We scaled PgBouncer to 4x throughput (clickhouse.com)

Article summary

The article discusses how ClickHouse Managed Postgres scaled PgBouncer to 4x throughput by running a fleet of processes, utilizing multiple CPU cores, and implementing peering to enable query cancellation across processes. This approach allows the pooler to handle increased concurrency without becoming a bottleneck. The setup uses the so_reuseport feature to load-balance incoming connections across the processes. The result is a significant increase in throughput, with the fleet of processes handling 4x more transactions per second than a single process.

Main themes

• PgBouncer scaling
• Connection pooling
• Concurrency handling
• Postgres performance
• Load balancing
• so_reuseport feature

What commenters say

• Using HAProxy with multiple PGBouncer instances may not be the most efficient approach due to the extra network hop.
• The so_reuseport feature is a more efficient solution than using an extra userspace hop.
• Other scalable PgBouncer alternatives, such as Odyssey, are available and worth considering.
• PgBouncer's peering feature is a built-in and simple solution for enabling query cancellation across processes.
• Some users have successfully used other connection pooling tools, such as pgdog, to scale their Postgres instances.
• The choice of connection pooling tool depends on specific use cases and requirements.
• Having too many connections to Postgres can be inefficient, and a smaller number of connections can be sufficient for scaling.
• The use of so_reuseport and peering can be applied to other scenarios beyond PgBouncer, such as torrent clients.