We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·24d ago·5 min read
We scaled PgBouncer to 4x throughput
TL;DR
PgBouncer's single-threaded nature limits throughput on multi-core systems, leading to inefficiencies. By deploying multiple PgBouncer processes with the SO_REUSEPORT option, ClickHouse Managed Postgres effectively utilizes all CPU cores.
✦ Why It Matters
Engineers should implement SO_REUSEPORT and peering in PgBouncer setups to maximize throughput and ensure effective query cancellation.
Key Takeaways
- Deploy multiple PgBouncer processes to utilize all CPU cores.
- Use SO_REUSEPORT to enable kernel load balancing of connections.
- Implement a peering mechanism for effective query cancellation.
- Monitor throughput improvements to validate performance gains.
Full Summary
PgBouncer, a connection pooler for PostgreSQL, operates as a single-threaded process, which restricts its throughput on multi-core machines. In ClickHouse Managed Postgres, multiple PgBouncer instances are deployed, each bound to the same port using the SO_REUSEPORT socket option, enabling the kernel to distribute incoming connections across these processes.
This setup allows for better CPU utilization, as all cores can handle connections simultaneously. However, a challenge arises with query cancellation requests, which can be directed to the wrong PgBouncer process.
To address this, a peering mechanism was implemented, allowing processes to communicate and forward cancellation requests to the correct instance. As a result, ClickHouse Managed Postgres achieved a fourfold increase in throughput while maintaining effective query cancellation across the fleet.