If I may, one thing that has been a constant troublemaker for us - running primarily on AWS - is timeouts. I would love to see a fundamental implementation of timeouts at the connection level, regardless of activity - e.g. in the middle of a query, between queries, etc.. If the connection hangs (because the server disappeared or failed over) or does not respond (particularly in the weird cases like when the destination IP just vanishes and the tcp timeout kicks in) - I'd like to know about it within a reasonable time and be able to react to that.
It would be interesting to set up integration tests wirh connections being disrupted in various ways and at different moments during connection and query processes. I'd rather do them in a test framework rather than one-off.
