unknown

Connection pool poisoning in Go's database/sql can escalate from row-level lock contention to complete database outages when combined with misconfigured PgBouncer deployments. When transactions opened via raw SQL aren't properly cleaned up, they poison the connection pool by holding row locks indefinitely. Without PgBouncer peering enabled across load-balanced instances, cancel requests route to the wrong instance and fail, causing clients to hard-close sockets. Postgres doesn't clean up connections waiting on locks when sockets close, leading to CLOSE_WAIT accumulation and connection exhaustion. The article demonstrates this failure mode through test scenarios and provides detection strategies (monitoring transaction duration, connection counts, lock waits) and prevention options including enabling PgBouncer peering, fixing application connection leaks, implementing session affinity, and configuring appropriate timeout defaults.