Connection pooling em postgresql: caracterização empírica do ponto de equilíbrio entre sobrecarga e contenção

Abstract

PostgreSQL database systems suffer severe performance degradation under high connection concurrency, owing to their process-per-connection model and contention in internal locking structures. Connection poolers are the industry-standard mitigation, but the decision to adopt one usually rests on heuristics rather than quantitative evidence, since there is no public characterization of the concurrency point beyond which a pooler offsets the overhead it introduces. This work investigates that transition through a controlled empirical analysis of six connection configurations over PostgreSQL 18 in an AWS cloud environment: a direct connection (baseline), PgBouncer and PgCat in transaction and session modes, and AWS RDS Proxy. Across 1,332 trials, the workload, query mode, concurrency, and pool size were varied, with instrumentation of both the client and the server internals. The results indicate a break-even point: at low concurrency the pooler merely adds latency; beyond a few dozen clients (about 30 under the write workload), the direct connection collapses under contention while the pooler keeps throughput stable. Server-side instrumentation attributes this gain to reduced internal lock contention rather than to a lower connection cost. The multi-threaded pooler (PgCat) outperforms the single-threaded one (PgBouncer) only when the pooler itself becomes the critical resource, and managed pooling (RDS Proxy) shows a lower throughput ceiling than the self-hosted alternatives. The study provides an open, reproducible characterization of this break-even for PostgreSQL, grounded in causal evidence collected at the server.