Performance Testing Done Right
summary
Argues that latency and throughput are distinct: latency tests fix throughput, throughput tests push limits; shows a round‑robin scheduler that lowers tail latency without improving throughput and a multithread change that boosts throughput but hurts latency. Uses HDR Histogram to show full percentile distributions, runs daily automated benchmarks in Perf Repo (Jenkins uploads, commit hashes) and finds concrete surprises like an 80/20→88/11 read/write drift, stressor context‑switch limits, and a hidden 1 Gbit cap that ruins scaling.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:45Why performance matters at Hazelcast
- 3:10Key metrics: throughput and latency
- 5:15Problem: mixing latency and throughput tests
- 13:34Problem: incorrect load stressing levels
- 22:49Problem: discarding performance data
- 29:55Problem: mismatched operation ratios
- 33:46Problem: ignoring regression
- 37:19Resources and demo
- 39:32Q&A
related talks
41mShift Left your Performance Testing
29mProgressive Testing to Meet the Performance Imperative
50mDesigning for Performance
38mCreating & Scaling Meaningful Performance Tests in CI
47mLee Barnes: Maximizing the Value of Performance Testing Throughout the Software Development Cycle
41mPerformance Engineering in the Age of DevOps and Agile
01h 00mPerformance In Modern Times: An Evolutionary Journey
12mPerformance In Context