Ran into something interesting: I’m running AB in GKE, but have the db backend in a Cloud SQL postgres instance. I kept noticing random blips when using airbyte, but it wasn’t until I wanted to dig into the db that I figured out what was going on. My postgres instance has a very low max concurrent connections (25) since it is a micro instance. Airbyte ends up using a lot of connections since PG isn’t very efficient at expiring old connections. When I looked at the server logs, there were a TON of 

FATAL: remaining connection slots are reserved for non-replication super-users

 errors. To solve this issue, I ended up deploying pgbouncer to my GKE cluster and reconfiguring the airbyte backend to use the pgbouncer instance.

Thank you for letting us know. This is interesting.

PG isn’t very efficient at expiring old connections.

Maybe you can add some parameters to the connection string related to keepalive and timeout? https://www.postgresql.org/docs/9.6/libpq-connect.html#LIBPQ-PARAMKEYWORDS Like

keepalives_idle

.