Use when designing, reviewing, or hardening PostgreSQL replication and high-availability posture against stated reliability targets. Produces streaming and logical replication topology, synchronous vs asynchronous trade-off decisions, replica lag monitoring and thresholds, automated failover behavior (Patroni, repmgr, or managed Multi-AZ), read-replica routing strategy, split-brain prevention, and multi-region posture. Do not use for schema design, query optimization, backup and restore procedures, or access-control hardening; use postgres-schema-and-migration, postgres-indexing-and-query-optimization, postgres-backup-and-operational-readiness, or postgres-security-and-data-access-hardening instead.
Use when designing, reviewing, or hardening PostgreSQL replication and high-availability posture against stated reliability targets. Produces streaming and logical replication topology, synchronous vs asynchronous trade-off decisions, replica lag monitoring and thresholds, automated failover behavior (Patroni, repmgr, or managed Multi-AZ), read-replica routing strategy, split-brain prevention, and multi-region posture. Do not use for schema design, query optimization, backup and restore procedures, or access-control hardening; use postgres-schema-and-migration, postgres-indexing-and-query-optimization, postgres-backup-and-operational-readiness, or postgres-security-and-data-access-hardening instead.
Postgres Replication and HA Readiness
When to use
Invoke when a PostgreSQL deployment has a stated availability or RPO/RTO target, when designing the replication topology for a new service, when validating an existing topology before a reliability review or scale event, or when remediating failover, split-brain, or replica-lag problems.
Do not use for: schema modeling or migrations (use postgres-schema-and-migration), query/index performance (use postgres-indexing-and-query-optimization), backup/PITR/restore drills (use postgres-backup-and-operational-readiness), or role/grant/TLS hardening (use postgres-security-and-data-access-hardening).
Inputs
Required:
The reliability targets in scope: availability SLO, RPO, RTO, and the failure domains that must be survived. Sourced from reliability-architecture.md where it exists.
The replica topology decision and consistency posture from data-architecture.md (single-primary, read replicas, multi-region, sync vs async expectation).
The deployment substrate: self-managed (Patroni/repmgr), managed (RDS/Aurora/Cloud SQL Multi-AZ), or Kubernetes operator.
Optional:
Current replication setup: physical vs logical, slots, synchronous_standby_names, replica count and placement.
Write and read throughput, and the read/write split the routing layer must serve.
Current replica lag measurements and failover history.
Network topology across AZs/regions and latency budget for synchronous commit.
Connection-pooling and routing layer in use (PgBouncer, Pgpool, application-side).
Operating rules
Replication serves availability and read scaling. It is not a backup. State this explicitly and hand off durability/PITR to postgres-backup-and-operational-readiness.
Every topology decision is justified against a failure domain from reliability-architecture.md: instance loss, AZ loss, region loss, or storage loss. A replica that does not survive a named failure domain is theater.
Synchronous vs asynchronous is an RPO decision with a latency cost. Synchronous commit gives RPO≈0 at a write-latency tax; async gives bounded data loss. State the RPO each choice yields and the latency it costs. Quorum/ANY synchronous standbys are the default over a single sync standby.
Failover must be automated and tested, or it does not count toward RTO. A manual runbook is not an RTO. Name the orchestrator (Patroni, repmgr, managed) and the measured failover time.
Split-brain prevention is mandatory: fencing/STONITH, a consensus layer (etcd/Consul for Patroni), or the managed provider's guarantee. A topology without an explicit split-brain answer is rejected.
Replica lag has a defined threshold tied to the read-consistency contract. State the lag budget, the monitoring signal, and the action when breached (route reads to primary, alert, depool replica).
Read-replica routing makes staleness explicit. Reads requiring read-your-writes go to the primary or use synchronous replicas; eventually-consistent reads may use async replicas with a stated staleness budget.
Logical replication is for selective/heterogeneous/zero-downtime-upgrade flows, not general HA. If proposed, name why physical streaming is insufficient and account for its limitations (no DDL, sequence, or large-object replication by default).
Replication slots prevent WAL loss but can fill the primary's disk if a replica is down. Every slot has a monitoring signal and a max_slot_wal_keep_size posture.
Do not co-locate all replicas in the primary's failure domain. AZ/region placement must match the failure domains the SLO requires.
If the topology cannot meet the stated RPO/RTO within the substrate's constraints, pause and raise an ADR candidate against reliability rather than asserting a target the design cannot hold.
Output contract
The replication and HA design MUST conform to:
observability-standards — replication lag, slot WAL retention, replica health, and failover events are named observable signals with thresholds, not manual checks.
deployment-standards — failover and topology changes are automated, repeatable, and do not require a not-yet-run manual step to hold the SLO.
Upstream contract: reliability-architecture.md is the source of truth for the availability SLO, RPO, RTO, and the failure domains to survive. data-architecture.md is the source of truth for the replica topology decision and consistency posture. If the RPO/RTO is unstated or unachievable on the chosen substrate, pause and raise an ADR candidate against reliability — do not invent or silently weaken the target.
Process
Establish targets and failure domains. Extract availability SLO, RPO, RTO, and the failure domains to survive from reliability-architecture.md. Extract the topology and consistency posture from data-architecture.md. If RPO/RTO is missing, pause and escalate.
Inventory the current (or proposed) topology: primary, standbys, replication type (physical streaming / logical), slots, synchronous_commit and synchronous_standby_names settings, replica placement across AZs/regions, and the routing layer.
Decide synchronous vs asynchronous per the RPO. If RPO≈0 is required, design quorum synchronous standbys (ANY n (...)), state the measured commit-latency cost, and confirm it fits the performance write budget (hand off the budget check if unknown).
Design the standby topology to match failure domains: number of replicas, AZ/region placement, cascading vs direct, and which replicas are failover candidates vs read-only. Every replica maps to a failure domain it lets the system survive.
Choose and specify the failover orchestrator: Patroni (+ etcd/Consul DCS), repmgr, a Kubernetes operator (CloudNativePG/Zalando), or managed Multi-AZ. Define the leader-election mechanism, health checks, and the expected automated failover time. Confirm it meets RTO.
Specify split-brain prevention concretely: DCS quorum and fencing for Patroni, watchdog/STONITH where applicable, or the managed provider's documented guarantee. State the behavior under network partition.
Define replica-lag posture: the lag budget tied to the read-consistency contract, the monitoring signal (pg_stat_replication, pg_wal_lsn_diff, replay lag), the alert threshold, and the automatic action on breach (depool, route to primary).
Define read-replica routing: which read classes are read-your-writes (→ primary or sync replica) vs eventually-consistent (→ async replica with stated staleness budget), and where routing is enforced (PgBouncer/Pgpool/app). Define behavior during failover (read-only window, reconnection).
Define replication-slot posture: slots per standby, max_slot_wal_keep_size, the disk-fill monitoring signal, and the cleanup action when a standby is permanently lost.
If logical replication is in scope, justify it against physical streaming, enumerate its limitations for this use (DDL/sequence/large-object handling, conflict handling), and define the publication/subscription topology and monitoring.
Define multi-region posture if required: sync vs async across regions (latency makes cross-region sync usually infeasible — state the RPO consequence), promotion strategy, and DNS/endpoint failover.
Produce the readiness assessment: topology diagram/narrative, the failure-domain coverage matrix (domain → survives? → how), the RPO/RTO the design delivers vs the target, monitoring signals with thresholds, and the failover test plan. Where the design cannot meet the target, state the gap and the ADR candidate.
Outputs
Required:
replication-ha-readiness.md: topology (narrative or diagram), failure-domain coverage matrix, sync/async decision with RPO and latency cost, failover orchestrator and measured/expected RTO, split-brain prevention mechanism, replica-lag budget and thresholds, read-routing rules, and slot posture.
Monitoring signal list: replication lag, slot WAL retention, replica health, failover events — each with a threshold and an action.
Failover test plan: what is exercised (planned switchover, primary kill, AZ loss), expected behavior, and pass criteria.
Conditional, when applicable:
Logical replication publication/subscription design with limitation analysis.
Multi-region promotion and endpoint-failover plan.
ADR candidate(s) where the substrate cannot meet the stated RPO/RTO, or for sync-commit latency trade-offs.
Output rules:
Every replica and topology element maps to a named failure domain it lets the system survive.
The RPO/RTO the design delivers is stated explicitly and compared against the target.
Failover is automated and has a test plan; a manual runbook is not counted as RTO.
Backup/PITR is explicitly out of scope and handed off to postgres-backup-and-operational-readiness.
Quality checks
Availability SLO, RPO, and RTO are sourced from reliability-architecture.md (or an ADR candidate is raised for the gap).
Every standby maps to a named failure domain it lets the system survive.
Sync vs async is decided per the RPO, with the delivered RPO and the write-latency cost both stated.
The failover orchestrator is named, with an expected/measured automated failover time compared against RTO.
Split-brain prevention is explicit and states behavior under network partition.
Replica-lag budget is tied to the read-consistency contract, with a monitoring signal, threshold, and breach action.
Read routing distinguishes read-your-writes from eventually-consistent reads with a stated staleness budget.
Replication slots have a WAL-retention bound and a disk-fill monitoring signal.
Logical replication, if used, is justified against physical streaming with limitations enumerated.
The design states the RPO/RTO it actually delivers vs the target, and raises an ADR candidate for any gap.
Backup and restore are explicitly deferred to postgres-backup-and-operational-readiness.
