The packed tier goes live inside GenerationStore:
- Two-tier reads: getDelta / readBeforeImage / readGenerationRecords
fall through live-tier → sealed segments (live-tier-wins: a crash
mid-fold leaves a duplicate representation, never a gap). Cold-open
seeds committedRanges from the segment manifest via interval merge —
packed generations resolve without their directories existing.
- repackHistory({timeBudgetMs, batchGenerations}): folds cold
generations (older than the newest 1024) oldest-first into sealed
segments, deleting per-generation directories only after segment +
manifest are durable. Public API + automatic time-bounded pass at
close() (before compaction, so reclaim can drop whole segments);
re-representation only — the sole history transform under the
archival profile. Early stop = consistent prefix, next pass resumes.
- compact(): packed generations reclaim logically in the loop and
physically at whole-segment boundaries via dropSegmentsBelow (the
frozen partial-segments-wait rule).
- generationDigest(g) (D8): deterministic content digest through g —
sealed-segment checksum chain + live-tier delta hashes; O(segments +
live window); RangeError out of range, GenerationCompactedError
below the horizon (a gate can never silently pin reclaimed history).
Four end-to-end pins: asOf answers byte-identical across fold + cold
reopen with folded dirs physically gone; repack+reclaim composition;
digest reopen-stability/divergence/loud-horizon; budget no-op+resume.
First stage of the co-frozen D1+D3+repacking unit: the format core,
self-contained under _generations/segments/.
- seg-<firstGen,20pad>.bgs: append-once packs of consecutive
generations (magic BGS1; frame = u32 len + u32 crc32c + msgpack
[generation, timestamp, delta, records, flags]; flags reserves
compressed-payload evolution without a format break). Sealed
segments are immutable — fold refuses overlap with sealed ranges.
- seg-<firstGen>.idx: DERIVED sidecar (per-generation frame offsets +
per-id generation postings + checksums); lost/corrupt sidecars
rebuild from their segment loudly; a damaged segment (frame CRC
mismatch) fails loudly, never serves wrong bytes.
- manifest.json: the one discovery path — open() reads it and never
lists the packed backlog (the scan-wedge class's cure); refuses a
newer manifest version rather than serving partial history.
- D3 semantics: dropSegmentsBelow reclaims WHOLE segments at
boundaries only and bumps compactedBelow durably; archival-profile
enforcement stays with the caller per the co-freeze.
- D8 rider: digestThroughPacked(g) — deterministic crc32c chain over
sealed-segment checksums (+ frame-level prefix mid-segment),
O(segments), reopen-stable.
Also fixes a cross-adapter contract bug the suite caught: memory
storage's deleteObjectFromPath ignored the raw-bytes store, so
deleteRawObject on a raw-bytes path (fact-log or segment files)
silently no-op'd — deletes now match filesystem unlink semantics.
Six pins. Wiring into GenerationStore (two-tier reads, the repacker,
cold-open manifest path) lands with the rest of the unit before its
release; cortex's fact-record/stamp shapes reconcile the sidecar
keying when they post.
Stage-2 D1 contract item (co-frozen): a fact scan may be slow, never
silent. batches() now races its FIRST pull against
SCANFACTS_FIRST_BATCH_MS (10s, exported; test-overridable) — a wedged
or unreadably slow store produces a loud abort naming the contract
instead of a consumer hanging indistinguishably from progress (the
production shape: a heal against a generations-backlogged brain
wedged silently on the first segment read).
Only the first pull is raced: the bound is time-to-first-batch (proof
the producer is alive), not per-batch pacing, and it runs only while
a pull is pending — consumer think-time between pulls never counts
against the producer (pinned).
Three pins: wedged-store loud failure within the bound, healthy scan
untouched end-to-end, slow-consumer immunity.
clearInterval() stops future heartbeat ticks but not one already in
flight: a straggler tick past its ownership guards could land its
atomic lock rewrite AFTER releaseWriterLock()'s unlink, re-creating
the lock file as a phantom that blocks the next writer until the
stale TTL expires (~60s) — the pool-eviction reopen case. Found as an
ENOENT heartbeat warning during benchmark teardown; the quiet variant
is the harmful one.
releaseWriterLock() now awaits the in-flight tick (tracked per tick,
self-clearing) before reading/unlinking, so a straggler's write always
lands BEFORE the unlink and gets removed with everything else. The
heartbeat's ENOENT is also now benign-by-contract (lock or directory
removed under us — the next acquire recreates it); other errors stay
loud.
Pin: straggler-past-guards simulation — lock file absent after close,
directory immediately claimable (fails on the undrained code).
flush() is durability work: it must cost what the current window's
deltas cost, never what the history backlog costs. Under adaptive
retention the byte budget derives from free memory, so bulk-load
pressure shrank the budget exactly at peak write volume and flush paid
actual reclaim inline — a production deployment measured single writes
blocked 25-191s behind reclaim-on-flush.
- flush() no longer calls autoCompactHistory(); close() is THE
auto-compaction site (already ran there; now alone).
- Every auto pass is time-bounded (CLOSE_COMPACTION_BUDGET_MS = 5s):
reclamation is oldest-first, so an early stop is a consistent prefix
and the next pass resumes. Explicit compactHistory() gains an
optional timeBudgetMs for caller-chosen maintenance windows.
- Documented trade stated where operators read: a long-lived writer
that never closes accumulates history until its next explicit
compactHistory() — predictable writes, explicit maintenance.
Pins: flush-never-reclaims + close-reclaims-durably (db-mvcc), bounded
pass stops-then-resumes as a consistent prefix (generationStore unit).
Four fixes from a consumer conformance report, one root disease — two
field-resolution regimes where there must be one:
- The delete/update aggregation hooks fed the engine a partial entity
view (type/service/data/metadata only), so a reserved-field groupBy
(subtype, visibility, ...) resolved to a nonexistent group on the way
down: counts drifted upward forever after deletes, and updates moving
an entity between reserved-field groups double-counted. The hooks now
pass the full-fidelity view via entityForAggFromRawRecord (every
reserved field top-level, mirroring the add path); the update sites
pass the full get() view instead of a hand-rolled subset.
- Aggregation source.where resolved fields only against the custom
metadata bag, so where on a reserved field silently matched nothing.
The matcher now resolves each filtered field through
resolveEntityField — the same single source of truth groupBy uses.
- removeMany() with no usable selector (bare array passed positionally,
empty params, ids: []) resolved successfully having deleted nothing.
All three now throw; the two legacy tests that pinned the silent
no-op as 'graceful' now pin the refusal.
- find() where keys accept both spellings: a metadata.-prefixed key
falls back to its flattened spelling when the prefixed one is not
indexed (metadata is flattened at index time). A literal nested custom
key named metadata still wins when indexed as spelled.
Five regression pins in aggregate-reserved-fields.test.ts (4 of 5 vary
red on the unfixed code).
Under default adaptive retention, every flush() recomputed total history
bytes by walking EVERY committed generation's delta — O(all generations)
with disk re-reads past the 4096-entry delta-cache bound. On a production
brain with 70,000+ accumulated generations this turned every write into a
full-tail scan (60-100s writes, escalating with history growth), even
though the free-RAM budget never tripped and nothing was ever reclaimed
(SELF-GENERATIONS-GROWTH).
historyBytes() now maintains a running total: seeded by one walk on first
use, then updated incrementally at both commit paths (+bytes) and the
compaction reclaim loop (−bytes), dropped on reopenAfterRestore. The
adaptive retention check on every flush is O(1). Invariant regression-
pinned: running total ≡ fresh walk through transact commits, single-op
group commits, and compaction.
New brain.historyStats() (exported HistoryStats): read-only generation
count / bytes / generation+timestamp range / horizon / retention mode /
effective budget — the one-call per-brain fleet audit for retention
exposure.
The post-import background deduplication pass (a merge-DELETE writer,
debounced ~5 minutes after import) had four lifecycle defects:
- enableDeduplication: false did not gate the background schedule — an
import that explicitly opted out could still have entities auto-removed
minutes later. The flag now gates both the inline and background passes.
- Each import() constructed its own coordinator-owned deduplicator, so the
debounce never spanned imports (N imports = N delete timers). The brain
now owns a single lazy instance (getBackgroundDeduplicator).
- close() never cancelled pending dedup; a delete pass could fire against
a closed brain. close() now cancels it first.
- The 5-minute timer held the process open (exit-hang class); now unref'd.
Four regression tests pin the contract (background-dedup-lifecycle);
import guides document that false disables both passes.
- New src/utils/osLimits.ts: reads RLIMIT_NOFILE (soft/hard, from
/proc/self/limits) and vm.max_map_count at open — once per process,
Linux-only, measurement-only — and warns loudly when either sits below the
pool-scale floors (soft NOFILE < 65536, max_map_count < 262144), with the
exact raise commands. On stock defaults the failure otherwise arrives as
EMFILE or a failed mmap deep inside an index open, long after the cause
stopped being visible. An unreadable limit produces NO warning — no
measurement, no claim — so non-Linux platforms stay silent.
- Exported for ops doors: checkOsLimits() returns the full OsLimitsReport;
floors exported as constants. Wired fire-and-forget in performInit after
storage init; the check can never affect open.
- Unit tests pin the parser (incl. 'unlimited'), the floor thresholds, the
null-never-warns rule, and the off-Linux silent path.
- acquireWriterLock now CLAIMS with an atomic create-exclusive write (O_EXCL)
inside a bounded retry loop: two processes racing an absent lock can never
both succeed (the old read-then-tmp-rename flow let the loser keep running
unlocked, silently). An EEXIST loser re-evaluates and either throws loudly
with the winner's details or performs a verified stale-takeover (re-read
before unlink so a lock that changed hands mid-deliberation is never
clobbered). Exhausted contention fails loudly instead of degrading into a
lockless open.
- BRAINY_WRITER_LOCKED passes through init() unwrapped: the error documents a
machine-readable contract (err.code + err.lockInfo with the holder's
pid/host/heartbeat), but init's blanket error wrapping stripped both,
leaving consumers a message to regex against.
- Two contract tests added: stale-foreign takeover installs OUR lock via the
atomic claim; the conflict error carries code + lockInfo at the public
init() surface.
- The transact apply budget scales with the batch — max(30s, opCount x 2s) —
or is exactly the caller's new TransactOptions.timeoutMs. A flat 30s cap
silently limited honest bulk work to ~15 operations on network-attached
storage (~2s/op measured in a production import) while looking generous for
small batches. Internal batch paths (removeMany chunks) get the same
scaling via the shared transactTimeoutBudget helper.
- TransactionTimeoutError now reports the operation it stopped at as i/N with
the operation's name, elapsed vs budgeted time, and states the batch rolled
back atomically and is retryable; context carries the fields
programmatically.
- The transact operational envelope is documented (optimistic-concurrency
guide): budget math, chunking with ifAbsent idempotency, when to reach for
addMany vs transact, and the precompute pattern (embedBatch + per-op
vector) that keeps inference out of the commit path.
- Embedding claims made honest per an end-to-end probe of the built dist:
batch and single embedding paths produce bit-identical vectors and a
vector-supplied add is fully searchable, but batch throughput on the
default WASM engine measures comparable to sequential (~160ms/text) — the
5-10x speedup claim in the addMany JSDoc is replaced with the measured
reality and the actual win (inference outside the budgeted write path).
- New public API: walks every canonical relationship record, queries the same
read path applications use (related() with all visibility tiers included),
and classifies every discrepancy into its failure family: missingFromReads
(records the read path omits — stale adjacency), danglingEndpoints
(endpoint entity gone — the partial-delete scar class), readOnlyVerbIds
(read-path edges with no canonical record — ghosts). Design-hidden
internal/system edges are counted separately so intentional hiding is never
misclassified as loss. Counts exact; example lists capped with an explicit
truncatedExamples flag; loud narration on incoherence; mutates nothing.
- Classification core lives in src/graph/graphAudit.ts behind injected seams
(canonical walks + the end-to-end read) so the discrepancy logic is testable
in isolation; brainy wires the seams to storage walks and related().
- getNounIdsWithPagination now refuses an undecodable resume cursor loudly —
the third and final pagination walk brought under the 8.5.2 cursor contract.
- Types exported: GraphAuditReport, GraphAuditDiscrepancy. Docs: the
inspection guide gains the audit -> repair -> audit verification ritual.
- Aggregation backfill walks build into a STAGING map and swap in atomically
on completion. A mid-walk failure drops the staging map — the previous live
state keeps serving, the aggregate stays flagged pending, and the storage
error surfaces to the failing query. Previously the walk wiped live state
before a scan that could throw, never cleared the pending flag on failure,
and re-ran a full walk on every subsequent query: a silent wipe/walk/throw
loop at the caller's retry rate.
- Failed walks are latched: retries within a 30s cooldown rethrow the recorded
error instantly instead of re-walking, so a tight caller-side retry loop
costs one loud error per query, never a full store walk per query.
- Persisted aggregation state is stamped with the store's committed generation
at flush; reopen adoption requires stamp equality. Stale state (unclean
shutdown) or over-counting state (a log truncation on a copied store pulled
the watermark back) triggers exactly one loud rescan, never a silent adopt.
- The backfill/adoption path narrates: adoption decisions, walk start/finish
with counts and duration, and failures all log by default.
- getNouns/getVerbs refuse a supplied-but-undecodable pagination cursor with a
loud error instead of silently restarting the walk at offset 0 (which
re-served page 1 forever to any while(hasMore) caller).
- The graph cold-load verb walk aborts loudly on a missing or non-advancing
cursor with hasMore=true.
- A versioned index provider whose generation is AHEAD of the committed
watermark (torn copy / crash-recovery truncation) is now named loudly at
open, alongside the existing behind-direction message.
- AggregationIndex: defineAggregate before init no longer forces a backfill.
init reconciles instead of clobbering: the app definition wins, persisted
state is adopted on hash match, a write landing pre-adoption forces an exact
rescan, and a successful state load clears the backfill flag. New ready()
settles every adoption decision before query paths consult backfill state.
- brainy: backfills are single-flight and batched. Concurrent queries share one
store walk and every pending aggregate fills from that same walk; the old
behavior let each concurrent query wipe the others' partial state and start
its own full walk, so a store under steady aggregate traffic never converged.
queryAggregate also waits for persisted definitions before deciding an
aggregate does not exist.
- paramValidation: recordQuery is telemetry-only. The duration-based cap
ratchet (x0.8 per recorded query while lifetime-average exceeded 1s, floored
at 1000 - below the documented 10000 auto floor, reported under a stale
basis label) is removed; the cap comes from its construction-time basis or
explicit overrides alone.
- storage: __aggregation_* and singleton system keys (brainy:entityIdMapper)
are recognized before the unknown-key warning fires; routing is unchanged.
- docs: find-limits cap-immutability note, aggregation reopen/adoption
semantics, RELEASES.md 8.5.1 entry.
setTimeout can fire a few ms early under load (timer coalescing) — a
10ms sleep measured 9ms and failed the >=10 assertion, tripping a
release gate. Sleep 25ms, assert >=20: the test verifies time is
MEASURED, not the OS timer's precision.
- storage.committedGeneration(): the committed watermark exposed on the
fact-scan capability, so a provider compares its stamp's
sourceGeneration against the store's truth without parsing the
private manifest format (the capability source now carries both the
live fact log and the committed closure).
- Pinned contract tests: fsync-before-ack (transact passes today; the
single-op path is pinned via it.fails as the documented target —
group commit must become LATENCY batching, ack waiting on the shared
fsync, never durability skipping; the pin flips red when that lands
so there is no cliff to discover) and scan-stability-under-rotation
(a scan snapshot yields exactly its facts — no gaps, duplicates, or
bleed-in — while segments rotate beneath it; the reclaim-during-scan
variant lands with fact compaction, which does not exist yet).
- FactLog accepts a rotateBytes option so tests exercise real rotation.
Three additive surfaces for native index providers, which hold only
`storage` and must never construct their own fact-log reader (the log's
open path is writer-side — it reconciles by truncating/rewriting):
- Fact-scan capability on the storage adapter (the getBinaryBlobPath
pattern): the host brain wires a closure over its LIVE fact log at
init; providers call storage.scanFacts()/factLogHeadGeneration()/
factSegmentPaths() and fall back to the enumeration walk on null.
Restore/reopen swaps the underlying log transparently.
- The family-stamp trio (readFamilyStamp/writeFamilyStamp/
verifyFamilyStamp + types) exported from @soulcraft/brainy/internals,
so 'one verifier' is literally one function shared with the native
side, never two synchronized copies.
- Rollup invariants widened to number|string: content fingerprints
(e.g. a per-tree SHA-256) are valid invariant values; strict equality
either way, and a type mismatch reads as incoherence, never a pass.
The canonical entity tree now carries a FAMILY STAMP
(_system/family-stamps/entity-tree.json, JSON — forensics stay
terminal-readable): which committed generation the tree reflects
(sourceGeneration) plus the rollup invariants (entity/relationship
counts) that verify a millions-of-files projection whole where per-file
checks cannot scale. Written at flush/close boundaries; open-time
coherence is a COMPARISON, not a walk:
- coherent / absent (legacy store) → silent
- behind → benign for the tree (it is written BY the commit; only the
stamp is stale after a crash between commit and flush) — refreshes at
the next flush
- incoherent (counts diverged at equal generation, or a stamp AHEAD of
the log head) → loud, names the failing invariant; repairIndex()'s
unconditional recount heals and RE-STAMPS so repair leaves a coherent
stamp behind
- a fault reading the stamp is UNVERIFIABLE — never conflated with
absence
One verifier (verifyFamilyStamp, exported) reads both member modes:
enumerated (exact byte size per member, bounded families) and rollup
(invariants, unbounded families). New exports: readFamilyStamp,
verifyFamilyStamp, ENTITY_TREE_STAMP_PATH, FamilyStamp, StampMembers,
StampVerdict.
Every committed generation now also appends a FACT — an after-image
commit record (what each touched entity/relationship became, or a
body-less tombstone for a removal) — to an append-only, crc32c-framed
segment log under _generations/facts/. The before-image history and the
canonical tree remain authoritative; the fact log gives consumers ONE
sequential, self-verifying stream (index heals, incremental replays)
in place of a per-entity directory walk.
- Wire format: positional msgpack facts [generation, timestamp, ops,
meta, blobHashes]; op = [kind u8, id bin16, record | nil tombstone];
32-byte segment header (magic, formatVersion, firstGeneration,
zeroed+verified reserved); length+crc32c frame per fact; zero-padded
segment names so lexicographic order == generation order; JSON
manifest with an atomic rename flip, manifest-first rotation.
- Commit protocol: facts append+fsync BEFORE the commit point inside
the existing durability window, so a crash can only leave the log
AHEAD of committed truth — open() truncates back (torn tails detected
by CRC). Absent generation = never committed; a scan can never see an
uncommitted fact. transact() facts are durable-on-return; single-op
facts ride the group-commit flush exactly like buffered history. A
fact-append failure fails the write, loudly — a silent gap would be a
lie a later replay discovers.
- New public surface: brain.scanFacts() (sequential batches with heal
telemetry: head/segments/approx up front, per-batch generation range
+ bytes + segment id, loud abort on gaps, summary cross-check) and
brain.factSegmentPaths() (immutable sealed segments for zero-copy
consumers; the mutable tail excluded). Exported types CommitFact,
FactOp, FactScanBatch, FactScanHandle.
- Storage: optional binary raw-byte primitives (appendRawBytes,
readRawBytes, writeRawBytes, rawByteSize) on StorageAdapter —
feature-detected; filesystem + memory adapters implement them; an
adapter without them hosts no fact log. Fact segments are byte-copied
(never hard-linked) into snapshots. The _generations/facts/ namespace
is registered as a protected family (rebuildable: false): no sweeper
or GC may delete under it.
- New crc32c (Castagnoli) utility with RFC known-answer tests.
Ports the fresh-brain probe that closed the type+subtype lens-drop
investigation into the permanent suite: a 7-pair corpus seeded through the
real write API (never restored bytes), every lens checked id-for-id against
an unfiltered canonical scan, warm AND after a cold reopen, plus the
type+subtype update-flip leg. Guards the under-inclusion class the egress
integrity guard structurally cannot catch.
A cross-directory rename added the new parent's Contains edge but skipped
removing the old one (a 'not critical' shortcut), leaving the moved entity
a child of BOTH directories: readdir(oldDir) kept listing it, re-creating
the old path showed the name twice, and tree-walking consumers saw the
file in two places. Reported live by a production deployment (37 stale
containment ghosts; blocks tree-sync integrations).
- rename() now removes the old parent's vfs containment edge(s) by edge
id resolved from the graph's own adjacency (the removal law: removal
never requires reading the removed thing), and a move to the root gets
its containment edge (previously skipped -> orphaned from readdir('/')).
- New vfs.repairContainment(): reconciles every VFS entity's containment
edges against canonical metadata.path — removes stale/duplicate vfs
edges, restores missing ones, never touches user knowledge edges. Wired
into repairIndex() so the one operator ritual heals existing ghosts.
- Regression: tests/integration/vfs-rename-containment.test.ts (7).
Persisted entity totals inflated permanently: a delete's count decrement
was sourced from re-reading the record being removed, so a null read
(replace race, or a ghost left by a pre-8.3.1 partial delete) silently
skipped the decrement while the paired add had counted — and
Math.max(persistedTotal, scanned) meant no disk cleanup could ever
lower the reported total. Reported by a production deployment with a
write→delete→re-create repro minting drift within hours.
- The caller's pre-delete read now rides through the entire delete path
(remove()/removeMany()/plan → DeleteNoun/DeleteVerb operations →
deleteNoun/deleteVerb → deleteNounMetadata/deleteVerbMetadata): a
null internal read falls back to the known prior record instead of
skipping the decrement. StorageAdapter.deleteNoun/deleteVerb gain an
optional priorMetadata parameter (additive).
- rebuildTypeCounts() rebuilt only the type-statistics arrays and
computed the total just to LOG it — the persisted scalar
(counts.json) survived every rebuild. One canonical walk now rebuilds
every counter rollup (scalar totals + per-type maps + type
statistics) and persists them; repairIndex() runs the recount
unconditionally, since counters can be inflated over clean shelves.
Two production-reported spine fixes:
- Canonical delete is a FULL removal: remove()/removeMany() deleted the
metadata leg but never the canonical vectors.json leg or the <id>/
directory, leaving ghost rows that inflated enumerated counts forever,
read as damage scars, and caused duplicate readdir entries for
re-created VFS paths (the stale-unpost path). The delete operation now
routes through storage.deleteNoun — both legs + the entity container —
with a full two-leg before-image rollback; deleteVerb symmetric; the
blind 'no metadata file' catch that masked real faults is gone. The
generation log still holds the delete's before-image, so asOf() history
is unchanged. repairIndex() gains a conservative, loud orphan-prune
sweep (containers with no metadata content leg) + count recompute for
stores damaged by earlier versions.
- Family-scoped migration gate: every read blocked on the whole-brain
migration lock even when the migrating index was irrelevant. The gate
now scopes to the index families a read actually consults — canonical
reads never wait; find() waits only on its query shape's families;
graph traversals wait only on the graph family. Writes keep the
conservative whole-brain wait. A read that needs the migrating family
still blocks (bounded) with the retryable MigrationInProgressError.
The canonical enumeration walk (getNounsWithPagination) hydrated each entity's
vector + metadata ONE-AT-A-TIME inside the shard loop — every enumeration paid
N x per-op-latency serially, and every index heal enumerates canonical, so this
was the dominant multiplier in the measured multi-minute heals (cortex heal-cost
decomposition).
- Hydration is now 16-way bounded-concurrency (matching the wave cor's native
rebuild uses): heal wall-clock becomes ~2xN/16 x per-op instead of N x per-op.
Order, cursor resume (skipped nouns still never read), filters, peek/hasMore
and totalCount are all preserved — pages are byte-identical to the serial walk.
- New getNounIdsWithPagination(): the id-only opt-out for callers that own their
IO schedule (an index heal). Unfiltered = ids straight from the shard paths,
ZERO per-entity reads; filtered hydrates metadata only (16-way). Same
cursor/offset/nextCursor contract, so it is page-compatible with the hydrating
walk.
Regression: pagination-parallel-hydration.test.ts pins paged==big-page identity,
ids==items order, zero-read id-only, and filter parity. 103 storage tests green.
The class-killer for the lost-main.dkann incident (ADR-004 §7). A provider can
declare a derived-index blob FAMILY (a set of members that are load-bearing
together, e.g. vector-base = main.dkann + main.slotmap + main.slotrev). Once
declared:
- deleteBinaryBlob / removeRawPrefix REFUSE to remove a declared member, throwing
the new ProtectedArtifactError — an in-process GC / sweeper is now INCAPABLE of
deleting a load-bearing index file (COLD != DEAD). Intentional retirement is an
explicit unregisterDerivedFamily(name) first.
- The declaration persists to _system/derived-artifacts.json, so protection
survives a reopen; clear() resets it with the rest of the derived footprint.
- checkDerivedFamiliesPresent() names any member missing on open (the catch for an
EXTERNAL deleter that bypasses the in-process refusal) → rebuild from canonical.
- Transients (*.tmp.*, *.rebuild-tmp, *.rotate-tmp) are never protected; a
namespace family protects a growing prefix (seg-*).
New StorageAdapter surface (optional): registerDerivedFamily / unregisterDerivedFamily
/ listDerivedFamilies + DerivedFamilyDeclaration; new exported errors
ProtectedArtifactError / DerivedArtifactMissingError. Enforcement is inert until a
provider declares a family (no regression). Cor declares its 6 families + does the
atomic set-swap in 3.0.15 (M2); brainy builds the contract now. 8 tests.
validateIndexConsistency() was blind to native providers — it only ran the JS
metadata index's own check, so a native provider whose manifest/segments/counts
had diverged still read as "healthy". Per ADR-004 §6 it now feature-detects and
aggregates each provider's optional validateInvariants() (a never-throwing,
<50ms self-report of its own cross-layer invariants), names any failing
invariant with its numbers in the recommendation, and exposes the per-provider
reports. A provider that violates the never-throw contract is surfaced as
unhealthy, never swallowed.
repairIndex() now reconciles NATIVE derived state from canonical too: it consults
each provider's invariants and calls rebuild() on any whose failing invariant
asks for heal:'rebuild' — the native counterpart of detectAndRepairCorruption().
New provider surface: optional validateInvariants(); new exported types
ProviderInvariantReport / InvariantResult / InvariantHeal. Additive, no break.
Cor implements the hook in 3.0.15 (M2); brainy builds against the shape now
(feature-detected, inert until a provider exposes it). 5 tests.
Closes the "dishonest readiness proxy" anti-pattern (Pattern A): size()>0 /
isInitialized were treated as "this index serves queries", but a cold native
index can load its COUNT before its SERVING structure, so a query returned a
silent [] indistinguishable from "no such data". A shared assessIndexReadiness()
now reads only the provider's honest isReady() signal (never size()), applied at
every site:
- Vector: a one-shot verifyVectorLive() guard on the semantic/proximity search
path (a pure semantic find({query}) has no filter, so nothing guarded it). It
prefers isReady(), else a known-vector self-match probe; self-heals via rebuild
or throws the new VectorIndexNotReadyError instead of a silent [].
- Graph: getVerbsBySource/ByTarget skip the fast path when the provider reports
not-ready (falling to the canonical shard scan), plus a one-shot probe that
self-heals a no-isReady provider whose adjacency did not cold-load.
- getIndexStatus(): folds in per-index honest `ready` (making `populated`
honest) + rebuildFailed/rebuildError/degradedIds, so a readiness probe never
200s a brain that is still warming up or degraded.
Unblocked by the native providers now reporting serving-truth (graph via
SSTable-residency readiness, vector via durableBaseLoadFailed). New export:
VectorIndexNotReadyError. getIndexStatus gains additive fields. No breaking API.
13 new tests; existing readiness guards green.
The Pass-1 hardening that makes loadBinaryBlob distinguish genuine absence
(ENOENT -> null) from a real fault (EIO/EACCES/... -> throw) was held out of
8.2.6 because the native accelerator's two column-store read sites still relied
on null-on-error. That accelerator release has now hardened those sites to
handle the throw (a faulted segment read marks the field unavailable and throws
a named error), so the fault-propagating form is restored and ships lockstep.
A present-but-unreadable index blob no longer masquerades as absent (which drove
needless rebuilds / empty reads). Adds the loadBinaryBlob leg to the blob
durability suite (absent -> null; present -> bytes; real fault -> throws).
The debounced materialization caught its failure with an empty `catch (() => {})`,
silently leaving the materialized Measurement entity stale; the aggregation-index
init() state-load failure was swallowed the same way, leaving aggregates reading
empty with no signal. Both are non-fatal (values rebuild via backfill-on-query),
but a silent stale/empty read violates loud-errors-never-quiet-losses. Both now
emit a loud warning naming the affected group / cause.
Two known-degraded states were recorded but never consulted by the read paths, so
a partial result looked authoritative:
- commitSingleOp returns `degraded` ids on an adopt-forward failed-rollback
recovery (the canonical record is durable but its derived-index entry may be
incomplete). persistSingleOp dropped that list on the floor — no health flag,
no read signal. It now records them in a queryable degraded set.
- A non-fatal index-rebuild failure at init (_indexRebuildFailed) was folded into
checkHealth()/validateIndexConsistency() but no read consulted it.
find() and get() now emit ONE loud warning per degraded window (reads still
return — canonical is the source of truth — but the caller is told results may be
partial and to run repairIndex()). Both degraded sources fold into the two health
surfaces, and repairIndex() reconciles from canonical and clears them.
With the unique per-writer temp suffix, a rename ENOENT can no longer mean "a
concurrent idempotent writer already renamed it" — nobody else holds this
writer's temp. It means our just-written temp vanished before the rename, so the
bytes did NOT land. The old code returned success on that ENOENT, acknowledging
a write that persisted nothing; the native provider mmaps these blobs, so a
phantom-acked blob is a silent-loss.
saveBinaryBlob now retries once with a fresh temp (self-healing a transient
external-sweeper/crash-cleanup race), and if the temp vanishes again it throws
loud rather than acknowledging a store-nothing write. Non-ENOENT rename faults
propagate verbatim as before. The unique-temp suffix already eliminated the
concurrent same-key collision that originally motivated the ENOENT shortcut.
The async group-commit flush that persists single-op generation history
swallowed persist failures as a bare warn: writes kept succeeding while their
before-images piled up in memory, never durable and unbounded, with no signal.
The generation store now accounts for every failed flush at one place
(flushPendingSingleOps), tolerates a transient blip (retry with capped
exponential backoff), and after PENDING_FLUSH_FAILURE_THRESHOLD consecutive
failures LATCHES a durability failure and refuses further single-op and transact
writes with a typed, exported PendingFlushDurabilityError — rather than promise a
durability it cannot deliver. Live canonical data is untouched; only the
immutable history is stuck. The latch self-heals: the moment a flush succeeds
(a retry, or an explicit flush()/close()) it lifts and writes resume.
Loud errors, never quiet losses.
clear() removed entities, indexes, system and _cas but left three top-level
trees on disk: _blobs (raw HNSW/LSM segment bytes and the native dkann index),
_id_mapper (the native shared mmap id-mapper), and _column_index (column-store
manifests). A cleared brain therefore re-read stale native state.
Worse, the column store splits its state across two of those trees — manifests
under _column_index/ and their segment bytes under _blobs/_column_index/ — so
removing one without the other stranded a manifest listing segments that no
longer exist, which the hardened segment-load path now (correctly) refuses with
ColumnSegmentLoadError. The three trees now fall together as a set, exactly as
_cas already does. locks/ is deliberately preserved: it is live coordination
state, not data.
ColumnStore silently skipped a manifest-listed segment it could not load: a
corrupt or missing segment dropped every one of its entities out of
filter/rangeQuery/sortTopK with no error, so an inconsistent index read as a
merely short result. It now throws a typed ColumnSegmentLoadError when a listed
segment yields undecodable or no bytes, and lets a genuine storage IO fault
propagate verbatim. Only a field with no manifest at all stays benign (nothing
was ever written for it).
baseStorage.getNoun_internal / getVerb_internal likewise caught every error and
returned null, reporting a present-but-unreadable entity as "not found". They
now return null only for genuine ENOENT-class absence (via isAbsentError) and
rethrow real faults and deserialize errors.
Pattern-B (blind catch) hardening: absence -> null, fault -> loud.
Write/index-spine hardening, first batch of Pass 1. Each fix restores an
invariant the surrounding code already intended; every one has a
fail-before/pass-after test.
- Pattern C, finding 5 (baseStorage): delete now decrements the user-facing
scalar total symmetrically — deleteNounMetadata was decrementing only the
per-type bucket, deleteVerbMetadata neither the bucket nor the scalar, so
getNounCount()/getVerbCount() inflated permanently (the stale scalar wins
pagination via Math.max and is persisted). Invariant now holds:
scalar total === Σ per-type across add/update/delete and reopen.
- Pattern A, finding 3 (graph/lsm/LSMTree): a partial SSTable-load failure no
longer publishes the manifest's full relationship count as healthy. Any
per-SSTable load failure throws after the batch, which resets to honest-empty
and lets the existing size()===0 self-heal rebuild run — size()/isHealthy()
can no longer lie about a partial load.
- Pattern B, finding 6 (hnsw/hnswIndex): deferred flush() no longer clears
dirty nodes whose connections failed to persist — failed nodes stay in the
retry set, and flush() throws HnswFlushError instead of returning a lying
node count. The immediate-mode first-noun saveHNSWSystem is un-swallowed, so
addItem() rejects rather than returning an id for a rootless index.
- Pattern B, finding 11 (part — storage reads): new shared isAbsentError()
helper (utils/errorClassification, ENOENT-only absence) applied to
loadBinaryBlob and readObjectFromPath — a real IO fault (EIO/EACCES/EMFILE)
now propagates loudly instead of masquerading as "absent", which had driven
needless rebuilds / empty reads (loadBinaryBlob feeds the native provider).
Regression: 78 green across the 3 new suites + db-mvcc, generationStore,
temporal-vfs, rollback-trapdoor, restore-nondestructive. Full gate runs before
the Pass-1 release (David-gated). Remaining Pass 1: finding 11 getNoun/getVerb
legs, finding 4 (ColumnStore), finding 8 (pending-flush), finding 10 (degraded),
finding 7 (clear). Pattern A guards (1,2,9) as a follow-up release.
A returned write must be immediately readable by id, metadata filter, vector
search, and graph traversal — no await, delay, or retry between the write
returning and the read. This holds by construction because every projection
(canonical + HNSW + metadata index + graph index) commits inside the write's
transaction; the generation counter is the {seq} a caller can pin. The test
locks that guarantee so index maintenance can never quietly move off the commit
path (which would turn a returned write into a not-yet-queryable one).
First brainy chapter of the write/index-spine hardening program.
When a transaction failed and its rollback ALSO failed to undo a canonical
write (retries exhausted), the old path logged, continued, threw
TransactionRollbackError, and set state='rolled_back' — while the record it
could not undo stayed durable on disk. The caller got an error implying the
write was undone; a read-back showed the record. A failed rollback had no
truthful response (the post-commit response lie).
Give a failed rollback a two-branch honest contract, decided by observation:
both commit paths already hold byte-identical before-images, so after a failed
rollback the store compares current canonical state to them and classifies each
touched record as reconciled, an additive orphan (present when it should be
gone), or a restorative loss (gone/wrong when it should have been restored).
- Adopt-forward: a single-op write whose only damage is a durably-present
orphan — the record the caller asked for — is committed forward (its
generation buffered) and returns success with a loud warning that the derived
index may be incomplete for that id until the next rebuild/repairIndex(). No
error, no double-write; the record is durable and get-able immediately.
- Fail loud + quarantine: a multi-op batch, or ANY restorative loss, throws the
new StoreInconsistentError naming every unreconciled record and its
disposition, and puts the brain into write-quarantine (reads keep working,
writes refused via assertWritable) until repairIndex() reconciles the derived
indexes against canonical and lifts it. The counter is not advanced, and
Transaction.rollback now reports state 'inconsistent' instead of the
'rolled_back' lie when any undo failed.
New: StoreInconsistentError + UnreconciledRecord exported from the root;
repairIndex() forces a rebuild and clears the quarantine. Regression
tests/integration/rollback-trapdoor.test.ts injects index-add-throws +
canonical-delete-undo-fails and pins adopt-forward (durable, get-able, not
quarantined), fail-loud (StoreInconsistentError + quarantine + reads work +
repairIndex lifts it), and the error's record naming.
restore() removed the entire live brain directory and then fs.cp'd the
snapshot in, so any copy failure left the store destroyed with only a partial
copy. The sharpest edge: fs.cp materialized the holes of sparse mmap blob
files, so a snapshot that fits on disk could balloon and ENOSPC mid-copy — the
recovery tool destroying the brain it was asked to recover.
Rewrite restoreFromDirectory as stage → verify → atomic swap:
- Copy the snapshot into a _restore_staging area BEFORE touching live data,
sparse-aware (copyTreeSparse/copyFileSparse skip all-zero 4 MiB chunks, so a
mostly-hole store restores at its true allocated size, not its apparent size).
- On any copy failure (ENOSPC included) remove only the half-written staging
area and throw — the live store is left exactly as it was.
- Only after the copy succeeds, fsync a completion marker naming the staged
entries, then swapStagedRestoreIn(): an idempotent per-entry
rm-old + rename-staged-in (same-filesystem renames that cannot ENOSPC),
removing stale live entries the snapshot lacks.
- completeInterruptedRestore(), wired into init() right after the root dir is
ensured, finishes a crash mid-swap FORWARD (resume a committed swap) or
discards an uncommitted staging area (live still authoritative).
_restore_staging is excluded from snapshots. persist() (hard-link snapshot)
was already safe and is unchanged; no public API change.
Regression (tests/integration/restore-nondestructive.test.ts): a forced copy
failure leaves live data fully intact and cleans staging; a normal restore
round-trips to the snapshot; an interrupted-but-committed restore completes on
reopen; an uncommitted staging area is discarded on open; the sparse copy is
byte-identical with allocated blocks far below apparent size.
A committed transact() reported success while its canonical entity writes were
still only in the OS page cache (tmp+rename, not fsync'd), even though the
generation counter and manifest were fsync'd. A hard kill in that window could
leave the durable counter ahead of the persisted entity bytes — phantom
progress for any generation-based consumer resuming from the counter. Reported
from a downstream migration's crash-lifecycle forensics.
Add an optional transaction durability barrier to GenerationStorage
(beginWriteBarrier / flushWriteBarrier). FileSystemStorage implements it:
writeObjectToPath records each successful canonical write and
deleteObjectFromPath records each delete's parent dir, between begin and flush;
flush fsyncs every recorded write (contents + rename dir entries, via
syncRawObjects) and the parent dir of every delete. commitTransaction opens the
barrier before running the planned operations and flushes it after, BEFORE
persisting the counter and manifest — so the batch's entire canonical footprint
is durable before the generation stamp advances. The barrier is optional: the
generation store calls it through optional chaining, so in-memory and
durable-per-call (cloud object-PUT) adapters treat it as a no-op.
The single-op group-commit path is unchanged (deferred durability is the
Model-B design that avoids a 3-5x per-write fsync regression), but its
durability contract is now documented explicitly on commitSingleOp: transact =
durable on return; single-op = durable at the next flush()/close(), with the
counter buffered alongside the data so a crash loses both together (never a
torn counter-ahead-of-state store).
Regression (tests/integration/transact-durability-barrier.test.ts): the entity
writes fsync in an earlier syncRawObjects batch than the manifest for single-op
and multi-op (add+relate) transactions; a precommit-rejected batch opens no
barrier and advances nothing; MemoryStorage exposes no barrier (optional-chain
no-op).
Transaction.execute() checked its time budget at the top of the operation
loop and threw TransactionTimeoutError from OUTSIDE the per-operation
try/catch, so a mid-flight timeout bypassed rollback entirely — only
per-operation failures rolled back. A bulk transact that crossed its 30s
budget mid-flight left the operations already applied to canonical storage
in place while the generation was never stamped: torn, generation-less
state. The generation-store commit path's abort cleanup explicitly assumes a
throw from execute() already restored the applied operations byte-identically
(it only discards the uncommitted staging directory), so the missing
rollback broke that invariant.
Give execute() a single rollback point: the operation loop is the sole
rollback-guarded region, and any error escaping it — an operation failure OR
a mid-flight timeout — rolls back every applied operation in reverse order,
then surfaces the original error (a rollback failure still supersedes it via
TransactionRollbackError). The per-exit-path rollback that let the timeout
throw slip past is gone; atomicity now holds by construction for every error
type. An aborted transaction leaves generation() unchanged and storage
byte-identical to its pre-transaction state.
Regression (tests/unit/transaction/timeout-rollback.test.ts): a mid-flight
timeout leaves an in-memory canonical store byte-identical with the tx in the
rolled_back terminal state; the operation-failure and rollback-failure paths
through the same single rollback point; and a clean transaction still commits.
transact() promises atomic forward references — add an entity and relate to
it in one batch — but the planner resolved relationship endpoint ints at
PLAN time, before the batch's add operations had applied. Asking the id
mapper about an entity that does not exist yet made the (correctly strict)
native mapper throw in EntityIdMapper.getOrAssign, so
transact([{op:'add', id:X}, {op:'relate', to:X}]) failed on native
deployments; the permissive JS mapper masked the bug and silently leaked an
id assignment whenever a batch was later rejected by a commit precondition
(normal control flow since the conditional-commit CAS landed).
Make endpoint resolution lazy: AddToGraphIndexOperation and
RemoveFromGraphIndexOperation take VerbEndpointInts — an eager
{sourceInt, targetInt} or a thunk evaluated when the operation EXECUTES,
mirroring the constructor's already-lazy generationFn. The four
transact-planner sites (relate, its bidirectional reverse edge, remove's
relationship cascade, unrelate) pass thunks, so resolution happens inside
the commit after same-batch adds have applied; the seven single-operation
sites keep eager objects (their endpoints provably pre-exist — relate
validates both, unrelate/updateRelation/remove read the existing verb).
The remove operation's rollback captures the ints resolved at execute so
its re-add uses the same mappings.
Byproduct fix: a rejected batch no longer pollutes the id mapper — the
regression suite pins this (mapper has no assignment for the phantom
entity after a precondition-rejected forward-ref batch), alongside the
exact reported shape, both-endpoints-in-batch, bidirectional,
add+relate+remove in one batch, and the split-transact control
(tests/integration/transact-forward-ref-graph.test.ts).
The temporal model had a hole exactly where files were concerned: every
entity write is an immutable generation with before-images, but VFS content
BYTES lived under an eager refCount GC left over from the pre-8.0 design —
unlink could physically destroy bytes that in-window history still
referenced, and overwrite never released the old hash at all (an unbounded
silent leak whose accidental byproduct was the only thing "preserving"
history). Reading the past could therefore return a stale field, a dangling
hash, or nothing, depending on luck.
Fix: blob reclamation becomes a HISTORY decision instead of a LIVENESS
decision. Each blob's metadata now carries historyRefCount alongside the
live refCount:
- The commit seam counts one history reference per persisted before-image
record carrying a content hash (commitTransaction staging and the
group-commit flush), recorded BEFORE the record-set persists and carried
in the generation delta (blobHashes — always present on new deltas, so
compaction only falls back to reading records for pre-contract
generations). An aborted transaction compensates best-effort.
- unlink/rmdir/overwrite drop ONLY the live reference (BlobStorage.delete →
release; overwrite finally releases the superseded hash — cancelling the
dedup increment on same-content rewrites and closing the leak), and only
AFTER the canonical mutation commits, so a failed delete can never leave a
live file whose bytes compaction might reclaim.
- History compaction is the ONE reclamation point: after deleting a
generation's record-set it releases that set's references and physically
reclaims any hash at zero live AND zero history references. Pins are
exempt automatically. Crash ordering is over-count-only in every path
(record before persist, release after delete), so a crash can leak until
the scrub recounts but can never reclaim bytes a retained generation
needs. scrubBlobHistoryRefCounts() restores exactness; existing stores get
a one-time marker-gated backfill on open, failing into leak-safe mode
(reclamation disabled) rather than guessing.
On top of the protected history, the temporal API the generational model
always implied:
- vfs.readFile(path, { asOf }) — the exact bytes as of a generation or Date,
materialized from the history (pinned view released so compaction is
never blocked by a read).
- vfs.history(path) — FileVersion[] ascending ({ generation, timestamp,
hash, size, mimeType? }), the newest entry being the live state.
- Overwrites now refresh the file entity's data/embedding text — semantic
search and the data field previously served the FIRST version's text
forever (the stale-field defect a consumer's incident recovery depended
on by luck).
Integration suite (temporal-vfs.test.ts): per-version exact reads +
history listing, leak-fix + history protection on overwrite, rm keeps bytes
readable, compaction reclaims past-window bytes and preserves in-window
(including the cross-file dedup case where an old file's history and a
newer file's removal share one hash), data freshness, and scrub exactness.
Subscribe once and receive one post-commit event per affected record for
EVERY canonical write, regardless of origin: direct calls, batch methods,
transact(), imports, and Virtual Filesystem writes all funnel through the
same commit points the feed is emitted from. This is the authoritative
in-process signal for live UIs, cache invalidation, and realtime sync layers
that forward it over their own transports.
Architecture — the post-commit dual of the ifRev precommit hook: mutation
methods hand lightweight event descriptors to the commit seam
(persistSingleOp / transact's plan), which stamps the committed
{generation, timestamp}, enriches entity deletes with the record's LAST
committed state from the commit's own before-images (free — Model B reads
them anyway; removeMany's id-only deletes gain full payloads this way), and
emits only after the commit succeeds — a losing CAS or rejected batch never
announces anything. Dispatch is a microtask FIFO after the mutex releases:
commit-ordered, a slow listener never delays a write, a throwing listener is
logged and isolated, and with no subscribers the write path constructs no
events at all. Single-point emission was chosen over per-method hooks
because the aggregation-hook pattern demonstrably drifted (relation ops and
removeMany were silently missing from it).
Coverage: add/update/remove (+ cascade unrelate per deleted relationship),
relate (both edges when bidirectional)/unrelate/updateRelation, per-item
events for addMany/updateMany/relateMany/removeMany, per-item events sharing
one generation for transact(), and transitively imports + VFS. clear() and
restore() — wholesale raw-state operations outside the per-record commit
path — emit a single store-level event meaning "refetch everything".
brain.close() drops all listeners.
New module src/events/changeFeed.ts (BrainyChangeEvent + ChangeFeed,
exported from the package root); guide docs/guides/reacting-to-changes.md.
Integration suite pins the contract: per-op payload fidelity, delete
last-state payloads, batch per-item emission, one-generation transact
batches, VFS-origin events, CAS-loser silence, ordering, listener isolation,
unsubscribe, and store-level events.
Sweeping the vestigial hnsw sharding machinery surfaced two real defects on
the counts-recovery path (the code that rebuilds totalNounCount/totalVerbCount
when counts.json is lost or corrupted — container restarts, partial copies):
- initializeCountsFromDisk counted files in entities/*/hnsw/ — directories the
8.0 write path never populates — so an established store recovered to ZERO
counts on real data: wrong getNounCount()/stats, a "New installation"-style
boot log, and a mis-sized rebuild-strategy decision at open. The scan now
walks the canonical entities/<kind>/<shard>/<id>/ tree (one entity per id
directory), with the same sampled type-distribution estimate as before.
- The sampler called getNounMetadata — whose ensureInitialized() re-enters
init() — from INSIDE init(), a latent deadlock reachable the moment the scan
found anything to sample. The sampled metadata files are now read directly
with fs (gz-transparent), no guarded accessors inside init.
The dead machinery itself (verified zero callers by reachability analysis
before deletion): the nine 7.x hnsw-layout entity/edge CRUD methods, the
sharding-depth prober + its depth-migration engine (unreachable — the probe
always returned null on 8.0 stores), an orphaned streaming verb paginator,
their path/scan helpers, and the now-unused type aliases and fields. The init
block keeps the truthful new-vs-established boot log introduced in 8.0.13,
now decided directly from the canonical layout. Net -1,138 lines; no public
API touched.
Adds a regression test: delete counts.json from a populated store, reopen,
counts recover exactly from the canonical tree.
The fast-follow to the ifRev CAS fix: the sweep for the same check-then-act
class found two more instances, both now closed with the same discipline —
the decision runs at the serialization point that guards the apply.
add({ifAbsent}) / add({upsert}): the absence check ran before the commit
mutex, so N concurrent same-id creates could all pass it and all write — the
second silently overwriting the first, violating ifAbsent's "return the
existing id WITHOUT writing" contract and upsert's merge-never-clobber
contract. The insert leg now carries a must-be-absent commit precondition
(the same conditional-commit primitive ifRev uses), verified under the commit
mutex against the authoritative before-image. A losing caller takes its
documented resolution instead of overwriting: ifAbsent returns the existing
id with zero writes; upsert merges into the now-existing entity via update()
(shared param mapping in upsertMergeParams so the planning-time branch and
the conflict retry can never drift), with a bounded retry if a concurrent
delete lands between the conflict and the merge. The planning-time checks
remain as fast-fails. transact()'s batch-level ifAbsent/upsert keep
planning-time semantics (documented, converges to a valid entity).
BlobStorage: write()'s dedup decision (exists → increment refCount, absent →
create at 1) and delete()'s decrement-then-remove were unserialized
read-modify-writes over blob-meta:<hash>. Concurrent writes of identical
content could lose references, so a later delete removed bytes another file
still referenced (data loss), or leaked unreferenced blobs. All
reference-count-bearing mutations now serialize through a per-hash
InMemoryMutex — distinct content never contends; incrementRefCount/
decrementRefCount are documented lock-assumed internals.
Both fixes are complete by construction in-process: storage enforces
single-writer-per-directory, so the process is the whole concurrency domain.
Tests (tests/integration/ifabsent-upsert-blob-concurrency.test.ts): 8-way
ifAbsent storm advances the store by exactly one generation with _rev 1;
8-way upsert storm on an absent id yields one create + seven merges
(_rev === 8); sequential ifAbsent/upsert semantics pinned unchanged; N
identical concurrent blob writes → refCount === N; the blob survives until
the true last reference drops; interleaved write/delete storm never loses a
landed reference.
N concurrent update({ ifRev }) calls carrying the same expected revision all
fulfilled (zero RevisionConflictErrors, last-writer-wins) — the check ran
before the commit mutex, so interleaved callers all passed it before any apply
landed. Sequential calls conflicted correctly, which hid the race. A production
cutover rehearsal caught it: 8 parallel ifRev-guarded ledger writes all
"succeeded" and 7 were silently lost. Advisory locks built on exactly-one-winner
semantics could hand the same lock to two workers.
Fix: conditional commit. commitSingleOp and commitTransaction accept an
optional precommit(beforeImages) precondition, invoked under the commit mutex
against the just-read authoritative before-images and before anything is staged
or applied — the per-record analogue of ifAtGeneration, which always ran there.
A throw aborts the commit atomically (generation reservation returned, zero
staging I/O in the transaction path). The store stays domain-ignorant; update()
and transact() supply the ifRev predicate:
- update(): the planning-time check remains as a fast-fail (avoids embedding
cost on an obviously stale expectation); the authoritative check re-verifies
the before-image's _rev and re-stamps the update's _rev from it, so the
counter is monotonic even for concurrent non-CAS updates (N plain updates
advance _rev by N). CAS against a concurrently-removed entity now throws
EntityNotFoundError instead of silently resurrecting it; plain updates keep
their last-writer-wins re-create semantics.
- transact(): per-op ifRev expectations registered during planning
(PlannedTransact.casUpdates) are re-verified in the batch's precommit; any
conflict rejects the whole batch before staging. Multiple updates to one
entity sequence through a running rev; ids the batch itself adds
(PlannedTransact.createdNouns — add resets the rev baseline even on
overwrite) keep their exact plan-time sequencing.
Regression suite (tests/integration/ifrev-concurrent-cas.test.ts): 8 parallel
same-rev updates → exactly 1 winner + 7 conflicts; same for transact batches;
_rev monotonicity; the documented read→CAS→retry ledger loop converging exactly
(8 workers, 8 decrements, 0 lost); sequential behavior unchanged; forward-ref
add+update in one batch unchanged.
The one-time 7→8 layout migration rescues the head branch's entities
(branches/<head>/entities/* → entities/*), then drained the entire
branches/<head>/ directory. Any durable state a 7.x engine wrote under the head
branch OUTSIDE entities/ — a branch-scoped index, blob area, or field registry —
would be silently deleted by that drain, the same failure class as the VFS
content blobs a 7.x store kept in _cow/.
Guard it: after the entity move, list what remains under branches/<head>/ and
exclude entities/. If anything survives, it is non-entity durable state, so
PRESERVE the branch (skip removeRawPrefix) and warn loudly with the leftover
keys — recoverable, not lost. A clean branch (only the moved entities) still
drains exactly as before. Adds a PARITY GUARD test to the 7→8 migration suite.
Every persisted 8.0 store logged "📁 New installation: using depth 1 sharding"
on every open — even brains holding thousands of entities — which is alarming to
read during a restart or incident. 8.0 stores nouns in the canonical
`entities/nouns/<shard>/<id>/vectors.json` layout (the path `saveNoun`/`getNouns`
use), but the legacy sharding probe `detectExistingShardingDepth()` inspects
`entities/nouns/hnsw` — a 7.x directory the 8.0 write path never populates (its
only writer, `saveNode`, is dead code with zero callers). So the probe returned
null for every 8.0 store and concluded "new".
Drive the new-vs-existing log from the layout the database actually reads and
writes: a new `hasCanonicalEntities()` checks for a real 2-hex shard directory
under `entities/nouns/`, and the known noun count short-circuits it. An
established store now logs "Using depth 1 sharding (N entities)"; only a genuinely
empty store reports a new installation. Behavior is otherwise unchanged — the
probe only ever set a log line, never triggered a rebuild or migration.
