Resource Control Admission - I have a date with my PSI
summary
Pressure Stall Information (PSI) exposes microsecond counters and 50 ms samples for CPU IO and memory so you can detect stalls rather than wait for loadavg; Linux has PSI since 2018 (Facebook pushed it) and Cloudflare supplies a Prometheus exporter. The pg_psi extension registers kernel triggers (example: 250 ms stall over 4 s) and can poll or run a background worker to cancel or suspend a VACUUM FULL when cgroup IO/CPU/memory exceeds thresholds; non‑root windows are limited to 2 s today and eBPF integration is planned.
AI-written summary · about
chapters (13)
- 0:00Introduction and agenda
- 2:38PostgreSQL processes categorized
- 4:46PostgreSQL resource requirements
- 7:41Pressure Stall Information (PSI)
- 9:49Observing pressure
- 11:38CPU avg vs loadavg
- 13:26PSI visualization (pseudo demo)
- 15:56Event notification
- 17:28Cgroup support
- 19:00Bringing PSI to PostgreSQL
- 21:18The future of PSI
- 23:56A “demo”
- 25:31Links & questions
related talks
27mPostgreSQL container groups, aka cgroups down the road
25mpg_stat_monitor: A cool extension for better monitoring using PMM
37mPractical Capacity Planning using Prometheus
24mModelling Postgres Performance Degradation on Burstable Cloud Instances
27mKubernetes SIG Storage: Intro & Deep Dive
50mHow To Improve PostgreSQL Monitoring and Observability With PMM
37mPerformance Diagnostics with Open Distro for Elasticsearch
56mPostgreSQL at low Level: stay curious!