Maximizing Error Injection Realism with System Calls
summary
A prototype called Phoebe uses eBPF-based monitors to record real syscall failures in production, clusters them into sporadic/fluctuating/steady patterns, and generates injection models by error-rate amplification. An injector rewrites syscall return codes to replay amplified real errors. Evaluations on an email server (Hadwig) and a BitTorrent client show some syscall errors cause crashes or state corruption while others are tolerated, and overhead stays low enough for production experiments.
AI-written summary · about
chapters (12)
- 0:00Chaos engineering challenges
- 3:24Efficiency in chaos experiments
- 4:49System calls as failure target
- 7:17Live demo: system call monitoring
- 12:53Designing Phoebe framework
- 15:20Synthesizing realistic error models
- 20:16Evaluating Hedwig email server
- 23:27Evaluating Transmission client
- 24:52eBPF for system call interception
- 26:58Phoebe limitations and future work
- 28:45Research contributions summary
- 29:27Open-source resources
related talks
37mReal-time Scheduling Fault Simulation
32mListen to your Engine
26mReinventing Seccomp for Fun and Profiles
42mCommunity Driven Breakout Session: eBPF and Auto-Instrumentation
14mWill eBPF Save Us From the Next Global Outage?
32mReal Time Security - eBPF for Preventing attacks
42mResolving Performance and Security Issues with eBPF
36meBPF for Runtime Security