Hot Standby vs Replica
Failover patterns.
Overview
Hot standby and read replica are both replicas, but they serve different purposes. Hot standby is for failover: synchronous (or semi-sync) replication, promotable on primary failure with low RPO. Read replica is for read scaling: asynchronous, serves read queries, no promotion path expected. Confusing the two produces architectures that fail at exactly the moment they were supposed to protect.
- Failover versus read scaling. Hot standby for failover; replica for read scale. Different replication modes serve different needs.
- Hot standby. Synchronous or semi-sync replication. Can be promoted on primary failure with low RPO.
- Read replica. Asynchronous replication. Serves reads; not designed for promotion.
- Replication mode matters plus per-engine differences. Sync versus async versus semi-sync drives durability; Postgres, MySQL, and SQL Server each handle this differently.
The approach
Three habits keep the choice rational: hot standby for HA needs, read replica for read scaling, RPO documented per replica so failover behaviour is understood before the failover happens.
- Hot standby for HA. Synchronous or semi-sync. Promotion path tested.
- Read replica for read scale. Asynchronous. Serves analytical or read-heavy workloads.
- Documented RPO. Per-replica the RPO target. Contractual commitments depend on this.
- Tested failover plus documented topology. Game-day exercises validate the failover path; per-database the replication configuration documented.
Why this compounds
Each correctly-architected replica supports its purpose for years. The team’s database HA fluency deepens; failover stops being a theoretical capability and becomes a tested one.
- HA improves. Hot standby supports real failover. RPO matches the documented commitment.
- Read scale improves. Replicas serve reads. Primary stays available for writes.
- Operational fit. Right replica for the use case. Failover does what the runbook says.
- Year-one investment, year-two habit. First replica is heavy lift. By the third, the topology pattern is settled.