Presented at PGConf.dev 2026 in Vancouver. The full title is Connection Pooling Beyond PgBouncer: A Custom Approach for Distributed PostgreSQL.
Existing poolers like PgBouncer work at the protocol level, with no deep integration into a distributed database. This talk is about the pooler we built from scratch for Multigres instead. It natively supports the extended query protocol (Parse/Bind/Execute), keeps per-user connection pools so Row-Level Security keeps working, and consolidates prepared statements across clients.
I go through the problems we hit along the way. Connections are routed by their session settings into hash buckets, with an LRU cache so comparing settings is a pointer-equality check. Pinning connections for transactions has to be balanced against keeping the pool busy. The extended query protocol has to survive a distributed query path. I also explain the three-tier pool layout (admin, regular and reserved) and where we traded connection reuse for protocol correctness.
Related reading: the Multigres connection pooling series, starting with Two jobs, two processes, then Per-user pools that share fairly, Pooling without choosing a mode and One parse per query