No video of the event yet, sorry!

This talk demonstrates how running Postgres on log-structured (LSM) file systems, for example F2FS, can significantly improve throughput in write-heavy workloads.

I'll show how I identified surprising bottlenecks in high throughput Postgres systems, where the main issue turned out to be the number of IOPS, due to small random writes. The system could not utilize the full storage bandwidth, due to overwhelming it with too many requests.

I experimented with various strategies to improve this by merging small I/O requests, to reduce pressure on the I/O queues, and better utilize the I/O bandwidth. One of those experiments was using log-structured file systems (LSM), particularly F2FS, which naturally coalesce small writes - this turned out to align well with how Postgres writes data.

This resulted in a substantial increase in write TPS, without any measurable (negative) impact on read performance.

In this talk, I’ll discuss:

• how to spot and diagnose IOPS bottlenecks in PostgreSQL
• the performance impact of small write patterns
• benefits of using LSM-style file systems for write throughput
• benchmark results and lessons learned from real-world testing
• pros/cons of using this in production systems