Production Profiling: What, Why and How
summary
Claims that profiling in production is now practical and low overhead and shows how sampling (SIGPROF, perf, Java Flight Recorder and hardware event sampling) reveals real CPU hot spots that safe point based stacks miss. Gives concrete business impact: 100 ms latency cuts sales ~1% and 500 ms cuts traffic ~20%. Explains instrumentation overhead and bias, contrasts CPU time versus wall clock, and argues for always-on profiling storing historical profiles for regression analysis across microservices — a workable approach that actually finds and fixes regressions.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:20Customer Experience
- 2:14Responsive Applications make more Money
- 5:32Unrepresentative Hardware
- 7:05Unrepresentative Software
- 9:25Unrepresentative Workloads
- 16:27Coarse Grained Instrumentation
- 21:11Normalisation of Deviance
- 23:37How to use Production Profilers
- 26:11Profiling Hotspots
- 28:39Instrumenting Profilers
- 36:22Advanced Statistical Profiling in Java
- 38:55Profiling Support in the Linux Kernel
- 40:58Historical Data
- 43:06Putting Samples in Context
- 43:43Summary
related talks
01h 13mContinuous Profiling: The Missing Piece in Your Observability Puzzle
01h 00mThe Basics of Profiling
40mContinuous Profiling, the missing piece in your observability puzzle!
48mProfiling is for Dev, Monitoring is for Ops – no more!
56mContinuous Profiling the missing piece in your observability puzzle
47mContinuous Profiling — what, why and how
39mProfiling and Benchmarking 101
01h 17mThe Basics of Profiling