Skip to content

platform-postgres-ha.md

Latest commit

 

History

History
executable file
·
33 lines (26 loc) · 1.55 KB

platform-postgres-ha.md

File metadata and controls

executable file
·
33 lines (26 loc) · 1.55 KB
Raw

Platform PostgreSQL HA runbook

Preconditions

The three nodes must be independently powered and labelled node-role.kubernetes.io/postgres. Their replicated storage must survive a single node loss. CloudNativePG 1.30.x, cert-manager, and the Barman Cloud plugin must be healthy before this chart is applied. The object-store endpoint, all replicas, logs, and backup keys must remain in India.

The chart requires custom server and client CAs. Workloads authenticate with role-specific TLS client certificates only. Never enable BYPASSRLS, password fallback, public database connection, or superuser application access to make a rollout easier.

Promotion gate

Before production promotion, record all of the following in the change ticket:

  1. kubectl get cluster aether-platform -n aethercode -o yaml shows three healthy instances, minSyncReplicas: 1, and a distinct hostname for each.
  2. A test write is visible on a synchronous standby before the client is acknowledged.
  3. A controlled loss of the primary node promotes a replica and services resume without lost committed writes. Run the judge queue replay test during this exercise.
  4. A fresh-cluster PITR restore from the India-resident object store reaches a timestamp selected in the drill, and application RLS tests pass against it.
  5. The restore time and recovered point meet the configured recovery objectives.

Do not claim readiness from a successful Helm install alone. A failed backup, unscheduled third instance, loss of synchronous replication, or failed restore blocks promotion.