Prometheus Sparse High-Resolution Histograms in Action
summary
Demonstrates a proof‑of‑concept that encodes entire histograms into a single sample so Prometheus stores one series per histogram instead of one per bucket, uses a protobuf scrape endpoint for targets, and yields ~90% index size savings and ~50% disk savings in benchmarks. Client libraries pick a power‑of‑two resolution, Prometheus TSDB stores compact spans of filled buckets, Grafana renders sparse high-resolution histograms as heatmaps live in the demo and the implementation currently runs in open‑source branches.
AI-written summary · about
chapters (8)
- 0:00Start
- 0:59What is a histogram?
- 2:53Problem with this design
- 6:00First property: Fixed bucket boundaries
- 7:14Fixed bucket boundaries, higher resolution
- 8:49Convert to a lower resolution
- 13:10One time series per histogram
- 14:25Instrumentation
related talks
22mA New Kid in Histogram Town
25mNative Histograms in Prometheus
36mHow to Use Prometheus's Native Histograms
21mPrometheus Native Histograms in Production
26mUsing OpenTelemetry’s Exponential Histograms in Prometheus
42mPrometheus Intro and Deep Dive
05mNative Histograms With Custom Buckets
36mPrometheus - Intro, Deep Dive, And Open Q+A