Understanding Microservice Performance
summary
Argues for recording timestamped per-request records (duration + success/error) and using percentiles not means, specifically 95th, 99th, 99.9th percentiles and maximum; shows JMeter truncates sub‑millisecond times to 0ms as a gotcha. Demonstrates that an average web page makes ~42 requests so there’s ~88% chance a user sees worse than the 95th. Highlights Little’s and Amdahl’s laws, warns latency explodes past ~75–80% utilization, and uses Guesstimate Monte Carlo modeling and R fits to the Universal Scalability Law to forecast retrograde performance.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:52Agenda
- 3:22Performance vs Latency
- 5:17Measuring
- 12:33Simulation
- 13:53Latency and throughput
- 15:18Littles Law
- 19:08Guesstimate
- 21:14Latency distributions
- 23:53SLA inversion
- 25:29Amdahls law
- 27:57Universal Scalability Law
- 28:57Can Increasing Capacity Reduce Performance
- 31:31Visualization
- 39:19Benchmarking
related talks
32mSoftware Performance: A Shape Not a Number
38mA Practical Look at Performance Theory
50mDesigning for Performance
38mTechniques and Tools for a Coherent Discussion about Performance in Complex...
40mPerformance Testing Done Right
47mResolving performance issues in practice
36mMoneyball for Performance Metrics
33mRedefining Web Performance