Performance engineering from kernel analysis to AI: Adrian Cockcroft’s take

The New Stack · 3 weeks ago · 21

Adrian Cockcroft discussed how performance engineers should move beyond relying on system metrics and percentiles by digging into kernel-level causes and then using new analysis techniques for response-time distributions. He said percentiles like P99 do not work for modern web service latency because a single number cannot show whether there are one or multiple latency peaks. As a result, he used an LLM to automate a custom R tool that identifies and tracks multiple peaks over time instead of collapsing data into averages or a single percentile value.