Platform under continuous improvement  —  Our platform is currently undergoing continuous improvements and expansion. While some sections are still being finalized, our infrastructure and services remain fully operational. We appreciate your patience as we continue building a faster, stronger, and more comprehensive platform for our clients.   ●   Platform under continuous improvement  —  Our platform is currently undergoing continuous improvements and expansion. While some sections are still being finalized, our infrastructure and services remain fully operational. We appreciate your patience as we continue building a faster, stronger, and more comprehensive platform for our clients.
Back to blog Engineering

PostgreSQL tuning for streaming workloads — what we changed and why

Clustrix Admin · May 11, 2026 · 6 min read

When a client migrated their video metadata workload to PostgreSQL 16, the default config burned 40% of available RAM on the wrong thing. Here's what we changed.

shared_buffers

The default 128MB is laughably wrong for a 128GB box. Set it to 25% of system RAM as a starting point.

work_mem

Per-operation memory. Multiply by max_connections × average concurrent ops to estimate worst case.

WAL compression

Turn it on. CPU cost is negligible; disk and replication bandwidth savings are dramatic.

The result

P99 query latency dropped from 240ms to 38ms. Disk throughput halved.

KEEP READING

Related posts

Engineering

Why we standardized on BBR for every server we deploy

Like this kind of detail?

That's the level of attention we put into every client's infrastructure.