Press n or j to go to the next uncovered block, b, p or k for the previous block.
| 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 | 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 28x 28x 7x 7x 21x 28x 28x 1x 1x 138x 138x 138x 138x 130x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 2290x 2290x 2290x 2290x 2290x 2290x 2290x 2290x 138x 5004x 5004x 5004x 5004x 5004x 5004x 5004x 5004x 5004x 5004x 5004x 5004x 138x 2085x 2085x 2085x 2085x 2085x 2085x 138x 3614x 3614x 138x 1x 1x 1x 1x 1x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 17x 5x 5x 5x 5x 5x 5x 5x 5x 5x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 1x 5x 5x 17x 8x 8x 4x 4x 4x 4x 4x 4x 17x 138x 138x 138x 4891x 4882x 9x 9x 9x 9x 9x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 4900x 4900x 4900x 4900x 4900x 4900x 4900x 4900x 4900x 4900x 11x 11x 11x 11x 11x 4889x 4889x 4889x 4889x 4889x 4900x 212x 212x 212x 212x 212x 212x 212x 212x 212x 212x 212x 4888x 4888x 4888x 4888x 4888x 4888x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 6x 6x 6x 6x 138x 138x 14x 14x 14x 14x 1x 1x 13x 13x 13x 14x 6x 6x 14x 1x 1x 6x 6x 6x 6x 6x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 5x 5x 5x 1x 1x 5x 1x 1x 1x 1x 1x 3x 3x 3x 3x 3x 3x 138x 138x 4x 4x 4x 2x 2x 2x 2x 138x 138x 138x 138x 138x 138x 138x 3x 3x 3x 3x 3x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 9x 9x 6x 6x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 3x 7x 1x 1x 2x 2x 2x 2x 2x 2x 2x 2x 2x 2x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 9x 9x 1x 1x 8x 8x 8x 8x 8x 8x 18x 18x 18x 18x 18x 18x 18x 8x 19x 19x 19x 19x 19x 19x 19x 19x 8x 8x 8x 8x 8x 8x 8x 138x 138x 138x 138x 138x 138x 138x 138x 138x 10x 10x 10x 10x 27x 27x 27x 27x 27x 10x 10x 10x 10x 10x 10x 10x 10x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 138x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 19x 138x 138x 138x | /**
* The cluster service (PRD §4: the central half of the hybrid design).
*
* Route handlers are wired here and implemented in their own modules, so every
* capability added by a later task is one `register` call rather than an edit
* inside a growing route file.
*
* The routes are the service's whole public surface for now: ingesting a batch
* of template events (R3/R4/R6/R13), reading the registry (R4) raw or through
* one of the read-time facet projections (R16), reading a retained exemplar
* (R13), the change stream (R8), template search and the deploy diff (R14), the
* correlated incidents (R15) and retiring a template
* (`POST /templates/:ref/retire`, the writer behind R8's retired-template
* delta), the flows observed and their sequence diagrams (`GET /flows`,
* `GET /flows/:reference/diagram`, R22/R23). Ingest acknowledges
* with 202 and tells the emitter nothing else — the emitter must not wait on the
* service (R18) — while anomalies and incidents travel on the digest (R8)
* instead of in the response. Only a payload that is not a batch of events fails
* the request, with 400; an individual event that cannot be ingested is counted
* in the result's `skipped` and logged.
*
* When `persistTo` is configured the **template registry** is restored before
* the first request and saved after every change (R4). The other state this
* service holds — digest, lineage, incidents, flow-drift history — is
* process-lifetime by design; see `persistence.ts`. A snapshot that cannot be
* read does **not** stop the service: it is quarantined and the registry starts
* empty, loudly (see `onReady` below and `.github/memory/decision.md`, Task 12).
*/
import Fastify, {type FastifyInstance} from 'fastify';
import {rename} from 'node:fs/promises';
import {isUlid} from '../ulid.ts';
import {DetectorSuite, type Anomaly} from './detectors.ts';
import {modelOfExecution, modelOfUnion, renderSequence} from './diagram.ts';
import {DigestLog} from './digest.ts';
import {EmbeddingCache} from './embedding.ts';
import {ExemplarStore} from './exemplars.ts';
import {FACETS, isFacet, project, type FacetAnomaly} from './facets.ts';
import {FlowLedger} from './flowLedger.ts';
import {IncidentStore} from './incidents.ts';
import {createIngest, FlowDriftHistory, FlowShapes, type IngestResult} from './ingest.ts';
import {LineageIndex} from './lineage.ts';
import {loadSnapshot, saveSnapshot, SnapshotError, type Snapshot} from './persistence.ts';
import {createProvider, providerIdentity, type EmbeddingConfig} from './provider.ts';
import {TemplateRegistry, type TemplateEntry} from './registry.ts';
import {deployDiff, searchRecords, searchTemplates} from './search.ts';
export interface ServiceOptions {
/** Embedding provider selection (PRD R5). */
embedding?: EmbeddingConfig;
/**
* Path to the JSON snapshot: the **template registry** and the per-kind union of
* observed calls — the artifacts a restart cannot re-derive (PRD R4, D17). The
* digest, causal lineage, incident store, flow-drift history and the per-execution
* flow ring are process-lifetime surfaces and are deliberately *not* persisted (see
* `.github/memory/decision.md`, Task 12). Omitted, the service is entirely in-memory
* and loses all of its state on restart. A snapshot this service cannot use — written
* by another version or another embedding provider, or unreadable — is moved aside
* under a suffix naming the reason, and the service starts empty rather than refusing
* to serve.
*
* D18 ruled that union writes are debounced. They are not, and the reason is the file:
* the unions share the snapshot the registry is written to, and R4 requires the registry
* to be durable before a batch is acknowledged — so a save per accepted batch is already
* the design, and a debounce would only add a window in which observations are not on
* disk. See `.github/memory/decision.md`.
*/
persistTo?: string;
/** Enable request logging. Off by default so tests stay quiet. */
logger?: boolean;
/** Full records retained per template (PRD R13). */
exemplarLimit?: number;
/**
* The digest's eviction budget (PRD R8, see `digest.ts`).
*
* A caller that generates a burst sizes it against what the burst publishes:
* the stream keeps the *newest* entries, so a run whose reading is an older one
* — the F2 fixture asserting on the template a retry burst was stamped on —
* would be asserting on an entry the stream is entitled to have dropped.
* Defaults to the service's own budget, which a deployed process stays inside.
*/
digestLimit?: number;
/** Injected so tests can count embedding calls (PRD R3). */
cache?: EmbeddingCache;
/** Injected so a restarted service can continue from a loaded registry (PRD R4). */
registry?: TemplateRegistry;
/** Maximum executions whose shape is retained, open or closed (PRD R6c). */
flowLimit?: number;
/** Maximum steps retained per in-flight execution (PRD R6c). */
flowStepLimit?: number;
/** Injected so a caller can observe or seed the flow observations the diagrams are drawn from (R22/R23). */
flowLedger?: FlowLedger;
/** Injected so a caller can observe the retention bound. */
shapes?: FlowShapes;
/** Injected so a caller can observe or share the per-flow-kind drift history (PRD R14). */
driftHistory?: FlowDriftHistory;
/** Injected so a caller can observe or seed the causal lineage the correlation reads (R7/R15). */
lineage?: LineageIndex;
/** Injected so a caller can observe or seed the correlated incidents (R15). */
incidents?: IncidentStore;
/**
* Recent anomalies retained for cross-batch correlation (R15); ignored when
* an `incidents` store is injected, since the store owns its own buffer.
*/
incidentBufferLimit?: number;
/** How far apart two anomalies on one trace may be and still merge (R15). */
incidentWindowMs?: number;
/**
* The detector suite, so a deployment can tune the windows to its own clock.
*
* The shipped default scores a rate baseline over five completed **60-second**
* windows, so a caller whose activity happens in milliseconds can never fill
* that baseline and would never see R6b — the window has to match the
* timescale of the traffic being scored, which only the caller knows.
*/
detectors?: DetectorSuite;
}
/** Drift epsilon is a distance, not a similarity: the shapes are unit vectors. */
const DETECTORS = {
drift: {epsilon: 0.25, learningRate: 0.2},
rate: {windowMs: 60_000, buckets: 5, zThreshold: 4},
};
/**
* How far apart two anomalies on one trace may be and still merge into one
* incident (R15). A single failure's cross-service anomalies surface within
* seconds to a couple of minutes — the span of one request path — so five
* minutes is generous without merging two unrelated failures that share a trace
* id over a long-lived connection. Overridable per service.
*/
const DEFAULT_INCIDENT_WINDOW_MS = 300_000;
/**
* Read a numeric bound from the query string, falling back when it is absent,
* blank, or not a finite number. `Number('abc')` is `NaN`, `Number('')` is `0`,
* and `Number(undefined)` is `NaN` too. A `NaN` bound would make every
* comparison false — turning the diff into an all-empty answer that looks like
* "nothing changed" rather than failing or falling back — and a blank bound
* parsed as `0` would collapse a window to nothing, which is the same
* silently-wrong answer. A blank value is therefore treated exactly like an
* omitted one. The same rule the digest applies to its cursor (see `/digest`):
* an unreadable bound reads the whole history instead of silently hiding it.
*/
function readNumber(value: string | undefined, fallback: number): number {
if (value === undefined || value.trim() === '') {
return fallback;
}
const parsed = Number(value);
return Number.isFinite(parsed) ? parsed : fallback;
}
export function createApp(options: ServiceOptions = {}): FastifyInstance {
const app = Fastify({logger: options.logger ?? false});
const registry = options.registry ?? new TemplateRegistry();
const cache =
options.cache ??
new EmbeddingCache(createProvider(options.embedding ?? {kind: 'offline', dimension: 64}));
const detectors = options.detectors ?? new DetectorSuite(DETECTORS);
const exemplars = new ExemplarStore({limit: options.exemplarLimit ?? 5});
const digest = new DigestLog({limit: options.digestLimit ?? 1000});
const flowDrift = options.driftHistory ?? new FlowDriftHistory();
const ledger = options.flowLedger ?? new FlowLedger();
const lineage = options.lineage ?? new LineageIndex();
const incidents = options.incidents ?? new IncidentStore(options.incidentBufferLimit);
const incidentWindowMs = options.incidentWindowMs ?? DEFAULT_INCIDENT_WINDOW_MS;
const ingest = createIngest({
registry,
cache,
detectors,
exemplars,
shapes: options.shapes ?? new FlowShapes(options.flowLimit, options.flowStepLimit),
driftHistory: flowDrift,
onAnomaly: (anomaly, entry) => {
// `anomaly.ref` is the flow *kind* for a drift anomaly and the
// template ref for novelty and rate-shift, so it is republished as
// `anomalyRef`: a consumer that read it as a template ref would be
// wrong for drift. `templateRef` is always the template whose event
// triggered the anomaly (see `digest.ts`).
const {ref, ...rest} = anomaly;
digest.publish('anomaly', {...rest, anomalyRef: ref, templateRef: entry.ref});
},
onAccepted: event => {
// The causal index is fed from the ingest rather than the route so a
// record is present before its own anomaly is published (see
// `IngestDependencies.onAccepted`); correlation is then correct even
// when a concurrent batch's anomalies are drained by another request.
lineage.add(event);
// The flow observation is fed from the same seam, so one accepted event
// produces exactly one lineage node and one ledger observation, in the
// same order — which is what lets both ends of a call be seen by one
// reader, in one order (PRD R22). The batch that carried it is then saved
// by the route below, on the same guarantee the registry has (R4).
ledger.observe(event);
},
onTemplateAdded: entry => {
digest.publish('template-added', {
ref: entry.ref,
service: entry.service,
signature: entry.signature,
});
},
onExemplar: (entry, id) => {
digest.publish('exemplar-retained', {ref: entry.ref, record: id});
},
onSkipped: (error, event) => {
app.log.warn(
{err: error, eventId: event.id},
'semantic-log: skipped an event that could not be ingested',
);
},
});
// Persistence (PRD R4, D17-D19). The snapshot holds the template registry and the
// per-kind union of observed calls; the provider identity travels with it, because
// vectors are only comparable within one provider at one width (see `persistence.ts`).
const snapshotPath = options.persistTo;
// The identity of the provider this service embeds with. Read from the configuration
// that built the cache rather than from the cache, so the label and the vectors come
// from one decision (`embedding` defaults to the offline provider, as `cache` does).
const provider = providerIdentity(options.embedding ?? {kind: 'offline', dimension: 64});
if (snapshotPath) {
// Restore before the first request: `onReady` is awaited by both
// `listen` and `inject`, so a service handed a snapshot is never asked a
// question before it has read it.
//
// A snapshot this service cannot use does **not** fail the service — a
// deliberate deviation from the brief, ruled 2026-09-13 (see
// `.github/memory/decision.md`, Task 12). A telemetry service that refuses to
// serve because one cache file is truncated is worse than one that starts fresh
// and says so, so the file is **moved aside**, never deleted (the evidence is
// kept), and the service starts empty. The suffix names *why*: an older version
// and another provider wrote files that are not broken, and calling them
// `.corrupt` would be a lie the operator then has to disprove. Both the
// quarantine and the empty start are logged; a silent empty start is the failure
// mode this replaces.
app.addHook('onReady', async () => {
let restored: Snapshot | undefined;
try {
restored = await loadSnapshot(snapshotPath, provider);
} catch (error) {
const fault = error instanceof SnapshotError ? error.fault : 'corrupt';
const quarantined = `${snapshotPath}.${fault}`;
app.log.warn(
{err: error, snapshotPath, quarantined},
'semantic-log: the snapshot could not be used; moving it aside and starting empty',
);
try {
await rename(snapshotPath, quarantined);
} catch (moveError) {
// The quarantine itself failed (for example the directory is
// not writable). Booting still wins: refusing to start over a
// file we cannot move is the outcome this rule forbids. The
// failure is reported rather than swallowed, and it is
// escalated to `error` because the file will be overwritten by
// the next save — the evidence is then gone.
app.log.error(
{err: moveError, snapshotPath, quarantined},
'semantic-log: the snapshot could not be moved aside; it will be overwritten by the next save',
);
}
return;
}
if (restored === undefined) {
return;
}
registry.replaceAll(restored.entries);
// The observed calls survive the restart too, and like the registry they are
// restored *whole*: `restore` replaces the unions, so a snapshot is the state
// the service had, not an addition to whatever it observed since booting.
ledger.restore(restored.kinds);
});
}
/** Write the snapshot as it stands now: the registry, the unions, the identity. */
const writeSnapshot = (): Promise<void> =>
snapshotPath === undefined
? Promise.resolve()
: saveSnapshot(snapshotPath, {
provider,
entries: registry.list(),
kinds: ledger.unions(),
});
// One save per *change*, not per response: only the routes that mutate the
// registry call this. Saving from the handler rather than from `onResponse`
// is deliberate — `onResponse` is not awaited by `inject`, so a save there is
// fire-and-forget and cannot be asserted through the routes; here it is part
// of the request, and a failure is reported to the caller as a 500. A
// missing path makes this a no-op, so the mutation routes need no branch.
const persist: () => Promise<void> = writeSnapshot;
app.get('/health', async () => ({status: 'ok'}));
app.post('/events', async (request, reply) => {
// The anomalies *this* request raises, owned by this request. The ingest
// collects them into this array (its optional second argument), so a
// failure discards exactly this batch's anomalies and cannot touch a
// concurrent request's — the shared buffer with a failure-time `length = 0`
// that this replaced would wipe another in-flight request's anomalies.
const raised: Anomaly[] = [];
let result: IngestResult;
try {
result = await ingest(request.body, raised);
} catch (error) {
// The rejected batch's anomalies are discarded with `raised` when this
// scope unwinds: half a batch is not evidence of anything, and the
// discard is this request's alone — it cannot lose another request's.
return reply.code(400).send({error: (error as Error).message});
}
// Correlate this batch's anomalies against the lineage the ingest has
// just fed (R15). Nothing awaits between the ingest's completion and
// this step, so no anomaly is correlated twice or left behind. It sits
// *outside* the ingest `try` on purpose: only an unusable payload is a
// 400, and correlation reads state the ingest has already committed.
for (const incident of incidents.absorb(raised, lineage, registry, incidentWindowMs)) {
digest.publish('incident', {
incidentId: incident.id,
trace: incident.trace,
rootCause: incident.rootCause,
services: incident.services,
kinds: incident.kinds,
severity: incident.severity,
firstAt: incident.firstAt,
lastAt: incident.lastAt,
});
}
// Persist the batch before acknowledging it (PRD R4). Also outside the
// ingest `try`: the batch is valid, so a snapshot write that fails is the
// service's failure (Fastify answers 500), never a malformed payload.
await persist();
return reply.code(202).send(result);
});
// The correlated incidents (R15): one per trace and burst, with a candidate
// origin. A read of the accumulated store; the *deltas* travel on `/digest`,
// so a poller can follow new incidents without re-reading this list.
app.get('/incidents', async () => incidents.list());
app.get('/templates', async () => registry.list());
/**
* The anomalies the digest still retains, reduced to what a drift
* attribution reads (R16). Bounded exactly as the digest is (1000 entries):
* there is deliberately no "anomalies ever seen" store, because a facet read
* must not be able to grow the service's memory without bound.
*/
const retainedAnomalies = (): FacetAnomaly[] =>
digest
.read(0)
.filter(entry => entry.kind === 'anomaly')
.map(entry => entry.data as FacetAnomaly);
app.get('/templates/:ref', async (request, reply) => {
const {ref} = request.params as {ref: string};
const {facet} = request.query as {facet?: string};
const entry = registry.get(ref);
if (!entry) {
return reply.code(404).send({error: 'unknown template'});
}
// No facet is the raw entry, as before. An unknown facet is a 400, not a
// silent fallback to the raw entry: a silent fallback is how a consumer
// ends up reading data it did not ask for (see `facets.ts`).
if (facet === undefined) {
return entry;
}
if (!isFacet(facet)) {
return reply.code(400).send({error: `unknown facet ${facet}`, facets: FACETS});
}
// A `diagnostic` read is handed the anomalies the digest still retains,
// so `drifted` can be derived (see `facets.ts`); the other facets ignore
// it. The collection is bounded by the digest's own cap by construction.
return project(entry, facet, retainedAnomalies());
});
/**
* Retire a template (PRD R8): the writer behind the `template-retired`
* delta, and the reason a digest poller can rely on that kind arriving.
* Retention here is an **explicit decision**, not a wall-clock TTL — see
* `.github/memory/decision.md` (Task 12) for why a `lastSeen` sweep was
* rejected. Retiring does not delete: it stamps `retiredAt`, which is what
* the deploy diff's `removed` bucket reads (R14), and the entry stays
* readable.
*/
app.post('/templates/:ref/retire', async (request, reply) => {
const {ref} = request.params as {ref: string};
const entry = registry.get(ref);
if (!entry) {
return reply.code(404).send({error: 'unknown template'});
}
if (entry.retiredAt !== undefined) {
// Already retired, and re-retiring is not a *change*: publishing
// again would be a delta storm for a polling consumer. The same rule
// the incident store keeps (see `incidents.ts`).
return entry;
}
const retiredAt = Date.now();
registry.retire(ref, retiredAt);
digest.publish('template-retired', {ref: entry.ref, service: entry.service, retiredAt});
await persist();
return entry;
});
app.get('/records/:id', async (request, reply) => {
const {id} = request.params as {id: string};
const record = exemplars.recordOf(id);
if (!record) {
return reply.code(404).send({error: 'not retained'});
}
return record;
});
// What has been observed, per flow kind and per execution (R22/R23). Two reads of
// one index: the unions are what the kind-level diagram is drawn from, and the
// summaries are what a caller picks an execution out of. The summaries carry no
// per-record detail — a thousand executions' observations would be a response no
// one asked for — so a caller asks for the one execution it wants next.
app.get('/flows', async () => ({
unions: ledger.unions(),
executions: ledger.executions(),
evictions: ledger.evictions(),
truncations: ledger.truncations(),
}));
/**
* The sequence diagram of one flow kind, or of one execution.
*
* One route rather than two, because Fastify cannot register the same pattern with
* two parameter names — and the id's shape is a sound discriminator either way: a
* flow id is the caller-minted ULID (`isUlid`) and a kind is a dotted lowercase
* name, so no reference can be both. What the shape decides is which store is
* asked, and the two failures stay legible: an execution that was never retained
* (evicted, or never observed) is not a kind nobody has heard of, and telling a
* caller the wrong one sends it looking in the wrong place.
*/
app.get('/flows/:reference/diagram', async (request, reply) => {
const {reference} = request.params as {reference: string};
if (isUlid(reference)) {
const execution = ledger.executionOf(reference);
if (execution === undefined) {
return reply.code(404).send({
error: 'flow execution not retained',
retained: ledger.size(),
evictions: ledger.evictions(),
});
}
return {
id: execution.id,
kind: execution.kind,
closed: execution.closed,
observed: {
services: execution.services,
refs: execution.refs,
legs: execution.legs,
},
diagram: renderSequence(modelOfExecution(execution)),
};
}
const union = ledger.unionOf(reference);
if (union === undefined) {
return reply.code(404).send({error: 'unknown flow kind', kinds: ledger.kinds()});
}
return {
kind: union.kind,
observed: {
executions: union.executions,
services: union.services,
legs: union.legs,
},
diagram: renderSequence(modelOfUnion(union)),
};
});
// Semantic search (R14, R24). Templates AND retained records are ranked, in one list
// with a `kind` discriminator: a caller asking "where does this show up" wants both
// "which template is this like" and "which occurrence looked like this", and two
// endpoints would make it ask twice and merge by hand. A request without `q` is an
// empty result rather than an error: the route is a filter, and "search for nothing"
// is not a malformed request. The query text is embedded through the same cache as a
// template, under a `query:`-prefixed key, so a repeated query costs nothing and can
// never collide with a fingerprint.
app.get('/search', async request => {
const {q, limit} = request.query as {q?: string; limit?: string};
if (!q) {
return [];
}
const vector = await cache.vectorFor(`query:${q}`, q);
// The same finite-bound rule as `/diff`: `Number('abc')` is `NaN`, and
// `slice(0, NaN)` is `[]`, so a malformed limit would read as "no
// matches" rather than falling back to the default.
const cap = readNumber(limit, 10);
const templates = searchTemplates(registry, vector, cap).map(result => ({
kind: 'template' as const,
ref: result.ref,
score: Number(result.score.toFixed(6)),
service: result.entry.service,
signature: result.entry.signature,
count: result.entry.count,
}));
const records = searchRecords(exemplars, cache, vector, cap).map(result => ({
kind: 'record' as const,
record: result.id,
score: Number(result.score.toFixed(6)),
service: result.event.service,
signature: result.event.template ?? null,
msg: result.event.msg ?? null,
time: result.event.time,
}));
// Merged on the score. Ties need no tie-break of their own: each half arrives
// already totally ordered (templates by score and ref, records by score and id),
// and `sort` is stable, so a tie keeps the order the halves were concatenated in
// — templates first, then records. That is deterministic, and it is the same
// answer every time rather than whatever order the two lists happened to build.
return [...templates, ...records].sort((a, b) => b.score - a.score).slice(0, cap);
});
// The deploy diff (R14): "what is new since the last release?" answered
// without a query or a rule, from the registry (added/removed) and the
// flow-drift history (drifted). `drifted` is keyed by flow **kind**, never by
// template: a template's embedding is keyed by the fingerprint that
// identifies it, so it is constant and can never drift (R6c, D4). The
// response carries the effective `range` back, so a caller whose bounds fell
// back can see which window it actually asked about.
app.get('/diff', async request => {
const {from, to} = request.query as {from?: string; to?: string};
const range = {from: readNumber(from, 0), to: readNumber(to, Date.now())};
const diff = deployDiff(registry, range, flowDrift);
const summarise = (entry: TemplateEntry): object => ({
ref: entry.ref,
service: entry.service,
signature: entry.signature,
count: entry.count,
});
return {
range,
added: diff.added.map(summarise),
removed: diff.removed.map(summarise),
drifted: diff.drifted,
unchanged: diff.unchanged,
};
});
// The change stream (R8): a poller asks "what changed since cursor N" and
// gets the deltas, its own cursor back, and the bound's state. The cursor is
// echoed so a consumer never has to predict it, and `stats` is returned with
// every page so a consumer that fell behind can see that it did — the bound
// tells rather than silently truncates. A query the service cannot read as a
// number is treated as "from the beginning" / "no cap" rather than failing
// the poll: the stream is advisory, and a rubbish query string must not deny
// the deltas it can still have. All five `DigestKind`s are now published,
// four from this file (`template-added`, `anomaly`, `incident`,
// `exemplar-retained`) and `template-retired` from the retirement route
// above, so a poller may wait on any of them (see `digest.ts`).
app.get('/digest', async request => {
const {since, limit} = request.query as {since?: string; limit?: string};
const cursor = Number(since ?? 0);
const cap = limit ? Number(limit) : undefined;
return {
cursor: digest.latest(),
stats: digest.stats(),
// An unreadable `limit` is dropped (no cap) rather than passed on as
// `NaN`, which `slice` treats as 0 and would return an empty page.
entries: digest.read(
Number.isFinite(cursor) ? cursor : 0,
cap !== undefined && Number.isFinite(cap) ? cap : undefined,
),
};
});
return app;
}
|