Ceph Durability: How Safe Is My Data?
summary
A new Monte Carlo simulator models Ceph MTTDL and nines across real configs and confirms some concrete surprises: doubling PGs per OSD drops a triple-replica MTTDL from ~1.8M years to ~1.1M years; tiny objects (compare 4MB to 4KB objects) collapse durability from millions of years to ~1,300 years; erasure coding K4+M2 is far weaker than replica-3; recovery sleep and rebuild rate dominate the risk. The tool is available in the repo and runs simulate and analyze modes against cluster JSON data.
AI-written summary · about
chapters (8)
- 0:00Introduction to Ceph durability
- 2:29Durability vs availability explained
- 9:14Introducing the reliability calculator
- 15:57Erasure coding vs replication
- 18:47Impact of object size and recovery
- 23:07Live cluster analysis demo
- 28:50Recommendations and future steps
- 30:34Q&A session
related talks
33mThe Challenge of Storing Small Objects on a Large Scale
31mAI, ML, and the Ceph Advantage: Scalable Storage for Smarter Workflows
24mThe Butterfly Effect of a Broken Disk: Top-Down Ceph Troubleshooting To Upstream Contribution
43mCeph and Flash
29mCeph Object Storage Overview, Capabilities and Future Plans
30mCluster storage systems gotta have HeART...
18mData Security and Storage Hardening in Rook and Ceph
36mYour Manila CephFS Share Backups Belong to S3