successful switchover, a way to terminate connections on an old master

[pere3]

Overview

Hello crunchy team! Thanks for a superb product on a postgres-operator landscape.

We currently are in a process of adapting this solution on our GKE-based infrastructure with a scale of 260-300 databases in a single cluster, and so far it's looking good and promising, thanks for a detailed documentation.

I wonder what yours take on a problem that we are currently facing:

• we are deploying postgres as two replicas cluster
• our application uses -ha svc to connect to the main replica
• we perform controllable switchover (we update the cluster spec and apply an annotation)
• switchover is performed well and a replica switched
• but our application (python + asyncpg) still manages old connections to now-read-only replica, which causing exceptions ReadOnlySQLTransactionError until amount of them reaches the limit to terminate connection by asyncpg driver

so switchover executes in a matter of seconds, but our application still uses old connections for much longer (maybe until restart, if limit of failed ops could not be reached) - what is proposed way to solve this behavior?

Environment

• Platform: GKE
• Platform Version: 5.1.2
• PGO Image Tag: ubi8-5.1.2-0
• Postgres Version 14
• Storage: PVC