Cluster-capable state

[smunini]

Cluster-capable state — discussion & design

Status: discussion draft (no code yet)
Source: PR #180 review comment,
which asks for a document describing the areas of hfs that must be modified —
behind an environment-variable config switch — to use a unified,
cluster-capable job store, plus operator documentation for running hfs in a
cluster. Relates to #169,
#170,
#150, and the ROADMAP
"Clustered / multi-instance deployment" item.

This draft is the output of an exhaustive sweep of the workspace for state that
lives only in one process's memory and would break, diverge, or silently produce
wrong results if hfs ran as multiple instances behind a load balancer. It
covers the three areas already flagged (SoF export jobs, WebSocket Subscription
delivery, the JWT jti cache) plus everything else the sweep surfaced.

---

1. What "cluster-capable" means here

A single hfs process today holds a lot of correctness-bearing state on its own
heap: async job registries, WebSocket connections, replay caches, terminology
caches, and background reaper tasks. Behind a load balancer with ≥ 2 instances,
each request may land on a different instance than the one that created the
associated state. Four failure modes recur:

1. Cross-instance invisibility — a poll / cancel / download for something
   created on instance A returns 404 on instance B (SoF export jobs, reindex
   jobs, WebSocket binding tokens).
2. Restart / redeploy loss — a rolling deploy silently drops in-flight state
   with no durable record (all in-memory job registries, delivery retries).
3. Silent divergence — instances serve inconsistent answers because a cache
   was invalidated on one node only (HTS terminology caches), or a counter is
   maintained per node (Subscription eventNumber, failure counters).
4. Security regression — a one-time credential replayed against a second
   instance is accepted because the first instance's replay cache is not shared
   (JWT jti).

The boundary we are drawing: process-local ephemera (connection pools, tokio
runtimes, immutable config loaded from disk, per-request scratch) may stay in
memory. Anything that outlives a single request and must be observed by
another actor — a poller, a worker, a second instance — must be externalized
to shared infrastructure when clustered.

2. The good news — HFS already has two working precedents

We are not starting from zero. Two subsystems are already cluster-correct, and
they establish the two shapes every fix below should take.

2a. DB-backed job store with lease + fencing tokens (Bulk Data)

Bulk Data $export and $bulk-submit are fully cluster-safe. Their job state
(status, tenant ownership, per-type progress cursors, output-file rows) lives in
the primary database, and workers compete for jobs through an atomic
lease-and-fence protocol:

• trait BulkExportJobStore (crates/persistence/src/core/bulk_export_worker.rs:215)
  and trait BulkSubmitJobStore (crates/persistence/src/core/bulk_submit_worker.rs:300)
  compose job-state storage + a claim strategy + fenced worker storage.
• ExportClaimStrategy::claim_next (bulk_export_worker.rs:124) atomically
  claims one job (SELECT … FOR UPDATE SKIP LOCKED on Postgres; a process-local
  mutex on SQLite). Every worker mutation carries a worker_id + monotonic
  fencing_token; a zero-row guarded write returns LeaseError::LeaseLost and
  the worker aborts, so a zombie can never mutate a job another worker now owns.
• Workers are spawned per-instance (spawn_export_workers,
  crates/hfs/src/main.rs:1035); every replica runs a pool that competes for
  DB-leased jobs. HFS_BULK_EXPORT_DISABLE_LOCAL_WORKER turns a node API-only.

This lease/fencing store is the reference architecture for the unified job
store in §4.

2b. Pluggable shared backend selected by env var (auth jti)

The jti replay cache is already a trait with swappable backends chosen at
startup:

• trait JtiCache (crates/auth/src/jti/mod.rs:15) with InMemoryJtiCache
  (memory.rs:14), RedisJtiCache (redis.rs:10, atomic SET NX EX), and
  DisabledJtiCache.
• Selected by HFS_AUTH_JTI_BACKEND (crates/auth/src/config.rs:73, default
  "memory"), Redis URL from HFS_AUTH_REDIS_URL.

The trait boundary and env-var selection are exactly the pattern §4 generalizes.
The gap (see §5, class C) is only that the safe backend is opt-in.

A third, already-written-but-never-wired piece exists too: JwksCoordinator
(crates/auth/src/jwks/coordinator.rs:11) — a Redis leader-lock
(hfs:jwks:refresh_lock) plus shared key store for cluster JWKS refresh. It has
no call sites.

3. Shared substrate — what "unified" backs onto

Three externalization primitives cover every finding. The proposal does not
introduce three separate dependencies; it selects among what a deployment
already runs:

Primitive	Purpose	Backing options already present in HFS
Shared job store	durable job lifecycle + lease/fencing	primary DB (Postgres; Mongo where transactional) — §2a
Shared KV/replay cache	replay tokens, coordination locks, coherency epochs	Redis (redis feature) — §2b; or a DB table
Shared pub/sub fan-out	deliver an event that originated on any node to sockets/caches on every node	Redis pub/sub, Postgres LISTEN/NOTIFY, or a message bus

The governing principle: when clustered, the primary backend must itself be a
shared database (Postgres / Elasticsearch / MongoDB). SQLite — the zero-config
default (HFS_STORAGE_BACKEND=sqlite, crates/rest/src/config.rs:739) — is a
single-writer local file and cannot be clustered at all. Cluster guidance
starts there.

4. The unified cluster-capable job store

Generalize §2a into one ClusterJobStore seam that the in-memory async
subsystems plug into, selected by a single env var.

Trait shape (mirrors BulkExportJobStore, kind-agnostic):

trait ClusterJobStore {
    async fn enqueue(&self, tenant, kind: JobKind, payload: Value) -> JobId;
    async fn claim_next(&self, worker_id, lease) -> Option<ClaimedJob>;   // FOR UPDATE SKIP LOCKED
    async fn heartbeat(&self, lease) -> Result<Expiry, LeaseLost>;        // fencing-token guarded
    async fn get_status(&self, tenant, job_id) -> Option<JobStatus>;      // tenant-checked → 404
    async fn cancel(&self, tenant, job_id) -> bool;                       // cooperative, DB-visible
    async fn list_expired(&self, now) -> Vec<JobId>;                      // for a single reaper
}

• One jobs table in the shared DB, discriminated by JobKind
  (bulk-export, bulk-submit, sof-export, reindex), so the four
  subsystems share migrations, the claim query, lease/fence logic, tenant
  scoping, and the cleanup reaper.
• Backend selection via env var, following the jti precedent:
  HFS_JOB_STORE_BACKEND = memory (default; in-process DashMap, single
  instance) | database (shared, cluster-capable). A convenience master switch
  HFS_CLUSTER=true flips this and the other cluster-safe defaults
  (§6) at once and fails fast if the primary backend is SQLite.
• Existing seams to retarget: the SoF path already declares its trait
  (ExportJobController, crates/rest/src/export/controller.rs:196) and a
  reserved-but-unread selector (HFS_EXPORT_CONTROLLER,
  crates/rest/src/config.rs:805). The database backend implements
  ExportJobController over ClusterJobStore and is wired where
  InMemoryController is unconditionally constructed today
  (crates/rest/src/lib.rs:517). Bulk export/submit already are this pattern
  and can migrate onto the shared table opportunistically or stay as-is.

5. Inventory of affected areas

Grouped by the fix class they need. Each item: location · what breaks · severity.

Class A — in-memory async job registries → move to the unified job store (§4)

• A1 · SoF async export ([Database-backed SQL-on-FHIR export job state #169]) — CLUSTER-BREAKING.
  InMemoryController (crates/rest/src/export/in_memory.rs:55) holds
  jobs: DashMap<JobId, JobStatus> (:56) and job_tenants: DashMap (:59);
  submit() runs each job in a detached tokio::spawn (:168) under a
  per-instance Semaphore (so N instances run N× the intended concurrency);
  a per-instance TTL reaper (spawn_cleanup:117 / reap_expired:323) sees only
  its own jobs. The 202/status URL, cancellation, download, and the completion
  manifest (JobStatus::Completed { files, .. }) all live only in one heap. A
  poll/cancel/download on another instance → 404; restart loses everything.
  The single clear code defect; the whole $…-export async surface is
  single-instance today. Severity: functional breakage / result loss.
• A2 · Persistence reindex jobs — degradation.
  ReindexManager (crates/persistence/src/search/reindex.rs:363) keeps
  jobs: RwLock<HashMap<String, ReindexProgress>> + cancel_channels; start()
  mints a UUID and tokio::spawns the work (:406). Status poll / cancel on
  another instance sees no such job; restart orphans the reindex. Admin-triggered,
  no corruption. Severity: degradation.

Class B — node-local connection registries & fan-out → shared pub/sub + shared state (Subscriptions, [#170])

Systemic: subscription reaction is per-instance, fire-and-forget, in-memory.
emit_subscription_event (crates/rest/src/handlers/subscription_event.rs:47)
tokio::spawns on_resource_event on whichever instance served the write,
against that process's registries. Subscriptions/topics are loaded only
reactively (nothing reads them from the DB at startup). So a Subscription
created on A produces zero notifications for a matching write served by B,
and all in-memory state vanishes on restart.

• B1 · WebSocket client registry — CRITICAL.
  WebSocketManager.clients (crates/subscriptions/src/channels/ws_manager.rs:19),
  a DashMap<(tenant, subscription_id), Vec<WsClientSender>> holding the
  mpsc-sender half of a live socket. Only the instance terminating the socket can
  deliver; WebSocketChannel::dispatch (channels/websocket.rs:45) hits the
  local manager and returns Success even with 0 local clients, so loss is
  invisible. Sockets are inherently node-local → needs a pub/sub fan-out where
  the event is broadcast to all instances and each delivers to its own sockets.
• B2 · WebSocket binding-token store — HIGH.
  WsBindingTokenManager.tokens (channels/ws_token.rs:24), single-use ~30 s
  tokens from $get-ws-binding-token. The token is minted on one connection and
  redeemed on a separate WS-upgrade connection the LB may route elsewhere →
  every cluster WS bind fails without sticky routing. Fix: shared KV (Redis TTL)
  or a signed/stateless (HMAC/JWT) token.
• B3 · Subscription & topic registries — CRITICAL/HIGH.
  SubscriptionManager.subscriptions (crates/subscriptions/src/manager/mod.rs:137),
  InMemoryTopicRegistry.topics (crates/subscriptions/src/topics/mod.rs:94),
  and SubscriptionEngine.topic_resource_index (engine/mod.rs:37) are all
  in-memory with no startup load. An instance that never saw the topic write
  rejects register with TopicNotFound. Fix: DB-backed load (the resources
  already persist) + startup reconciliation.
• B4 · Per-subscription counters — HIGH/MEDIUM.
  events_since_start → Subscription.status.notificationEvent[].eventNumber
  (manager/mod.rs:124, built in notification/builder.rs:172) and
  consecutive_failures driving error/off transitions (manager/mod.rs:127,
  engine/mod.rs:854) are per-instance. Result: non-monotonic/duplicated
  eventNumber across the cluster (breaks subscriber gap-detection); a dead
  endpoint whose failures scatter across nodes never reaches the off threshold.
  Fix: shared atomic counters (DB row / Redis INCR).
• B5 · Delivery retry loop — MEDIUM.
  dispatch_with_retry (engine/mod.rs:722) keeps the retry "queue" on a
  fire-and-forget task's stack (backoff up to 60 s × 10). A redeploy drops all
  pending retries with no record. Fix: a durable delivery outbox table
  processed by workers — i.e. the §2a lease pattern again.
• Cluster-safe already: rest-hook, messaging, and email channels
  (channels/{rest_hook,messaging,email}.rs) are stateless push-to-external —
  any instance delivers identically. No heartbeat/reaper loop exists yet
  (heartbeat_check_interval, config.rs:33, is unused) — but if one is added
  it must be lease-guarded, not a plain per-instance interval.

Class C — shared caches / replay with local-only invalidation → shared store or cross-instance invalidation

• C1 · JWT jti replay cache — SECURITY HOLE (default-on).
  InMemoryJtiCache (crates/auth/src/jti/memory.rs:14) is the default
  (config.rs:73; built in crates/hfs/src/main.rs:617). A one-time client
  assertion accepted on A and replayed to B is honored again — B's cache never
  saw the jti. Blast radius scales with instance count. Secondary bug even
  single-instance: it ignores the token's expires_at and uses a flat 1 h TTL
  (memory.rs:26/43), so a longer-lived token is replayable after eviction. Fix
  already exists: RedisJtiCache (redis.rs:10). Recommend: when clustered,
  make a shared backend mandatory / fail-closed. Severity: security.

• C2 · JWKS refresh — MEDIUM (dead code ready to wire).
  JwksCache (crates/auth/src/jwks/cache.rs:16) is per-instance — functionally
  fine (every node fetches the same public keys) but each node independently
  hammers the IdP on a kid miss. JwksCoordinator (jwks/coordinator.rs:11)
  was built to fix this (Redis leader-lock + shared key store) and is never
  wired. Fix: wire it under the cluster switch.

• C3 · HTS terminology response caches — HIGH (silent wrong clinical answers).
  Per-process, no TTL, invalidated only on the instance that received the
  write/import:

  • AppState caches — expand_cache, not_found_urls, *_validate_code_*,
    expand_handler_cache, lookup_handler_cache, etc.
    (crates/hts/src/state.rs:190); cleared only by local clear_expand_cache
    (state.rs:280, called from import_bundle.rs / crud.rs).
  • Postgres backend caches — inline_compose_cache, lookup_response_cache,
    cs_resolved_meta_cache, subsumes_response_cache, translate_response_cache
    (crates/hts/src/backends/postgres/mod.rs:97; cleared :172), and
    process-global OnceLock statics CLOSURE_COUNT_CACHE /
    CLOSURE_PREFIX_CACHE (postgres/value_set.rs:37,97).

  A CodeSystem/ValueSet/ConceptMap update on instance A leaves B serving stale
  $expand / $validate-code / $lookup / $translate / $subsumes results
  indefinitely — a correctness hazard in a clinical system. Fix: cross-instance
  invalidation (PG LISTEN/NOTIFY on terminology writes, or a shared
  terminology-version epoch keyed into the caches, or short TTLs). The SQLite-side
  equivalents (sqlite/value_set.rs:73, code_system.rs:119) are Low — a
  cluster can't share a SQLite file anyway.

Class D — once-per-instance background tasks → leader-election / leasing

• D1 · HTS bootstrap sync — MEDIUM.
  bootstrap_sync runs on every boot (crates/hts/src/main.rs:78,134,187),
  deduped by the shared bootstrap_imports ledger — but the check-then-import-
  then-record sequence has no lock. N instances cold-booting together (rolling
  deploy, autoscale) each see "not imported" and import the same heavy
  SNOMED/LOINC/RxNorm/ICD-10 file concurrently. Idempotent upserts keep the end
  state correct, but you pay N× cost + write contention. Fix: pg_advisory_lock
  around bootstrap, or leader-election.
• D2 · Bulk export/submit cleanup reapers — LOW (already tolerable).
  Unleased per-instance reapers (crates/hfs/src/main.rs:1063,1244) scan the
  shared DB and race on deletes, but deletes are idempotent → duplicated work
  only. Optional: gate behind a single-owner lease for tidiness.

Class E — durability queues

• E1 · Composite async sync queue — HIGH (silent search divergence).
  In async mode, SyncManager buffers secondary-backend (e.g. Elasticsearch)
  propagation in an in-process mpsc::channel(1000) with status in
  RwLock<HashMap> (crates/persistence/src/composite/sync.rs:157,160). A
  crash/redeploy with events still queued permanently loses those secondary
  writes → the search index silently diverges from the primary, recoverable only
  by a full reindex. Both a crash-durability and a cluster issue. Fix: a durable
  outbox (DB table) drained by workers.

Class F — configuration-level cluster caveats (no new code, must be documented)

• F1 · SQLite default cannot cluster. HFS_STORAGE_BACKEND=sqlite
  (crates/rest/src/config.rs:739) is single-writer local file. Clustering ⇒
  Postgres / Elasticsearch / MongoDB. HTS SQLite backend: same.
• F2 · Bulk output stores default to node-local disk. LocalFsOutputStore
  is the default for both bulk subsystems (crates/hfs/src/main.rs:955,1124); a
  download routed to a different node than the writer 404s. Set
  HFS_BULK_EXPORT_OUTPUT_BACKEND=s3 / HFS_BULK_SUBMIT_OUTPUT_BACKEND=s3 for
  shared, presigned output.
• F3 · Sidecar bulk_export.db under Mongo/S3 primary.
  When the primary backend can't host transactional job state (MongoDB, S3),
  bulk-export job state falls back to a per-process local SQLite file
  (build_embedded_job_store, crates/hfs/src/main.rs:886, worst case a
  per-PID temp path). Under those backends in a cluster, jobs are invisible
  across instances — the SoF failure mode. Fine for Postgres primary (shares the
  primary DB).
• F4 · Audit file sink is node-local. FileSink
  (crates/audit/src/sinks/file.rs:19) writes NDJSON to local (often ephemeral)
  disk → a fragmented, restart-lossy audit trail. Use DatabaseSink or
  CloudWatchLogsSink (both write immediately to shared infra) when clustered.
• F5 · Unconditional version-ID increment race. New version_id =
  current.parse::<u64>() + 1 is a read-then-write with no lock on the
  unconditional update/delete path (crates/persistence/src/backends/postgres/storage.rs:291,396;
  mongodb/storage.rs:110). Two concurrent writers (more likely across
  instances) can both assign N+1, colliding ETags / losing a history version. The
  conditional path is safe (guards on expected_version,
  postgres/storage.rs:946). Fix: make the increment atomic (… SET version = version+1 … RETURNING) or require If-Match. Severity: LOW-MEDIUM.

Confirmed benign (checked and cleared — for coverage confidence)

Server-assigned resource IDs are Uuid::new_v4() (collision-free, no shared
counter). The AtomicU64/AtomicUsize counters found (sqlite/mod.rs:484,
persistence sqlite/backend.rs:23, sof/remote_fetch.rs:62) name ephemeral
in-memory test DBs or count within a single operation. Per-backend
search_registry is immutable config loaded identically per instance. Composite
HealthMonitor, per-tenant pool LRU bookkeeping, fhirpath tracer/runtime
statics, outbound OAuth token cache (crates/rest/src/bulk_submit_oauth.rs:30),
and CDS Hooks types are per-instance-local or read-only and safe. Audit recording
is fire-and-forget with UUID IDs (no per-instance sequence), only a minor
crash-durability window.

6. Environment-variable configuration surface

One master switch plus per-subsystem overrides, all following the existing
HFS_* conventions.

Variable	Values	Default	Effect
HFS_CLUSTER	true / false	false	Master switch: selects cluster-safe backends below and fails fast if the primary backend is SQLite or a required shared dependency is unset.
HFS_JOB_STORE_BACKEND	memory / database	memory (database when HFS_CLUSTER)	Unified job store (§4) — SoF export, reindex; bulk export/submit already DB-backed.
HFS_AUTH_JTI_BACKEND	memory / redis / disabled	memory	Must be shared (redis) when clustered (C1). Consider fail-closed under HFS_CLUSTER. (exists)
HFS_AUTH_REDIS_URL	URL	—	Redis for jti + JWKS coordinator. (exists)
HFS_SUBSCRIPTIONS_FANOUT	memory / redis / nats / pg-notify	memory	Shared pub/sub for WS delivery + WS binding tokens + counters (class B). (new)
HFS_TERMINOLOGY_CACHE_INVALIDATION	local / pg-notify / redis	local	Cross-instance HTS cache invalidation (C3). (new)
HFS_BULK_EXPORT_OUTPUT_BACKEND	local-fs / s3	local-fs	Set s3 when clustered (F2). (exists)
HFS_BULK_SUBMIT_OUTPUT_BACKEND	local-fs / s3	local-fs	Set s3 when clustered (F2). (exists)
HFS_AUDIT_SINK	database / file / cloudwatch / …	varies	Avoid file when clustered (F4). (exists)

Design intent: an operator sets HFS_CLUSTER=true, points at a shared primary
DB, and provides Redis (or accepts DB-notify equivalents); the switch turns the
individual backends to their cluster-safe variants and refuses to boot on an
unsafe combination rather than silently degrading.

7. Operator documentation deliverable

Ship a "Running HFS in a cluster" page (book chapter + skill note) covering:

1. Prerequisites — a shared primary backend (Postgres/ES/Mongo, not
   SQLite), shared object storage (S3) for bulk output, and Redis (or the
   pg-notify alternatives) for replay/fan-out/invalidation.
2. The one-switch setup — HFS_CLUSTER=true + the shared-infra URLs, and the
   fail-fast checks it enforces.
3. Per-subsystem behavior matrix — what each $operation / channel does with
   1 vs N instances, and which env var makes it cluster-safe.
4. Known single-instance-only features until their fix lands — with the same
   ⚠️ pinning guidance Database-backed SQL-on-FHIR export job state #169/Cluster-aware Subscription notification delivery #170 already carry (pin to one instance or use sticky
   sessions).
5. Load-balancer notes — WebSocket sticky sessions are not sufficient for
   Subscription delivery (the triggering event originates on an arbitrary node);
   pub/sub fan-out is required.

8. Suggested phasing

Each phase is independently shippable and leaves the tree green.

• Phase 0 — framing & guardrails. Add HFS_CLUSTER with fail-fast validation
  (reject SQLite primary, memory jti, local-fs bulk output, file audit)
  and publish the §7 operator doc describing current single-instance limits. No
  behavior change to the happy path. (Highest value / lowest risk — documents
  and enforces the boundary immediately.)
• Phase 1 — unified job store + SoF export (Database-backed SQL-on-FHIR export job state #169). Land ClusterJobStore
  (generalized from BulkExportJobStore), implement ExportJobController over
  it, wire HFS_JOB_STORE_BACKEND/HFS_EXPORT_CONTROLLER. Fold reindex jobs
  (A2) onto the same table.
• Phase 2 — auth hardening. Make jti shared-mandatory under HFS_CLUSTER
  (C1); wire JwksCoordinator (C2). Small, security-first, mostly existing code.
• Phase 3 — Subscriptions cluster delivery (Cluster-aware Subscription notification delivery #170). DB-backed
  subscription/topic load + startup reconciliation (B3), shared pub/sub fan-out
  (B1), shared/stateless WS binding tokens (B2), shared counters (B4), durable
  delivery outbox (B5).
• Phase 4 — HTS cache coherency (C3) + bootstrap lock (D1) and
  composite durable sync outbox (E1).
• Continuous — config caveats (F1–F5) documented in Phase 0, code fixes
  (F5 atomic version increment) folded in where cheap.

9. Out of scope / open questions

• Redis vs DB-only. Can we keep the shared-infra footprint to just the
  primary DB (using LISTEN/NOTIFY + a DB table for replay/fan-out) so a
  cluster needs no Redis? Trade-off: DB load & NOTIFY fan-out limits vs one
  fewer dependency. The env-var surface above keeps both open.
• Leader-elected singletons. Reapers (D1/D2) and any future heartbeat loop
  could share one leader-election primitive rather than each rolling its own
  lease.
• Observability in a cluster ([feat(observability): uptime, Prometheus /metrics, and OTLP traces across all servers #150]). Per-instance /metrics with tenant as
  a span attribute (never a metric label) is already correct; the open item is
  documenting cross-instance aggregation, not code.
• Bulk export under Mongo/S3 primary (F3). Decide whether to route its job
  state onto the unified database job store (needs a transactional home) or to
  document it as Postgres-primary-only for clustering.

[smunini]

Status update — 2026-08-24

The body above is preserved as the original design record. Since it was written, main has moved on several of its findings and PR #269 has implemented the phased plan, so here is a point-by-point correction of what no longer holds, what narrowed, and where the effort stands.

Corrections — claims that main has invalidated

C1 / §2b — the jti replay cache is gone, and the finding is reversed.
The "SECURITY HOLE (default-on)" assessment was wrong in direction. HFS is a resource server: it has no token endpoint and never receives a private_key_jwt client assertion, so every jti it saw belonged to a reusable bearer access token — and replay-checking those rejected legitimate traffic (unusable with Keycloak, #205). #230 defaulted the cache off; #268 removed the subsystem outright (JtiCache, RedisJtiCache, HFS_AUTH_JTI_BACKEND, HFS_AUTH_REDIS_URL, and the auth redis feature). Genuine bearer anti-replay requires sender-constrained tokens (DPoP/mTLS), which is out of scope here. Consequences for this document: the two jti rows in the §6 env table are dead, the §2b "pluggable shared backend" precedent no longer exists on main, and auth is off the cluster-state list entirely (the ROADMAP now says so explicitly).

C2 — JwksCoordinator was deleted, not wired.
It went in the same cleanup (#268), since its only reason to exist was the removed redis feature. The concern it addressed (N instances independently hammering the IdP on boot/rotation) is real but warn-level — a thundering-herd cost, not a correctness bug. PR #269 rebuilt it as a database-backed single-flight refresh (ClusterRefreshCache + JwksCoordination), no Redis required.

F5 — narrower than written.
The MongoDB half was a misread: every Mongo write has been version-guarded CAS since the backend's original commit. On Postgres, the delete path was fixed on main by #312 (UPDATE … WHERE … AND version_id = …). Still non-atomic on main: Postgres update (crates/persistence/src/backends/postgres/storage.rs — SELECT version_id then unguarded UPDATE) and restore_deleted. PR #269 carries the transactional CAS for those. This remains the one live single-instance bug in the inventory.

Narrowed — landed on main, cluster gap remains

B3 — subscriptions are no longer reactive-only.
#305 (PR #354) added full startup rehydration: SubscriptionTopic, R4-backport Basic, and Subscription resources are paged from storage into the registries at boot, with bounded-concurrency handshake replay (HFS_SUBSCRIPTION_REHYDRATE*). #357 (PR #441) persists server-driven status transitions (HFS_SUBSCRIPTION_PERSIST_STATUS). The topic registry is also tenant-scoped now. What remains per-instance: a Subscription created on instance A after instance B booted is still invisible to B until B restarts (no fan-out — the B1/B2/B5 findings stand unchanged), and the B4 counters (eventNumber, consecutive_failures) are still in-memory and reset on restart.

C3 — local eviction fixed; cross-instance staleness still open.
#304 (PR #352) closed the same-instance gap (SQLite CRUD writes bypassed import_bundle and never evicted) and introduced TerminologyCaches::invalidate_caches with exhaustive destructuring so a new cache field can't be forgotten. The handler-layer cache count has since grown from 5 to 8. Cross-instance invalidation — the actual cluster hazard — is exactly as described: no TTL, no epoch, no NOTIFY.

F3 — now export-only.
MongoDB and S3 gained native $bulk-submit job stores (2026-08-10), so the sidecar-SQLite caveat applies only to bulk export under Mongo/S3 primaries.

Cosmetics. The SoF operations were renamed to $viewdefinition-export / $sqlquery-export (SQL-on-FHIR 3.0.0-ballot), and most cited line numbers have drifted.

Unchanged — re-verified against main @ b6a8200

A1 (SoF InMemoryController still unconditionally constructed; HFS_EXPORT_CONTROLLER still parsed-but-unread), A2 (in-memory ReindexManager), B1 (WS registry, dispatch still Success with 0 clients), B2 (in-memory binding tokens), B5 (in-memory retry; heartbeat_check_interval still unused), D1 (bootstrap sync still lock-free — the only HTS advisory lock guards schema migration), E1 (composite sync still a crash-lossy mpsc(1000)), F1, F2, F4, and the §2a bulk-data lease/fencing description.

Implementation status and plan

PR #269 (feat/cluster-capable-state) implements Phases 0–4 with their gates met: HFS_CLUSTER fail-fast validation, the unified ClusterJobStore (SoF export #169 + reindex), database-backed JWKS single-flight, full subscriptions cluster delivery (#170: fan-out, shared counters, shared binding tokens, durable delivery outbox), HTS cross-instance cache invalidation, the bootstrap import lock, and the durable composite sync outbox — plus the operator chapter (book ch. 15).

Rather than merging it wholesale (~22k lines), the plan is to peel it into phase-sized PRs against current main, in value order:

1. F5 first — the Postgres update/restore_deleted CAS (live single-instance bug; main's delete.rs ignores If-Match entirely — version preconditions are silently dropped on a destructive method #312 fixed only delete).
2. Phase 0 + Phase 1 — HFS_CLUSTER guardrails, the unified job store, SoF export + reindex on it. Closes Database-backed SQL-on-FHIR export job state #169, the only CLUSTER-BREAKING item.
3. Phase 2, database mode only — the Redis coordination variant is dropped: main deliberately removed its Redis dependency, and a second implementation of a warn-level optimization doesn't justify reintroducing it.
4. Phase 4 (C3 invalidation, D1 lock, E1 outbox).
5. Phase 3 (subscriptions, Cluster-aware Subscription notification delivery #170) last — the largest phase; rehydration on main already fixed its restart-loss half, and WS delivery stays documented as sticky-session/single-instance until it lands.

A merge simulation against today's main auto-merges 97 of the branch's 102 files; the 5 conflicts are confined to the SoF export controller, rest/lib.rs, and the subscriptions engine — expected fallout from the operation renames and delivery instrumentation, not a design collision.