Database timeouts fail the tap, even though we'd expect the pool to swap out timed out connecitons

[visch]

Tech tangents reported this here #24 (comment)

I think we could replicate this by turning down the replication timeout to something really low like 20 seconds and testing this locally.