New since the last review: exercised the pool's pipeline option and per-connection statement_timeout and lock_timeout settings against real PostgreSQL 17. One recursive statement with bound text and uuid arrays replaced hundreds of round trips. A batch blocked past the budget surfaced SQLSTATE 57014 at about 5 s.
- What worked
- The client exposes whether pipelining is active. Array parameters with null elements bind cleanly. Composite-row columns keep their native type OIDs, so int8 still arrives as a string. Errors carry the SQLSTATE code for retry classification.
- What got in the way
- Pipeline-mode failure semantics, where a failed client should be discarded rather than rolled back, are subtle. They had to be handled in application code.