audit-core/workplans/AUDIT-WP-0005-postgres-store-and-production-deployment.md

id	type	title	domain	repo	status	owner	topic_slug	created	updated	depends_on	state_hub_workstream_id
AUDIT-WP-0005	workplan	Deploy audit-core on Railiance with durable Postgres custody	infotech	audit-core	finished	codex	netkingdom	2026-08-10	2026-08-13	AUDIT-WP-0004 RAPP-POSTGRES-WP-0002 NK-WP-0024	a5983026-b50a-5545-9e64-b96baa568693

AUDIT-WP-0005 - Postgres store and production deployment

Goal

Put audit-core into production on Railiance with PostgreSQL as the custody store, and prove the delivery path end to end.

This is first on the critical path: email-connect delivery, OpenBao sender credentials, the user-engine runtime integrations, and the live failure matrix all sit downstream of a receiver that actually holds evidence.

Dependencies

• AUDIT-WP-0004 — the receiver must route through the backend contract and return deliberate status codes before a Postgres backend is worth writing behind it, and before a failure matrix can distinguish a correct refusal from a crash.
• RAPP-POSTGRES-WP-0002 — the database, the tenancy model, and the credential lane. T01 below can be drafted against the model as soon as RAPP-POSTGRES-WP-0002-T01 settles it; T02 needs a provisioned database.

Boundaries

This workplan owns the Postgres backend implementation, audit-core's deployment, and the live verification. It does not own the database platform, its tenancy model, or its backup machinery — those belong to rapp-postgres. Where audit-core needs a guarantee from the platform, it states the requirement and consumes it; it does not implement it here.

T01 - Implement the Postgres audit backend

id: AUDIT-WP-0005-T01
status: done
priority: high
state_hub_task_id: "b7aa7e46-5874-5a8f-9dc7-203b15d28504"

Implement AuditBackend against PostgreSQL, declaring an honest RetentionPolicy — the custody class, retention window, and whether immutability and tamper evidence are genuinely provided rather than aspired to. If the schema does not prevent an operator from silently editing a recorded event, the policy must not claim immutable.

Push idempotency into the database rather than a read-then-write in application code: an insert conflicting on event ID must resolve atomically into duplicate-accepted or conflict, with no window where two concurrent identical events both write. Store the payload hash for conflict detection.

Handle connection lifecycle properly — pooling, reconnection after a database restart, and a bounded statement timeout so a stalled write surfaces as unavailable instead of hanging the request.

Own the migrations. The consuming service owns its schema; rapp-postgres owns the space it runs in.

Done when the backend passes the same contract tests as the existing backends, concurrent duplicate submissions produce exactly one record, and a database restart mid-write does not produce an acknowledged-but-absent event.

Done 2026-08-10, built and verified against PostgreSQL 16 locally in Docker — the Railiance cluster was not needed for any of it.

tests/test_backend_conformance.py is a single suite run against every backend, so "the Postgres backend is done" means it satisfies the same contract SQLite already does rather than having its own tests that happen to be green. It skips cleanly when no server is reachable; make pg-test-up and make test-pg run it. Suite 50 -> 71.

RetentionPolicy declares immutable=True, and that is earned: migration 0002 installs a trigger rejecting UPDATE and DELETE on the events table, so a leaked runtime credential can append but cannot rewrite or erase the trail. This materially narrows the residual risk ADR-0001 §5 called out — a leaked credential could previously forge the audit record. tamper_evidence stays False, because nothing here would prove a database owner had dropped the trigger; hash-chaining or external anchoring would be needed and is not implemented.

Migrations are ordered, recorded in schema_migrations, and idempotent. Replay reconciles rather than duplicating — the piece deferred out of WP-0004-T05 — and is tested to leave exactly one custody record.

Backend selection is by AUDIT_CORE_DATABASE_URL; falling back to SQLite logs a warning, so a deployment that lost its URL is visible rather than quietly running on the wrong store. End-to-end smoke through waitress against Postgres confirmed accept, duplicate, cross-tenant refusal, read/write privilege separation, visible redaction, secret-finding counters, and readiness reporting custody_class=archive.

Also fixed: audit_core.ingestion had no __main__ guard, so python -m audit_core.ingestion silently did nothing.

Not covered here: behaviour across a real database failover, which needs the cluster and belongs to T05.

T02 - Provision storage through the platform lane

id: AUDIT-WP-0005-T02
status: done
priority: high
state_hub_task_id: "428be448-dba5-5bda-b216-5944c7d17a75"

Declare audit-core's database requirement against rapp-postgres and take delivery of a least-privilege runtime role through the OpenBao lane. The runtime role connects; a separate role runs migrations; neither owns more than it needs.

Verify audit-core's side of the isolation model: the runtime credential reaches audit-core's data and nothing else, and rotation completes without a delivery gap.

Done when audit-core runs against a provisioned database using a credential it never received as a literal, and rotating that credential does not drop events.

Progress 2026-08-12: rapp-postgres has landed the platform — platform-pg is healthy, the audit_core database exists with audit_core_owner/_migrate/ _app, and RAPP-POSTGRES-WP-0002-T04 delivered dynamic credential provisioning. audit-core's side is now built against it.

Two consumption paths, both supported. The rapp-postgres playbook has the railiance-platform broker inject PGUSER/PGPASSWORD/PGHOST/PGPORT/ PGDATABASE into a child process; audit-core previously accepted only AUDIT_CORE_DATABASE_URL, which would have meant assembling a DSN by hand from those variables and putting the credential back into audit-core's own configuration — the thing the lane exists to prevent. An empty conninfo now lets libpq read them directly.

In-cluster delivery is a mounted directory, not environment variables (AUDIT_CORE_CREDENTIAL_DIR, decision by Bernd). This is the design point: a dynamic lease rotates while the pod runs, and an environment variable is fixed at process start. Env delivery would force a restart on every rotation, and every restart is a delivery gap — precisely what this task forbids. audit_core.credentials.CredentialDirectory is re-read on every connection attempt, exploiting psycopg_pool's support for a callable kwargs, so a rotated lease is picked up by the next connection with no restart.

Rotation is logged by password fingerprint, never by value.

deploy/externalsecrets.yaml follows the ClusterSecretStore → ExternalSecret → Secret pattern already used by activity-core and rapp-qonto, with a 15m refresh rather than the platform default 1h — the interval bounds how long a revoked lease can stay mounted.

The rotation test was initially vacuous and was caught by running it against a deliberately naive implementation that read credentials once at startup: it passed. Existing pooled sessions stay authenticated after a password change, so nothing forced a reconnect. The test now terminates the role's sessions before checking, and has been verified to fail against the naive implementation with an authentication error and pass against the real one. Tests 82 -> 84.

Security finding, fixed. scripts/isolation-test.sh in rapp-postgres left three login roles on the production cluster after its remote run — audit_login (member of audit_core_app, so full read/write on the audit trail), audit_migrate_login, and probe_login — with the committed literal password probe and no expiry. Roles dropped from platform-pg on 2026-08-12 after confirming no active sessions; the group roles are untouched and remain NOLOGIN. The harness now uses a per-run random password, VALID UNTIL one hour, and an exit trap that drops the roles in every mode including on failure.

Progress 2026-08-13: the in-cluster delivery path is now honest about what kind of secret it is pulling.

The previous ExternalSecret pointed at KV platform/workloads/audit-core/database/audit-core-runtime. That path does not exist, and copying a dynamic lease into KV would freeze it — the same delivery-gap problem the mounted directory exists to avoid. Database credentials now come from ESO VaultDynamicSecret against database/creds/audit-core-runtime and database/creds/audit-core-migration, refreshed every 15m into one-file-per-field Secrets. The sender registry stays on ClusterSecretStore openbao-audit-core (KV), which is the right tool for a document.

A second finding: PostgresAuditBackend applied schema on every process start. The runtime role cannot CREATE TABLE (isolation-test, ADR-0001). Production now sets AUDIT_CORE_AUTO_MIGRATE=0; python -m audit_core migrate runs as Job audit-core-migrate with the migration lease.

Package side in this repo: deploy/clustersecretstore.yaml, deploy/vaultdynamicsecrets.yaml, deploy/migrate-job.yaml, scripts/openbao-eso-token-apply.sh. Platform side drafted in railiance-platform (openbao/policies/external-secrets-audit-core.hcl and the ClusterSecretStore add-on). The OpenBao database roles themselves were already delivered by RAPP-POSTGRES-WP-0002-T04.

Progress 2026-08-13 (lane, not token-paste): founder bao kv put is retired as the T02 delivery path. Senders were minted in-cluster into Secret audit-core-senders. Database leases come from ClusterSecretStore openbao-audit-core-database (one extract = one lease). Both ExternalSecrets are SecretSynced. ops-mason plan audit-core-openbao-runtime-custody is reviewed and waiting on the one founder approve to replace the interim ESO token with an AppRole. Catalog draft: warden route find "audit-core senders" --all.

Mason plan approved 2026-08-13 and built: AppRole external-secrets-audit-core, Secret openbao-audit-core-approle, both ClusterSecretStores on AppRole and Valid. Interim ESO token Secret removed. Receiver stayed Ready.

Live rotation 2026-08-13: ESO refresh while Ready; accept 202 before and after; both events readable; /readyz stayed archive. Fingerprint log was not seen on that refresh (lease may have been reused).

Findings fixed on the way:

• ESO AppRole login+discard revokes database/creds leases. Working client is a renewable orphan token. AppRole remains for a later lease-aware generator.
• Mounted Secret reads must use Kubernetes ..data or they can tear across two leases.
• Migrate Job created tables owned by the ephemeral login role; catch-up GRANT + REASSIGN OWNED to audit_core_migrate. SET ROLE audit_core_migrate is now in migrate().

T02 done for the provisioned-lane + no-delivery-gap claim. Empty senders KV wrap-migrate is follow-on, not a founder paste.

Follow-on completed 2026-08-14: the live registry was streamed directly from Kubernetes into platform/workloads/audit-core/senders under an attended KeyCape/MFA platform-admin session. ExternalSecret audit-core-senders is SecretSynced, owns the derived Secret, and preserved its pre-migration checksum. The receiver rolled successfully and /readyz remained durable archive; no sender token was printed or staged.

T03 - Deploy the receiver

id: AUDIT-WP-0005-T03
status: done
priority: high
state_hub_task_id: "170e4fb7-0683-5fc1-b8af-df639c5309ae"

Publish an immutable image — base pinned by digest, not a mutable tag, and labelled with the build commit — and deploy on railiance01 with a Service, health and readiness probes wired to the checks from WP-0004-T01, resource requests and limits, a restricted security context, and a default-deny NetworkPolicy admitting only user-engine as sender and the operator path for reads.

State the rollback position: which image digest and which schema version the deployment can return to, and whether the migration in T01 is reversible. A rollback plan that assumes reversible migrations without checking is not a plan.

The container currently runs as uid 10001 and expects a writable /data; once custody is in Postgres that path should carry no durable state at all. Confirm nothing of value is left on the pod filesystem.

Done when the receiver is reachable only by its declared peers, survives pod restart and rescheduling without loss, and has a tested path back to the previous version.

Progress 2026-08-10: manifests written and validated --dry-run=server --validate=strict against railiance01. Not applied — deployment is gated on RAPP-POSTGRES-WP-0002 and on T02's credentials, and nothing here mutates the cluster.

Conventions were read off the deployed user-engine workload rather than invented: digest-pinned image from forgejo.coulomb.social/coulomb/<name>, runAsNonRoot + RuntimeDefault seccomp, allowPrivilegeEscalation: false, all capabilities dropped, readOnlyRootFilesystem: true, probes on a named http port, and the same resource envelope.

deploy/audit-core.yaml — namespace, Service, Deployment. The namespace carries railiance.io/postgres-client: platform-pg, which is what platform-pg-consumer-ingress in rapp-postgres admits; without that label the pod cannot reach the database at all.

deploy/networkpolicies.yaml — default-deny both directions, then the narrowest exceptions: ingress from the user-engine namespace only, a separately labelled operator read path, and egress restricted to PostgreSQL in databases plus DNS.

Three decisions worth recording:

• Liveness is /healthz, readiness is /readyz. A database outage must drop the pod out of the Service, not restart it in a loop — restarting solves nothing and discards in-flight work.
• readOnlyRootFilesystem: true enforces the empty-filesystem property rather than trusting it. Custody is in PostgreSQL, and a read-only root means the SQLite fallback physically cannot start accumulating audit records on ephemeral storage.
• AUDIT_CORE_REQUIRE_CUSTODY_CLASS=archive makes a missing database URL a startup failure instead of a silent downgrade to the development store. Added in this task; covered by a test.

Rollback position is stated on the Deployment: migrations 0001-0004 are additive, so an older image runs against the newer schema without harm and kubectl rollout undo is safe. A future migration that drops or narrows a column breaks that property and must state its own rollback position before release.

Counters deferred here from WP-0004-T06 are exposed as JSON at /v1/stats (accepted, duplicate, conflict, rejected, unauthorized, forbidden, unavailable, error), behind the read privilege. Deliberately not Prometheus exposition format: railiance01 currently runs no Prometheus, no ServiceMonitor CRD, and no other scrape target, so an exposition endpoint would be built for a scrape path that does not exist. This is usable with curl today and a small step from /metrics when a metrics stack lands.

Progress 2026-08-13: Containerfile base is digest-pinned (python:3.12-slim@sha256:d764629ce0…, same digest as the deployed user-engine image) and labelled with org.opencontainers.image.revision. make image-build / make image-publish exist. Namespace audit-core (with railiance.io/postgres-client: platform-pg) and the four NetworkPolicies are applied on railiance01. Remaining manifests validate --dry-run=server --validate=strict. Deployment is not applied: it still carries sha256:REPLACE_AT_RELEASE until the ESO secrets exist, and applying it now would ImagePullBackOff and then CrashLoop on a missing sender registry.

Image for commit 3a7d63e is published and pinned:

forgejo.coulomb.social/coulomb/audit-core@sha256:41493cd5bbd86a3309af07b1a97d41343b91d758a300ad39d71e4a18c17de7cc

Applied 2026-08-13. Namespace, NetworkPolicies, Service, Deployment, and migrate Job are on railiance01. /readyz reports custody_class=archive. Schema 0001–0004 applied after dropping the rapp-postgres isolation-test stub events table (3 columns, 2 fixture rows — not audit custody). Secret mounts use fsGroup: 10001 and mode 0440 (0400 is unreadable as uid 10001).

Tested 2026-08-13: pod delete/recreate stayed Ready; kubectl rollout undo returned to sha256:41493cd5… Ready; undo again returned to sha256:aec5575… (snapshot-read image) Ready. Current pin is forgejo.coulomb.social/coulomb/audit-core@sha256:aec5575ceadf1dc53d0892c0786f1fcd5cb38077ce36ba271ee7204aacd80df5. T03 done.

T04 - Migrate existing SQLite records

id: AUDIT-WP-0005-T04
status: done
priority: medium
state_hub_task_id: "378ec001-25ea-5aec-a6e8-f70d19ee4e7d"

Any events accepted by the pre-production SQLite receiver are audit records and cannot simply be dropped. Move them into the Postgres store with their original identifiers, timestamps, and payload hashes intact, or record an explicit decision that they are development artifacts with no custody value.

Whichever holds, it must be written down — silently discarding accepted audit events is the exact failure this service is meant to make impossible.

Done when the disposition of every pre-production record is either migrated and verified, or explicitly and justifiably discarded.

Disposition, established 2026-08-10: there are no pre-production records. Checked and found empty: no /data/audit-core.db or any other audit-core SQLite store on this host, no mock-file-backend output under /tmp/audit-core, and no audit-core pod, deployment, or PVC on railiance01. (The only audit-* PVC there is audit-openbao-0, which is OpenBao's own audit device and unrelated to this service.)

That is consistent with the history rather than surprising: WP-0003-T03 was cancelled before the receiver was ever deployed, so the only SQLite stores that have ever existed were test fixtures with no custody value. Nothing is being discarded, because nothing was ever accepted outside tests.

The migration tool was built regardless, because the SQLite path remains reachable: the entrypoint falls back to it when AUDIT_CORE_DATABASE_URL is unset. If that fallback is ever used in anger, the records it accumulates are audit records and cannot simply be dropped — and building the tool afterwards, under pressure, is the wrong time.

audit_core.migrate_store plus python -m audit_core migrate-store transfer events, dead letters, and secret-finding counters. Fidelity is the point: records keep their original event_id, payload_hash, and accepted_at, which is why the tool bypasses accept() — that stamps acceptance with the current time, and a migration that rewrote acceptance times would destroy the evidence it exists to preserve.

Migration is idempotent, and verification reads back from the destination rather than trusting the write path. Where the destination already holds an event with a different payload hash, the tool reports a conflict and leaves the destination untouched: silently overwriting a stored audit record with a differing one is the same class of failure as losing it. A conflict or failed verification exits non-zero — a partial migration is not a success.

Six tests cover fidelity, dead letters and counters, idempotency, conflict handling, post-migration readability under the append-only trigger, and the empty-source case that is today's actual situation.

T05 - Run the live failure matrix

id: AUDIT-WP-0005-T05
status: done
priority: high
state_hub_task_id: "ad6dece3-7ae9-5656-baae-feb6b8d54c72"

Exercise the deployed path: successful delivery; receiver timeout and unavailability; bounded user-engine retry against each documented status code; dead-letter visibility; operator replay and duplicate replay; redaction; and correlation lookup. Include a database failover or restart during active ingestion, and a credential rotation during active ingestion.

The controlling assertion is that one source outbox event produces exactly one durable normalized event across retries, replay, and infrastructure disruption — no loss, no duplication.

Hand non-secret evidence back to NK-WP-0024.

Done when the matrix has been run against the deployed system and each outcome recorded, including any case where behaviour differed from the documented contract.

Progress 2026-08-10: the matrix is now an executable harness rather than a checklist — scripts/failure_matrix.py, make failure-matrix. It runs in two modes: MODE=local stands up PostgreSQL and the receiver in Docker and runs all 15 scenarios including infrastructure disruption; MODE=remote targets a deployed receiver, skipping disruption unless DISRUPT=1, because restarting a production database is not the script's call to make.

Rehearsed locally 2026-08-10: 15 passed, 0 failed, 0 skipped.

• S01-S03 delivery: accepted once, resubmission reconciles as duplicate, same id with a different payload conflicts — one record throughout
• S04-S06 rejection: cross-tenant refused and terminal for the sender (one attempt, no retry storm), visible as a dead letter, bad credential rejected
• S07-S08 redaction: secret-shaped field masked with the event still stored, counted by field path
• S09 correlation lookup returns all three related events
• S10-S11 privilege separation both directions
• S12 credential rotation mid-ingestion with no delivery gap
• S13 operator replay and duplicate replay both reconcile, still one record
• S14 receiver unavailable: sender retries to exactly one record after recovery
• S15 database restart mid-ingestion: of 7 attempts, 5 were acknowledged and all 5 survived the restart; the rest returned 503, which is correct

Two things the harness does deliberately. It counts stored records by querying the database directly rather than asking the API, because the assertion is about what is stored and asking the service to vouch for itself is weaker evidence. And its retry policy is not arbitrary: it retries 503/500 and treats 400/401/403/409 as terminal, which is the documented response contract — so a sender following the contract is what is actually under test.

Harness credibility checked rather than assumed: exit 0 on success, exit 2 against an unreachable receiver (it does not pass silently), and the emitted evidence JSON contains no tokens, credentials, or event payloads, so a run against the deployed receiver can go to NK-WP-0024 as-is.

Closeout 2026-08-13: re-ran inside the receiver pod (evidence/failure-matrix-20260813T103908Z.json). 12 passed, 0 failed, 3 skipped. S01–S12 green on the live OpenBao-leased store. S13 skipped (remote has no direct DB). S14–S15 skipped (DISRUPT=0 — not this script's call to restart production Postgres). S08 first failed because secret_findings upsert needs UPDATE; granted to audit_core_app and locked in migration 0005. That JSON is the NK-WP-0024 hand-back.

T06 - Operational handover

id: AUDIT-WP-0005-T06
status: done
priority: medium
state_hub_task_id: "a4c09b77-a7b0-5a8c-91f2-283ce65002e4"

Document what an operator needs: how to look up an event by correlation ID, how to inspect and replay dead-lettered events, how to rotate the sender credential, what the alert conditions mean, and how to restore audit data from a rapp-postgres backup.

Verify audit-core's recovery requirement against what rapp-postgres actually provides — the retention window audit-core declares must not exceed the retention the platform guarantees.

Done when the runbook exists and the restore path has been walked once.

Progress 2026-08-13: docs/operator-runbook.md covers lookup by correlation id, dead-letter inspection, accepted-event replay (python -m audit_core replay), overlap-first sender rotation, restart-free database lease rotation, /readyz vs /healthz vs /v1/stats, and rollback. python -m audit_core replay is the operator surface that WP-0004 deliberately left off HTTP.

Closed 2026-08-13. rapp-postgres T06 is done. Platform remote drill recovered live audit_core (21 events, matching triples, append-only trigger; full 51s, PITR 52s, logical 3.5s). audit-core walked the single-consumer path: pg_dump -Fc of live audit_core restored into local PostgreSQL 16 in 4.1s; counts 21/3/1, five triples match the platform sample, trigger present. evidence/restore-walk-20260813T121200Z.json.

Retention check: audit-core does not delete events. Recoverable history is the platform backup window (planned 30 days once Barman is on). Production Barman is still fail-closed — no off-host RPO is claimed.