Describe the bug

Requests to hydra (running with at least postgres) can sometimes hit an "idle" (but closed) connection to the DB. In this scenario the caller gets back:

error_hint=write tcp SRC:37738->DST:5432: write: broken pipe

I've run into similar errors in other apps, and the issue is upstream in lib/pq as well (lib/pq#870).

I was able to work around this issue by setting max_conn_lifetime=10s in the DSN -- as it avoids the connections being held open but actually closed on the remote.

Reproducing the bug

This has been annoying to reproduce on-demand. I've noticed this when tokens go to get refreshed from the IDP -- specifically after a long time (8hrs in this case) such that the connection can close out.

Expected behavior

I would expect that L4 errors such as this would be transparently retried instead of bubbling all the way to the client.