Failover
Failover transfers a service or resource function to an alternate component or site when the previous one cannot or should not continue serving it.
Scope: Failover can be planned or unplanned, and manually or automatically initiated. The mechanism may transfer a primary role or redirect requests to a replacement instance. A changed route does not by itself exclude a previous writer. [1] [2]
Data boundary: With asynchronous PostgreSQL replication, promotion can omit committed transactions that never reached the selected standby. Restoring access and satisfying a recovery-point objective are separate checks; the failover label promises neither zero loss nor a fixed interruption time. [3]
Afterward: Rebuilding a standby restores redundancy; returning the service to its original primary is a separate failback decision. Some deployments keep the new primary instead. [2] [1]
Research contrast: Remus delayed network output until backup checkpoint acknowledgement, but its 2008 prototype lacked execution fencing after loss of both replication paths. Output protection within a fail-stop model does not establish partition safety. [5]
Relations
- related
- Failback, I/O fencing, Asynchronous replication, Storage availability, Recovery point objective, Recovery time objective
Limitations
- Promotion does not itself prevent the former primary from returning as an independent writer. [2]
Citations
- Microsoft Azure reliability: failover and failback, planned/unplanned failovers and failback
- PostgreSQL 18, section 26.3: Failover, former-primary exclusion and standby reconstruction
- PostgreSQL 18, section 26.2.8: Synchronous Replication, asynchronous loss window
- Cully et al. Remus: High Availability via Asynchronous Virtual Machine Replication. NSDI 2008, sections 2.1, 2.4 and 2.6.
Further reading
Sources & caveats
- Primary documentation reviewed on 2026-11-18; product-specific examples do not establish a universal failover procedure or measured recovery time.