feat(open-path): init never gates on the embedding model; open goes concurrent; slow opens narrate

A production restart storm measured 90,017ms for a single brain init vs
1,117ms quiet (~80x contention multiplier), traced to performInit() eagerly
awaiting the process-global WASM embedding engine before the VFS root even
existed. Every writer's open() queued on the one throttled model compile
(90-140s on throttled CPUs).

- VirtualFileSystem.doInitializeRoot() no longer embeds '/'. The root is
  system-tier plumbing nothing ever searches; when the default WASM engine
  is active it now gets an explicit all-zero placeholder vector
  (cosineDistance returns max distance for a zero vector, so it never ranks
  ahead of real content). deferEmbedding was considered and rejected: its
  landing path kicks the embed worker synchronously right after commit,
  which would still force the cold compile within milliseconds — just off
  the awaited path, not avoided. A registered native 'embeddings' provider
  (no cold-start cost, possibly a different dimension) still embeds the
  root for real, via the new Brainy.usesDefaultWasmEmbedder() seam.

- performInit()'s eager-embedding step now only STARTS the WASM engine warm
  in the background instead of awaiting it inline. embed()/embeddingManager
  already serialize concurrent callers on one shared init promise, so the
  first real embed() converges correctly either way; a failed warm narrates
  loudly instead of surfacing as a silent latency spike or an unhandled
  rejection. eagerEmbeddings: false still means no warm at all.

- FileSystemStorage.init() batches its ~8 independent bootstrap mkdirs
  (each creates its own full subtree via recursive:true, so none depend on
  the others existing) into one Promise.all. The restore-completion step
  and initializeCounts() stay strictly sequential — they have real order
  dependencies on rootDir and systemDir respectively.

- performInit() now times five phases (storage init / generation-store
  open+fold / index init+gate / VFS bootstrap / embedding-warm-started) and
  logs one warning with the per-phase breakdown when total open exceeds
  2000ms; silent otherwise.
This commit is contained in:
David Snelling 2026-08-25 10:09:45 -07:00
parent f8f64780b1
commit 96624f408c
5 changed files with 446 additions and 53 deletions

View file

@ -742,6 +742,15 @@ export class Brainy<T = any> implements BrainyInterface<T> {
private _pendingEmbedIds = new Set<string>()
private _embedWorkerFlight: Promise<void> | null = null
// OPEN-PATH FIX: the background embedding-engine warm kicked off (never
// awaited) by `performInit()` when `eagerEmbeddings` resolves true. Stored
// for observability only — `embed()`/`embeddingManager.embed()` already
// await the engine's OWN singleton init promise internally, so nothing
// needs to explicitly await this field for correctness. Never rejects on
// its own: a `.catch` narrates the failure and swallows it so a failed
// warm never surfaces as an unhandled rejection.
private _embeddingWarmPromise: Promise<void> | null = null
/** The stored log-authority switch, read once at open (default: tree). */
private _logAuthority: LogAuthorityRecord = { authority: 'tree' }
// A failed walk latches its error: retries within the cooldown rethrow it
@ -1079,6 +1088,23 @@ export class Brainy<T = any> implements BrainyInterface<T> {
configureLogger({ level: LogLevel.DEBUG }) // Enable verbose logging
}
// OPEN-PATH NARRATION: lightweight phase timing across the five
// named stretches of init — storage init / generation-store open+fold /
// index init+gate / VFS bootstrap / embedding-warm-started. Each
// `markPhase()` call records elapsed ms SINCE THE PREVIOUS checkpoint,
// so the buckets always sum to the pre-integration/warmOnOpen total.
// Silent under 2s; one `prodLog.warn` line naming every phase's ms
// above it, so the operator's next restart storm names its own slow
// phase instead of re-deriving it from a stack of raw timestamps.
const initStart = Date.now()
let lastPhaseCheckpoint = initStart
const phaseTimingsMs: Record<string, number> = {}
const markPhase = (name: string): void => {
const now = Date.now()
phaseTimingsMs[name] = now - lastPhaseCheckpoint
lastPhaseCheckpoint = now
}
try {
// Auto-detect and activate plugins BEFORE storage setup
// so plugin-provided storage factories (e.g., filesystem override from cor) are available
@ -1150,6 +1176,12 @@ export class Brainy<T = any> implements BrainyInterface<T> {
}
}
// PHASE 1 of 5 — "storage init": plugin/legacy-layout bootstrap,
// storage adapter construction+init, the OS-limit check, and the
// writer-lock claim, all folded into one bucket (everything above this
// line since performInit started).
markPhase('storage-init')
// 8.0 generational MVCC: open the record layer BEFORE any index is
// created or loaded. Crash recovery may rewrite canonical entity files
// (restoring before-images of an uncommitted transaction), and every
@ -1224,6 +1256,12 @@ export class Brainy<T = any> implements BrainyInterface<T> {
await this.createMigrationBackupIfNeeded()
}
// PHASE 2 of 5 — "generation-store open+fold": GenerationStore
// construction+open (crash-recovery replay/rollback fold), the
// derived-family registration, the fact-scan seam, the entity-tree
// stamp check, the brain-format handshake, and the pre-upgrade backup.
markPhase('generation-store-open-fold')
// Provider: embeddings (reassign embedder if plugin provides one)
const embeddingProvider = this.pluginRegistry.getProvider<EmbeddingFunction>('embeddings')
if (embeddingProvider) {
@ -1461,6 +1499,14 @@ export class Brainy<T = any> implements BrainyInterface<T> {
// Check for pending data migrations
await this.checkMigrations()
// PHASE 3 of 5 — "index init+gate": provider wiring (embeddings,
// cache, roaring, msgpack, sort:topK, distance), HNSW/metadata/graph
// index construction, the eager cold-load, id-resolver + connections-
// codec wiring, crash-recovery index rebuild, the replay-gap check,
// legacy VFS blob adoption, blob-history backfill, and the
// rebuildIndexesIfNeeded() gate + migration check.
markPhase('index-init-gate')
// Register shutdown hooks for graceful count flushing (once globally)
if (!Brainy.shutdownHooksRegisteredGlobally) {
this.registerShutdownHooks()
@ -1612,15 +1658,40 @@ export class Brainy<T = any> implements BrainyInterface<T> {
}
}
// Eager embedding initialization.
// PHASE 4 of 5 — "VFS bootstrap": shutdown-hook registration, blob
// storage init, the provider-summary log, flipping `initialized`,
// the migration-lock wait, VFS construction+init, flipping generation
// stamping active, the log-authority adopt/oracle check, and
// pending-embed crash recovery.
markPhase('vfs-bootstrap')
// Eager embedding initialization — BACKGROUND WARM (open-path fix).
//
// Adaptive default (8.0): the WASM embedding engine eagerly initializes
// Adaptive default (8.0): the WASM embedding engine eagerly WARMS
// during init() WHENEVER it is the active embedder — i.e. no native
// 'embeddings' provider has taken over — and the instance is a writer
// (not reader-mode) outside of unit tests. The WASM module (≈93MB with
// the embedded model) takes 90-140s to compile on throttled CPUs; paying
// that during boot rather than on the first embed()-driven call is the
// right default for the overwhelmingly common single-process server.
// the embedded model) takes 90-140s to compile on throttled CPUs.
//
// Historically this AWAITED `embeddingManager.init()` INLINE, so every
// writer's open() blocked on the compile — N concurrent opens all
// queued on the ONE process-global singleton (an ~80x contention
// multiplier measured in a production restart storm: 90,017ms busy vs
// 1,117ms quiet). The engine only needs to be ready before the FIRST
// REAL embed() call, not before init() returns, so this now only
// STARTS the warm and moves on — init() never waits for it.
//
// No double-await needed for correctness: `this.embed()` (~line 15420)
// delegates to `this.embedder`, which for the default engine is
// `embeddingManager.getEmbeddingFunction()` → `embeddingManager.embed()`
// (src/embeddings/EmbeddingManager.ts). That method calls `await
// this.init()` FIRST, and `init()` itself serializes every concurrent
// caller onto ONE shared `globalInitPromise` — so the first real
// embed() automatically waits for whichever finishes first: this
// background warm (if still running) or a fresh init() (if the warm
// hasn't reached this code yet, e.g. `eagerEmbeddings: false`).
// Verified by reading both call sites; `_embeddingWarmPromise` below
// is stored for observability only, never re-awaited by embed().
//
// Skipped automatically when:
// - a native 'embeddings' provider is registered (it owns embeddings;
@ -1628,8 +1699,8 @@ export class Brainy<T = any> implements BrainyInterface<T> {
// - reader-mode (readers don't embed — they query existing vectors),
// - unit-test mode (tests must stay fast and use the mock embedder).
//
// `eagerEmbeddings: false` is the explicit override to force lazy init
// (first-embed) even when this instance is the active embedder.
// `eagerEmbeddings: false` keeps meaning "no warm at all" — fully lazy,
// the first embed() call pays the full cost inline, same as before.
const isUnitTestMode = isDeterministicEmbedMode()
const eager = this.config.eagerEmbeddings ?? true
if (
@ -1638,9 +1709,45 @@ export class Brainy<T = any> implements BrainyInterface<T> {
this.config.mode !== 'reader' &&
!isUnitTestMode
) {
console.log('Eager embedding initialization enabled...')
await embeddingManager.init()
console.log('Embedding engine ready')
const warmStart = Date.now()
console.log('Background embedding-engine warm started (init() does not wait for it)...')
this._embeddingWarmPromise = embeddingManager
.init()
.then(() => {
prodLog.info(
`[Brainy] background embedding-engine warm complete in ${Date.now() - warmStart}ms`
)
})
.catch((err) => {
// Loud, never silent: a warm that fails to compile must be
// heard NOW, not discovered as a mystery latency spike on
// whichever request happens to trigger the first real embed().
// That first embed() call still retries init() itself (the
// singleton promise contract above) and surfaces its own typed
// error to its caller — this is the immediate, background echo.
prodLog.warn(
`[Brainy] background embedding-engine warm FAILED: ` +
`${(err as Error).message} — the first embed() call will retry ` +
`initialization and surface the error there`
)
})
}
// PHASE 5 of 5 — "embedding-warm-started": just the synchronous cost
// of kicking off the background warm above (the warm's own compile
// time is NOT included — that's the whole point of backgrounding it).
markPhase('embedding-warm-started')
{
const totalOpenMs = Date.now() - initStart
if (totalOpenMs > 2000) {
const phaseList = Object.entries(phaseTimingsMs)
.map(([name, ms]) => `${name}=${ms}ms`)
.join(', ')
prodLog.warn(
`[Brainy] slow open: ${totalOpenMs}ms total (${phaseList}) — see the ` +
`phase breakdown above to find which one to investigate first`
)
}
}
// Integration Hub initialization
@ -15695,6 +15802,25 @@ export class Brainy<T = any> implements BrainyInterface<T> {
return embeddingManager.isInitialized()
}
/**
* Whether the process-global WASM embedding engine (all-MiniLM-L6-v2,
* fixed 384-dim output, 93MB with the bundled model, 90-140s cold compile
* on throttled CPUs) is this instance's active embedder `false` when a
* plugin has replaced it via the `'embeddings'` provider key. A native
* provider has no such cold-start cost and may use a different output
* dimension, so it is never worth avoiding.
*
* Used by init-path bootstrap writes (the VFS root see
* `VirtualFileSystem.doInitializeRoot()`) to decide whether embedding a
* value during `init()` risks paying the WASM engine's cold compile.
*
* @returns true when the default WASM engine is active (no native
* `'embeddings'` provider registered).
*/
usesDefaultWasmEmbedder(): boolean {
return !this.pluginRegistry.hasProvider('embeddings')
}
/**
* Setup embedder
*/