Operations, Security & Reliability

Failover, Disaster Recovery, and Reliability

Reliable failover is a rehearsed control loop that detects real impact, prevents two writers through fencing, promotes a sufficiently current copy, reroutes clients, measures loss, rebuilds redundancy, and learns within an explicit reliability budget.
  • Detection must balance speed and false promotion.
  • Fence before promoting.
  • Promotion chooses a data-loss boundary.
  • Client routing is part of recovery.
  • Failover ends only after redundancy and protections return.
  • DR includes regional dependencies and human operation.
  • Reliability consumes a budget.
Failover and DR runbook
1 DETECT: declare impact, commander, timeline; freeze risky automation.
2 ASSESS: primary reachability, candidate receive/flush/replay positions, archive health.
3 FENCE: revoke route/lease/storage/credential or power off old writer; verify from independent path.
4 PROMOTE: choose most suitable eligible candidate; record expected RPO/data-loss window.
5 ROUTE: update proxy/DNS/service discovery; drain stale pools; verify writes/read-your-writes.
6 VERIFY: business invariants, unknown commits, audit/security, backup/archive, lag and objectives.
7 REBUILD: create replicas from new timeline; restore capacity and monitoring.
8 FAILBACK: only as a planned migration after evidence review.
9 LEARN: postmortem, owners, tested actions, game-day update.
Fencing prevents split brain
Regional dependency analysis
DependencyFailure questionExercise evidence
Identity/secrets/KMSCan responders and apps authenticate/decrypt?clean-room issuance and rotation
Network/DNS/proxyCan clients converge without stale writer paths?TTL/pool drain and fence checks
Backup/archive/object storeAre bytes, catalog, keys independently reachable?isolated restore to selected target
Artifacts/config/extensionsCan compatible database/app be rebuilt?versioned manifests and bootstrap
Observability/auditCan responders see and preserve evidence?out-of-region dashboards and immutable sink
People/vendorsWho has authority and access off-hours?game-day contacts, approvals, handoff