Downsampling in Prometheus
summary
Downsampling is defined as reducing sample rate while keeping recent full-resolution data; current practical options are federation or recording rules with remote_write to a long‑term Prometheus. An experimental promtool offline TSDB command implements both simple decimation and the LTTB (largest‑triangle‑three‑buckets) algorithm and runs on a 21 GB, 1‑year, 15s dataset, but a bug prevents trustworthy space‑saving numbers. Findings: counters and histograms downsample well, gauges do not, and fewer samples hurt TSDB compression efficiency.
AI-written summary · about
chapters (14)
- 0:00Downsampling in Prometheus
- 1:26What downsampling means
- 3:09Prometheus as multi-rate system
- 4:13Defining downsampling
- 6:38Visualizing counters and gauges
- 8:25Current downsampling methods
- 11:40Trade-offs of approaches
- 12:24Selecting samples for downsampling
- 13:52Largest triangle three buckets
- 16:40Experimenting with algorithms
- 17:46Challenges in storage
- 19:56Future plans and feedback
- 20:17Key takeaways
- 21:22Q&A
related talks
21mSparse High-resolution Histograms in the Prometheus TSDB
35mPrometheus Sparse High-Resolution Histograms in Action
36mPrometheus - Intro, Deep Dive, And Open Q+A
30mLifecycle of a Sample in the Prometheus TSDB
22mA New Kid in Histogram Town
36mDeep Dive: Prometheus
42mPrometheus Intro and Deep Dive
23mPrometheus – What it is, What is New, and What is Coming