Thanos Deep Dive: Inside a Distributed Monitoring System
Talkat KubeCon + CloudNativeCon North America 2019 · on Wednesday, 2019-11-20· at 17:20· Room 6C - San Diego Convention Center, Upper Level·
summary
Thanos extends Prometheus with a store API that lets queriers fan out to sidecars and a store gateway serving object storage, enabling global queries and multi-year retention. Long-range queries hit huge compressed gorilla chunks (≈11x), and decompression costs (tens of nanoseconds per sample) drive latency, so a compactor produces downsampling to 5m/1h with five aggregation chunks (count,sum,min,max,counter) to cut query cost at the price of more storage; scaling uses time partitioning.
AI-written summary · about
chapters (8)
- 0:00Start
- 2:39Let's quickly reiterate on Thanos
- 8:14Thanos Query: Store Discovery
- 9:33Thanos Query: Life of a query
- 20:00Downsampling: What chunk to use on query?
- 23:07U Downsampling: Caveats
- 27:24Time partitioning
- 28:00Block Sharding
related talks
36mReal World Architecture - Building a Global Cross-Cloud Monitoring...
24mThanos Receiver Deep Dive
18mArchitecting Growth: Scaling Tactics for Prometheus Metrics...
33mPredicting the Future of Distributed Systems
35mSystems @Scale Summer Q&A Week 3
26mTorchStore: What We Learned Building Distributed Storage Sol...
16mFrom Sampling To Full Visibility: Scaling Tracing To Trillions of Spans
38mLog Analytics: Understanding Complex Systems at Scale