PgBouncer vs Pgpool-II vs Odyssey: choosing a PostgreSQL connection pooler

PostgreSQL uses one backend process per connection. Past a few hundred concurrent connections, memory overhead and context switching dominate and throughput collapses. A connection pooler becomes architecturally mandatory at scale. PgBouncer, pgpool-II, and Odyssey make different trade-offs around session state, prepared statements, throughput, and operational complexity. The wrong choice leads to session-leak bugs, throughput ceilings, or split-brain failures unrelated to PostgreSQL itself.

flowchart TD
    A[Choose a PostgreSQL pooler] --> B{Session state required?}
    B -->|Yes| C[Session mode: PgBouncer, Odyssey, or pgpool-II]
    B -->|No| D{Built-in load balancing or HA?}
    D -->|Yes| E[pgpool-II]
    D -->|No| F{Prepared statement caching?}
    F -->|Yes| G[Odyssey]
    F -->|No| H[PgBouncer transaction mode]

Pooling modes and session state

Start by determining what session state must survive across client requests. PostgreSQL session features include SET variables, temporary tables, advisory locks, and LISTEN/NOTIFY. A pooler can preserve these in session mode or strip them in transaction mode.

PgBouncer supports session, transaction, and statement pooling. Session mode preserves all session features but multiplexes least. Transaction mode is the default recommendation: it reuses a backend per transaction, maximizing concurrency and minimizing backend memory. Statement mode breaks multi-statement transactions and is rarely used in production.

pgpool-II supports session pooling only. Every client connection holds a dedicated backend for its lifetime, so session features work as expected. You lose the aggressive multiplexing of transaction mode, which limits connection scaling.

Odyssey supports transaction and session pooling. Like PgBouncer, it can run in transaction mode for maximum efficiency or session mode when applications require stateful behavior.

The critical operational gotcha is session state leakage in transaction mode. When server_reset_query = DISCARD ALL is set, PgBouncer clears state after each transaction. This means SET LOCAL, temporary tables, and advisory locks do not survive the transaction boundary. If your application assumes these persist, you will see intermittent, hard-to-reproduce failures that look like application bugs but are pooler misconfiguration. Size pool_size to roughly two or three times the CPU core count on the PostgreSQL server; this gives enough backend parallelism without memory bloat. See How PostgreSQL actually works in production: a mental model for operators for how MVCC and session state interact.

Prepared statement support

Modern drivers and ORMs rely heavily on prepared statements. Whether they work depends on the pooler and the mode.

In PgBouncer transaction mode, prepared statements do not work. Because each transaction may run on a different backend, a prepared plan created in one transaction is not visible in the next. The only workaround is session mode, which eliminates most of the pooling benefit. If your stack uses an ORM that prepares statements aggressively, PgBouncer transaction mode will cause errors or silent fallback to unplanned execution.

Odyssey offers prepared statement caching in transaction mode. It can track prepared plans across backend assignments, which makes it compatible with drivers and ORMs that rely on them without forcing you into session mode.

pgpool-II uses session pooling, so prepared statements work transparently.

If prepared statements are non-negotiable and you want transaction-mode efficiency, Odyssey is the only option among the three. If you can disable prepared statements at the driver level, PgBouncer transaction mode remains viable.

Throughput and concurrency architecture

PgBouncer is a lightweight, single-threaded event loop. It adds minimal overhead. Its simplicity is a feature: less code, fewer locks, and predictable latency. The trade-off is that a single PgBouncer process scales vertically up to a CPU core, after which you run multiple instances behind a load balancer. This is well-understood operational behavior: you scale out the pooler layer, not up.

pgpool-II is a heavier middleware stack. It parses queries, manages replication topology, and implements its own watchdog for HA. That feature surface adds latency and reduces raw throughput. If your primary goal is to reduce connection overhead, pgpool-II gives you more than you asked for.

Odyssey uses a multi-threaded architecture. Unlike PgBouncer’s process-per-worker horizontal scaling, Odyssey handles higher concurrency inside a single process. This can simplify deployment at very high connection counts, though it is newer and less battle-tested at extreme scale than PgBouncer.

Load balancing, failover, and routing

This is where pgpool-II differentiates itself, for better and worse.

pgpool-II includes built-in read load balancing, query caching, automatic failover, and watchdog-based HA. If you need a single binary that pools connections and routes reads to replicas, pgpool-II is the only one that offers it natively. The cost is operational complexity: watchdog split-brain is notoriously difficult to debug, query caching has concurrency issues, and the configuration surface is large enough that misconfiguration is common. It also carries stability concerns, including an authentication bypass vulnerability (CVE-2025-46801). It should be patched and reviewed carefully.

PgBouncer does no load balancing, failover, or query parsing. It is strictly a connection concentrator. If you need HA or read routing, you pair it with Patroni for failover and HAProxy or DNS for routing. This keeps concerns separated and failure modes simpler, but it means more moving parts in the architecture.

Odyssey likewise does not provide built-in load balancing or automatic failover. Plan to use external tooling for topology changes and read distribution.

Operational complexity and failure modes

PgBouncer is the simplest to operate. Configuration is a single INI file. The admin console exposes pool state with SHOW commands. Deploy it in an HA pair behind a load balancer to eliminate a single point of failure. The most common production failure is pool exhaustion during deployments, when new application instances open connections before old instances drain. Fix this by sizing max_client_conn and default_pool_size correctly, enforcing connection limits in application connection strings, and draining old instances before scaling down. Session state misconfiguration is the other major source of incidents.

pgpool-II is a full middleware stack. The surface area creates risk: watchdog leader election races, query cache invalidation bugs, and configuration sprawl. Teams often end up managing pgpool-II alongside separate failover tooling because the built-in HA is hard to trust under pressure. If your need for load balancing is not absolute, the operational cost usually outweighs the benefit.

Odyssey sits between the two: more features than PgBouncer, but a smaller community and fewer public war stories. Yandex developed it, giving it credible engineering lineage, though its production footprint outside that ecosystem is still growing. If you adopt it, plan to build more internal runbook coverage and accept that public debugging guidance is thinner than for PgBouncer.

Selection reference

No pooler fixes bad application behavior. If you see connection exhaustion, the first response should not be to switch poolers. It should be to check whether the application is leaking connections, holding transactions open, or opening new pools during health checks. The table below assumes your application side is already behaving correctly and you are deciding which pooler semantics fit your workload.

If you are already seeing connection saturation, review PostgreSQL connection exhaustion: detection, diagnosis, and prevention and PostgreSQL FATAL: too many connections - causes and fixes before changing poolers.

How Netdata helps

Netdata correlates pooler and PostgreSQL signals so you can tell whether a slowdown is in the database or the connection layer.

• Connection saturation: Correlate pg_stat_activity counts with application error rates to distinguish between pooler exhaustion and true max_connections saturation.