Profiling and Flame Graphs

Suppose an API deployment increases CPU time per request by 40%. Host and container metrics confirm CPU pressure, and thread-level inspection identifies the busy workers.

The next question is:

Logs usually cannot answer this. They capture selected events, not every function executed between them.

A CPU profiler solves this by sampling the running program at regular intervals and recording the active call stacks. Those samples reveal which functions consume the most CPU time.

A flame graph visualizes the same data, making hot code paths easy to spot.

Profiling shows where CPU time is actually being spent, without adding instrumentation to every function.

This chapter develops a bounded CPU-profiling workflow, covering target and event selection, call-graph capture, symbol and stack validation, flame-graph construction and interpretation, fair profile comparison, and regression diagnosis.