A3S Docs
A3S CloudCloud 0.1.0

Dependency and Process Runbooks

Roles, prerequisites, verification, recovery, and rollback for every current Cloud dependency.

Dependency and Process Runbooks

Status: Experimental

Snapshot: cloud-0.1.0

These runbooks describe the current integration snapshot. Provider-specific HA topology, retention, RPO, and RTO remain operator policy until H0 is verified.

Process roles

RoleOwnsRequired dependenciesSafe verification
apiREST, OpenAPI, authentication, accepted intentPostgreSQL, configured security providerslive/ready endpoints and a scoped read query
workerdurable operations, deployments, builds, edge/certificate convergence, retentionPostgreSQL, Flow, node control, Runtime nodes, Gateway; object storage as configuredlease age, operation progress, node observations, Gateway acknowledgements
relayPostgreSQL outbox publicationPostgreSQL and NATS when selectedoutbox watermark, publish latency, NATS stream health
allall three roles in one processunion of the aboveuse for development, not an HA claim

Start data services first, then relay/worker, then API ingress. Stop ingress first, wait for mutations, then workers/relay. Never run workers from two restored sites without proven fencing.

PostgreSQL

Prerequisites: supported PostgreSQL, TLS/network policy, backups, capacity for migrations and outbox growth. Alert on ready-health failure, connection exhaustion, long transactions, stale worker leases, and outbox lag.

On failure, close mutation ingress, preserve the database and logs, restore or fail over through the provider, then start Cloud with workers paused. Verify migration level and authoritative queries before workers resume. Roll back by returning to the last verified database authority; never merge two writable histories.

NATS JetStream

Status: Optional when the in-memory event provider is selected. PostgreSQL remains authoritative. Verify stream existence, retention, consumer health, and the Cloud outbox watermark. On failure, keep accepted intent in PostgreSQL, stop or back off relay publication, restore NATS through its provider, and resume idempotent publication. Consumers must tolerate redelivery. Do not mark an operation failed solely because a derived event relay is temporarily down.

Object storage

Status: Optional for the local profile and required by production policies that select S3-compatible logs, artifacts, backups, SBOMs, or provenance. Require TLS, bucket versioning/retention as policy dictates, least-privilege credentials, and digest verification. On corruption or absence, keep metadata truthful and report an explicit gap; never substitute unverified bytes. Rollback a storage migration by restoring the prior endpoint and version IDs before workers resume writes.

Gateway

Gateway owns ingress, complete route snapshots, managed public certificates, and exact reload acknowledgements. Verify management authentication, service health, installed revision/digest, routes, certificate SAN/expiry, and a routed probe. On failed reload, keep the previous complete snapshot active. Roll back through a new command that installs the previous verified snapshot; never edit only one route behind Cloud's revision authority.

Runtime nodes

Nodes make outbound mTLS connections, persist command journals, execute provider-neutral Runtime units, and ship observations/logs. Verify certificate validity, heartbeat freshness, capability set, free resources, Runtime state, command receipts, and log watermarks. Drain before maintenance. Preserve state directories during repair so idempotent commands can replay. If identity is lost, re-enroll and reconcile; do not copy another node's private key.

Flow workers

Flow advances durable deployment, build, cleanup, and recovery operations. Verify worker lease age, run/event sequence, retry backoff, command TTLs, and terminal cleanup. On a crash, restart against the same PostgreSQL authority and let the persisted run resume. Do not manually skip a failed step or issue a new logical operation until the idempotency and resource receipts prove the old one terminal.

Control-plane recovery gate

After any dependency incident, require all of the following before reopening mutations: live and ready health, PostgreSQL authority, bounded outbox lag, non-stale worker leases, current node observations, exact Gateway acknowledgement, verified object digests, and no unexplained non-terminal operation. If any check fails, keep ingress closed and restore the previous verified dependency or snapshot.

On this page