Protocol-Aware Recovery for Consensus-Based Storage
Talkat USENIX ATC '19 · on Wednesday, 2019-07-10· at 16:30· USENIX ATC Track I: Grand Ballroom I–VI·
summary
Paper introduces protocol-aware recovery and implements CTRL to exploit RSM protocol rules to recover from disk corruption. Implementation in LogCabin and ZooKeeper fixes logs, provides safety and high availability where unmodified systems are unsafe or unavailable — baseline systems fail in about 30% unsafe injection cases while CTRL always preserves committed data. Evaluation shows low cost: about 4% overhead on SSDs for write-heavy workloads, plus broader fault-injection results and practical recovery rules.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:21Redundancy Enables Reliability
- 0:53How to Recover Faulty Data
- 2:20Problem: Approach is Protocol-Oblivious
- 3:10Results Overview
- 3:47Outline
- 4:53Background - RSM Persistent Structures
- 5:43Current Approaches to Handling
- 6:25Protocol-oblivious: Crash
- 8:17RSM Recovery Taxonomy
- 9:09CTRL Overview
- 9:32CTRL Local Storage
- 10:07Crash-Corruption Entanglement in the Log
- 10:58CTRL Distributed Recovery
- 11:09Follower Log Recovery
- 11:53Leader Log Recovery
- 12:22Leader Log: Determining Commitment
- 13:59Reliability Experiments Example
- 14:34Reliability Experiments Summary
- 15:07Reliability Results Summary
- 15:18Update Performance (SSD)
related talks
01h 12mRedundancy does not imply fault tolerance
16mPilot Execution: Simulating Failure Recovery In Situ for Production Distributed Systems
46mKubernetes Stateful Application Disaster Recovery with Zero Data Loss
15mTesting Consensus Implementations using Communication Closure
19mSpeculative Recovery: Cheap, Highly Available Fault Tolerance with Disaggregated...
33mStrong and Efficient Consistency with Consistency-Aware Durability
35mMaking Ceph Fast in the Face of Failure
38mTowards consensus on distributed consensus #bbuzz