High-Performance PostgreSQL Connection Pooler

Problem

Standard connection poolers (PgBouncer, PgPool-II) lack modern features: query-level routing, prepared statement lifecycle management, adaptive pool sizing based on workload patterns, and detailed observability.

Architecture

Rust-based proxy server using tokio for async I/O. Implements PostgreSQL wire protocol. Features: query parsing for routing decisions, statement cache with LRU eviction, adaptive pool sizing using PID controller, Prometheus metrics, and distributed tracing.

Technologies

RustTokioPostgreSQL ProtocolPrometheusOpenTelemetryClapSerde

Key Highlights

Sub-millisecond overhead vs direct connections
Query-based routing: read replicas for SELECT, primary for writes
Prepared statement cache with automatic invalidation on DDL
Adaptive pool sizing responds to workload changes in seconds
Zero-downtime configuration reload via SIGHUP
Comprehensive metrics: latency histograms, pool utilization, query classification
Supports PostgreSQL 14+ features (pipeline mode, COPY streaming)

Challenges

Implementing full PostgreSQL wire protocol correctly
Handling prepared statement lifecycle across pool boundaries
Achieving zero-copy query parsing for routing decisions
Balancing latency vs throughput in adaptive algorithms
Ensuring correctness under connection failures and failover

Solutions

Used property-based testing with protocol specification
Designed statement cache with generational IDs for invalidation
Built custom query parser using nom for zero-copy parsing
Implemented PID controller with workload classification
Added comprehensive chaos testing in CI pipeline

Lessons Learned

Protocol implementation correctness requires exhaustive testing
Adaptive algorithms need guardrails, not just optimization
Observability is not optional for infrastructure software
Rust's type system prevents entire classes of connection bugs
Performance optimization requires measurement at every layer

Key Highlights

Sub-millisecond overhead vs direct connections

Query-based routing: read replicas for SELECT, primary for writes

Prepared statement cache with automatic invalidation on DDL

Adaptive pool sizing responds to workload changes in seconds

Zero-downtime configuration reload via SIGHUP

Comprehensive metrics: latency histograms, pool utilization, query classification

Supports PostgreSQL 14+ features (pipeline mode, COPY streaming)