Database operations · Lesson 3 of 6 · 4 min read
Replication and high availability
Synchronous and asynchronous replication
- Synchronous: the primary waits until a replica confirms it has the change before acknowledging the commit. No data loss on failover (RPO 0), but every write pays the network round trip, and if the replica stalls, writes stall.
- Asynchronous: the primary acknowledges immediately and ships changes afterwards. Fast, but replicas lag behind, and a failover can lose the last few transactions.
- Replication lag is how far behind a replica is (seconds, or bytes of WAL). Reads from a lagging replica return old data, so "I saved it but the page doesn't show it" bugs appear when reads go to replicas.
Enlarge diagram
Swipe sideways to see all of it, or tap to open it full screen.
PostgreSQL uses streaming replication (physical, the whole cluster) and logical replication (selected tables, across versions, useful for migrations and upgrades). MySQL replicates using the binary log, with GTIDs to track position.
RDS and Aurora options
| Option | How it works | Failover | Readable? |
|---|---|---|---|
| RDS Multi-AZ (instance) | Synchronous standby in another AZ | Automatic, typically 1–2 minutes | No, the standby only waits |
| RDS Multi-AZ DB cluster | One writer plus two readable standbys in different AZs | Typically under 35 seconds | Yes |
| RDS read replicas | Asynchronous copies, same or another region | Manual promotion | Yes |
| Aurora | Shared storage keeping 6 copies across 3 AZs; up to 15 replicas | Usually under 30 seconds, to a replica | Yes |
| Aurora Global Database | Storage-level replication to other regions, typically under a second behind | Managed switchover or failover to another region | Yes, in secondary regions |
Read replicas are for scaling reads and some DR; Multi-AZ is for availability. They solve different problems and are often used together.
Enlarge diagram
Swipe sideways to see all of it, or tap to open it full screen.
Surviving a failover
During failover, the database endpoint's DNS record is moved to the new primary. Applications must cope:
Enlarge diagram
Swipe sideways to see all of it, or tap to open it full screen.
- Reconnect and retry: connections to the old primary break, so the app's connection pool must detect dead connections and retry with back-off.
- Don't cache DNS forever: the JVM in particular can cache DNS; set a short TTL (
networkaddress.cache.ttl=5or similar) so the new address is picked up. - Use the right endpoint: the cluster (writer) endpoint for writes, the reader endpoint for reads, never an instance endpoint for the primary.
- RDS Proxy keeps client connections open and reconnects to the new primary itself, cutting the visible failover time considerably.
- Test it:
aws rds reboot-db-instance --force-failoveror Aurora's failover action in a staging environment, while running load, and watch what the app does.