ServiceLab: Preventing Tiny Performance Regressions at Hyperscale through...
Talkat 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI '24) · 2024-07-10..07-12·
summary
ServiceLab is a pre-production A/B performance platform that replays sampled RPCs across on-demand heterogeneous machines to catch tiny regressions before rollouts, detecting regressions as small as 0.01% and reporting a 0.14% median regression on a serverless front-end running on >500k machines. A study of 21M hardware benchmarks finds five factors account for 87% variance (server type, CPU, kernel, turbo, datacenter), Trials multiply parallel replicas for tighter CIs, and static analysis selects 23% of commits to test, with 3% signaled.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:15Challenges at hyperscale
- 3:57ServiceLab overview
- 6:07Experiment design
- 7:54Dependency management
- 8:53Allocating machines
- 10:02Detecting tiny regressions
- 11:05Adaptive statistical methods
- 12:30Production experience
- 13:03FrontFaaS case study
- 13:42Results and feedback
- 13:56Key takeaways and conclusion
related talks
01h 08mPerformance testing in the cloud
23mStop Performance Regressions Before Release Valkey’s CI Benchmarking
53mEarly Performance Testing: Finding News You Can Use
56mBetting on Performance: A note on Hypothesis-driven Performance Testing
33mPerformance Testing in Agile and DevOps
45mPerformance-by-Design: Embedding Intelligent Scaling and Guardrails Into Platform Engineering
41mPerformance Engineering in the Age of DevOps and Agile
38mA Practical Look at Performance Theory
