The work-around for the first issue might work well in practice, but due to asymmetry of tcp streams, is still open for a race where the server will idle-close a connection while a new (non-idempotent!) query is just in-flight.

The correcter solution would let the client manage the idle-timeout and disconnect once reached. Depending a bit on how controlled the client is, this might be a good or bad idea.

Note that this problem is very generalizable to http (1) upstream servers. If you need to support non-idempotent requests and want persistent connections to your upstream, it's not a good idea to have the upstream manage the idle timeout (and disconnect if reached).

In practice, I would have the client manage an idle-timeout of 30s, and server of 40s as an extra protection against misbehaving clients.