When Prometheus Can’t Take the Load Anymore
summary
Compares three Prometheus scaleouts—Thanos, Cortex, and M3—by performance, HA, cost and ops complexity and ends up choosing Thanos for simpler deployment and lower cost with acceptable performance. Highlights architecture differences: Thanos sidecar uploads 2‑hour TSDB blocks to object storage, which keeps ops simple but risks losing the last two hours if Prometheus dies; Cortex chunk storage is fastest but needs DynamoDB and heavy caching, and M3 stores replicated data on local disks with etcd coordination, which gives low latency but adds a complex DB to operate.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:38The Problem
- 3:58The Right Path
- 10:03Cortex
- 17:21Cortex Advantages
- 21:27Atmos Architecture
- 27:04Atmos Disadvantages
- 28:07Summary
- 29:01Other Aspects
- 30:32Conclusion
related talks
26mWhen Prometheus Can’t Take the Load Anymore
26mPrometheus That Scales
20mIntro to Scaling Prometheus with Cortex
37mPrometheus Deep Dive
41mHigh Available + Scalable Prometheus with Thanos in Alibaba
37mIntro: Cortex
35mDoing Things Prometheus Can’t Do with Prometheus
31mLong-Term Metrics with M3 and Prometheus