Supporting 3 Availability Zones Stretch Cluster
summary
Shows an implemented Ceph upstream fix that prevents a specific data‑loss edge case where a replication size 6 pool with min_size 3 can accept writes from a lone datacenter then lose those writes when the others rejoin and the lone site fails. It exposes tunables — peering crush bucket count, target and barrier — and a new acting‑set selection rule that caps per‑site replicas so a PG won’t go active with replicas only from one site. Net‑split detection is added as a short‑term remedy; code is in Squid and main.
AI-written summary · about
chapters (12)
- 0:00Introduction and agenda
- 1:04What is a stretch cluster
- 3:11Monitor election in stretch clusters
- 4:57Net split problem in two-site setup
- 6:20Peering rule challenges
- 8:05Stretch mode with three availability zones
- 12:27Expanding stretch mode to three sites
- 13:28Peering crush bucket count explained
- 16:02Peering crush bucket target explained
- 18:54Limitations of three availability zones
- 20:00Solutions for net split issues
- 21:49Q&A session
related talks
32mStretch Clusters in Ceph: Algorithms, Use Cases, and Improvements
25mExplicitly Supporting Stretch Clusters in Ceph
27mDetect, Decide, Fence: Resolving Netsplits in 3-AZ Stretch Cluster
46mKubernetes Stateful Application Disaster Recovery with Zero Data Loss
34mModernizing Ceph Deployments at CERN: CephFS and Object Storage Across Data Centres
27mFrom Chaos To Clarity: Systematic Ceph Cluster Analysis for Rapid Troublesho...
26mImplementing a Resilient Ceph RGW Object Storage with Erasure Coded Data Pool as the Foundation...
06mNobody Knows What PGs are Good For, Only I Do