake that errors is logged and |
| 200 | // does not stop, delay, or fail the others. |
| 201 | // |
| 202 | // wake may be nil (no activator armed on this box): a working no-op. |
| 203 | func dispatchSidecarWakes(moved []registry.Route, wake func(rt registry.Route) error) { |
| 204 | dispatchBoundedWakes(moved, wake, func(rt registry.Route, err error) { |
| 205 | if err != nil { |
| 206 | log.Printf("agent: sidecar wake requested: wake %s: %v", rt.Host, err) |
| 207 | return |
| 208 | } |
| 209 | log.Printf("agent: sidecar wake requested: woke %s", rt.Host) |
| 210 | }) |
| 211 | } |
| 212 | |
| 213 | // sidecarWakeLeg is sidecarWakeChanges and dispatchSidecarWakes glued |
| 214 | // together — the leg fastConvergeLoop wires into fastDeps.wakeLeg beside |
| 215 | // the restart leg, so a wake request rides the same 2s serial-driven |
| 216 | // convergence `carlos restart` already gets. Reports whether anything |
| 217 | // moved (for logging/tests); the actual wake calls happen off this |
| 218 | // function's return path (dispatchSidecarWakes), so it never blocks on a |
| 219 | // slow or unresponsive sidecar. |
| 220 | func sidecarWakeLeg(ctx context.Context, routes []registry.Route, store release.Store, mem *sidecarWakeMemory, wake func(rt registry.Route) error) (moved []registry.Route, readErrs []error) { |
| 221 | moved, readErrs = sidecarWakeChanges(ctx, routes, store, mem) |
| 222 | if len(readErrs) > 0 { |
| 223 | log.Printf("agent: sidecar wake check: %d sidecars unreadable: %v", len(readErrs), readErrs[0]) |
| 224 | } |
| 225 | if len(moved) == 0 { |
| 226 | return moved, readErrs |
| 227 | } |
| 228 | if mem.claimsEnabled { |
| 229 | captured := make(map[string]release.SidecarWake, len(moved)) |
| 230 | for _, rt := range moved { |
| 231 | captured[rt.Host] = mem.identities[rt.Host] |
| 232 | } |
| 233 | original := wake |
| 234 | wake = func(rt registry.Route) error { |
| 235 | if err := verifySidecarWake(ctx, store, rt, captured[rt.Host]); err != nil { |
| 236 | return err |
| 237 | } |
| 238 | rt.WakeIncarnation = captured[rt.Host].Incarnation |
| 239 | if original == nil { |
| 240 | return nil |
| 241 | } |
| 242 | return original(rt) |
| 243 | } |
| 244 | } |
| 245 | dispatchSidecarWakes(moved, wake) |
| 246 | log.Printf("agent: sidecar wake requested for %d sidecar(s) on this box", len(moved)) |
| 247 | return moved, readErrs |
| 248 | } |
| 249 | |
| 250 | // sidecarWakeSeed is sidecarWakeChanges' counterpart to restartSeed: at |
| 251 | // arming, read each served sidecar's wake-object etag once, silently, so |
| 252 | // the memory sidecarWakeLeg compares against exists before the first serial |
| 253 | // move. Without it the first wake request after a boot is the sweep that |
| 254 | // establishes the baseline, and is therefore the one request that does |
| 255 | // nothing. |
| 256 | // |
| 257 | // This pass IS the agent-restart protection: only a sidecar this box already |
| 258 | // served when the agent armed has its history recorded silently. Everything |
| 259 | // that appears afterwards is a route minted under a running agent, and a |
| 260 | // wake object already standing at one of those is acted on (see |
| 261 | // sidecarWakeChanges). |
| 262 | // |
| 263 | // A LIST failure leaves the memory unseeded on purpose: the first leg pass |
| 264 | // then establishes the baseline itself rather than reading an empty cache as |
| 265 | // "every sidecar on this box is brand new" and waking the lot. A per-sidecar |
| 266 | // read error is narrower — that one sidecar is left out of the cache and may |
| 267 | // be woken once on its next sighting. |
| 268 | func sidecarWakeSeed(ctx context.Context, reg *registry.Registry, store release.Store, mem *sidecarWakeMemory) { |
| 269 | routes, err := reg.List() |
| 270 | if err != nil { |
| 271 | log.Printf("agent: sidecar wake baseline: list routes: %v (the first request may need a second try)", err) |
| 272 | return |
| 273 | } |
| 274 | _, readErrs := sidecarWakeChanges(ctx, routes, store, mem) |
| 275 | if len(readErrs) > 0 { |
| 276 | log.Printf("agent: sidecar wake baseline: %d sidecars unreadable: %v", len(readErrs), readErrs[0]) |
| 277 | } |
| 278 | } |
| 279 | |
amadan