- last_updated annotations containing ': ' ('family composition: ...',
'feature graph re-pinned: ...', 'REQ-5 added: ...') broke Gitea's
frontmatter parser in docs/architecture/README.md,
backends-and-dispatch.md, and requirements.md
- last_updated is now a bare date; the annotation lives in a quoted
notes field across all annotated architecture docs
- convention documented in .opencode/agents/architect.md: bare
last_updated date, quoted notes, quotes required for ': ' scalars
Verification: all docs/**/*.md frontmatter parsed with yaml.safe_load
(16/16 OK); cargo test, clippy --all-targets -D warnings, fmt --check pass
Full-tree review before task decomposition found five composition
defects (mechanisms specced correctly in isolation, composition
unruled) and a set of caller-facing gaps. ADR-012 rules each:
- Sweeps are never errors: aborts are GcAbortCause report data in
Ok(SweepReport) (ProtectFailed / SweeperLock / NoLivenessSources);
GcAborted retired from the error enum; direct-delete refusal is
the GcRefuse error covering the full protection set
- Pin token gains its wire shape: blobs/put response {token, digest};
blobs/have's token-renewal form (one digest + token); token
validity domain = the minting serving node, process-lifetime
mapping
- Fleet mode is an explicit constructor declaration (fleet: true),
never inferred from engine choice
- All fleet GC state hosts on the fleet's kv engine (postgres) — one
arbitration domain; large=pg-lo fleet nodes required onto the same
pg instance (composite predicate's new clause); large=local fleet
puts pin-row-first, publish-second
- Engine-state seam reduced: sqlite pin/sweep-lock bodies dropped
(dead machinery); non-SQL engines stage delete-window candidates
in-process; one-window-host rule per instance
- Facade clarifications: fall-through for all key-addressed ops,
kv-only put-time rejection, mem+local dual-tier valid, error-model
member/return-shape ruling (trait/facade family split), has ->
bool, PinState variants, fleet liveness-table registration form,
window executor = the next sweep
Alignment edits across all specs and ADR-005/008/009/010/011
(bracketed corrections per the established pattern); OQ-11 (pg-only-kv
feature graph) added to the parked index for auditability.
Verification: two independent review passes; all findings resolved;
verdict READY for task decomposition.
ADR-010 (new): the Backend trait's I/O seams pinned before any code —
'get' returns a crate-internal read cursor (kv engines materialize;
'local' pread loop; pg-lo lo_get windows), 'put' has two named forms
(whole-value kv / staged-put large); GC state is store-core-owned via
a crate-internal contract-tested engine-state seam on the SQL-backed
engines (backends never learn liveness — ADR-005 verbatim); the store
core holds the joint entry+pin transaction (corrects ADR-009 §2's
ownership statement); kv-tier fleet-validity rule added (mirror of
ADR-008 §3); ops pin-token TTL race resolved (the token IS the pin;
renewal rides have/re-put; pause-past-TTL falls to delete-then-recover);
feature graph pinned (pg-lo does not imply postgres).
ADR-011 (new): vocabulary pinned — tier (contract, exactly two) vs
engine (concrete impl, exactly five); store instance / node / fleet
split (fleet = pool-sharing, not node count — fixes requirements.md's
self-contradiction with REQ-2); mem demoted from 'backend'/'testing
tier' to the kv tier's third engine (contract-reference engine);
fs tier renamed 'large' including the feature name (last free moment
before code exists); constructor modes pinned (dual-tier default,
kv-only, mem-only) plus the composite fleet-validity predicate.
Consistency round across all specs and ADR-003/004/007/008/009: stale
pg-lo passages resolved per ADR-009's supersession; ADR-008 owner
'node id' → store-instance id; threshold wording corrected (conservative
edge, not midpoint, of the 128-256 KiB crossover zone); SweepReport
shape specced; mem classification unified; 008/009 ADR files renamed to
match the tier rename; README deferral-policy recap aligned (third
category = decided-but-sequenced work, not a parking kind); POC crate
list completed.
Verification: two independent architecture review rounds (the first
found 4 criticals — unpinned trait I/O shapes, missing fleet-state
seam, node/fleet self-contradiction, pin-token/TTL conflict — all
resolved; final round: zero criticals); all markdown links resolve;
ADR tables complete (11 ADRs).
The ADR-008 pg-lo admission POC ran in a standalone crate
(/workspace/alkblobs-pglo-poc): PgLoBackend over the ADR-003/008 trait
contract (including size), 10/10 exact-count sweep-outcome contract
tests, clippy/fmt clean; dockerized postgres:16-alpine on :15432,
POC #5 driver stack (tokio-postgres + deadpool) via SQL lo_* functions,
no new dependency.
Gate verdict: passed, with named deltas.
- Performance: durable put 60-65 MB/s at >=1 MiB, within 1.5x of — and
below 1 MiB beating — durable local fs on this fsync-slow disk;
cached gets 70-180 MB/s single-stream, ~0.7 GB/s aggregate over 16
readers (20-50x behind page-cache fs — the honest named delta)
- Contract: companion table is the list()/size()/CAS authority (never
the catalogs); stage-then-commit; GC-participating lo_unlink delete
- Handles: the tx-scoped descriptor is real but pool-compatible via
descriptorless lo_get(oid, off, len) windows — window gets keep
handle-acquire p99 at 1-6 ms under readers <= pool; held descriptor
is the fallback posture
- Vacuum: pg_largeobject pages churn-reused, never returned; tracked
by autovacuum; rel-size monitoring named as an ops requirement
- Crash/orphan: LO creation is transactional — kill/terminate
mid-write-tx leaves zero orphan pages; the only orphan class is a
committed LO bypassing the companion table (planted, reaped by the
~7 ms/oid sweep; committed content survives byte-exact)
- Harness lessons: lo_lseek is int4 — the 64 variants are the
>2 GiB discipline; shared-table parallel tests are unsound (per-test
CREATE DATABASE isolation)
Docs: new poc-pglo-findings.md; poc-pglo-spec.md status passed;
register OQ-BL-06 #7 marked passed; ADR-008 pg-lo bullet updated
(duplicate bullet removed) + backends-and-dispatch/open-questions
cross-references.
Verification: cargo test --release (10 passed), clippy -D warnings,
fmt --check in /workspace/alkblobs-pglo-poc.
- requirements.md (new): the three use cases pinned as REQ-1..4 with
the node/pool/fleet/engine vocabulary defined once — ends per-session
re-derivation of consumer facts. REQ-2 (replicator fleet over one
shared pg pool, incl. large-blob serving) is recorded as a planning
fact predating all POCs, on the operator's authority.
- ADR-008: Backend trait gains size(key) length probe (before any
backend ships data — ADR-002 one-way-door discipline); fleet GC
mechanism (DB-backed pin rows committed atomically with entries,
TTL+renewal semantics, liveness = embedder-owned table, protect
callback is single-node-only, advisory-locked single sweeper,
staged re-arbitrated delete window on both engines); fs tier becomes
engine-selectable (local default; pg-lo named candidate) with the
fleet locality contract (shared media or re-routing; mixed tiers are
a documented deployment invariant, not a constructor-provable one).
- poc-pglo-spec.md (new): POC #7 spec — pg Large Objects as the fs
tier's pg-lo engine; instruments, decision gate, registered in
phase-0.md OQ-BL-06.
- ADR-003/005 status amendments point to ADR-008's extensions; specs
ripple (backends/store-api/gc/ops/overview/README).
- open-questions.md: deferral-policy header gains the decisions-vs-
sequenced-work distinction; pg-lo's why-not-parked audit trail
recorded.
- research fixes: postgres POC renumbered #4->#5 to the canonical
register (phase-0 OQ-BL-06), redb cross-refs fixed, thinking-
artifact sentence in B1 replaced with the honest reading.
Verification: docs-only change; reference-integrity sweep across the
tree (ADR/REQ/POC refs resolve); architecture-reviewer pass on the
delta — original 3 criticals addressed, its follow-up (fleet liveness
form, pin TTL, staged-delete semantics, enforceability, shared-media
caveats) fixed in this commit.
Dissolve OQ-10's circular "standing offer" framing and record the
engine decision the POC evidence already supported:
- ADR-007 (new): kv tier ships sqlite (default) + postgres (feature
`postgres`, default-off), constructor-selected per node; records the
deployment economics (classical kv+fs+relational downstream stack
collapses to one relational engine + fs), the pg impl contract
(POC #5 B5 posture deltas), the CI sweep-safety gate under
--all-features, redb ruled out (POC #6), and posture defaults as
deferred cost in the ADR-006 pattern
- ADR-003: amended status (engine pin widened by ADR-007); kv entry
notes the widened engine set
- ADR-004: reconciliation note — the "two backends" count is tiers,
not engines
- backends-and-dispatch.md: tier≠engine distinction, pg engine
section, co-tenancy note (same-file consumer writes share sqlite's
single-writer ceiling), substitution-door precedent
- open-questions.md: OQ-10 resolved-by-ADR-007, with the
why-this-is-not-Schrödinger's-code record (a deployment running
alkblobs cannot pre-exist the pg engine — OQ-09 precedent); parked
index narrowed to OQ-07/OQ-08
- overview.md: layer map + dependency posture reflect the `postgres`
feature; README.md: ADR-007 in tables + corollary applications
- phase-0.md / pg+redb findings: promotion + supersession notes
Trigger framing survives only as deployment guidance (which engine a
node chooses — topology, not throughput), not as a gate on the crate's
own work.
Verified: cargo test, clippy --all-targets -D warnings, fmt --check.
Docs-only change (Phase 1 architecture); no implementation yet.
The backends spec read as if the Backend trait were a general storage
abstraction and 'two backends, exactly' banned other storage in a
deployment — so a downstream wanting sqlite manifests appeared forced
into implementing a third backend, the exact weld ADR-004 forbids.
- backends-and-dispatch.md: state plainly that the trait is a CAS-tier
contract (immutable digest-keyed entries, GC-shaped list/delete) and
never a host for manifest/index storage; consumer storage is
beside-the-pool. 'Two, exactly' counts trait impls shipped, not
storage per deployment. kv: tier name vs pinned sqlite engine made
explicit.
- ADR-004: add the explicit corollary under the consumer-seams point.
glm-5.3-flash