CPU Profiling¶

"A function that's 30% of the profile is the one to look at. A function that's 0.3% is not — even if the change is satisfying to make."

CPU profiles attribute where the CPU spent its cycles — wall-clock vs on-CPU, sampling vs instrumenting, what's actually a "hot" function vs what just runs a lot. This is the most commonly read profile shape, and the one where bad reading habits do the most damage.

Scope¶

• Sampling profilers: how they work (signal-driven stack capture, perf events), accuracy vs overhead
• Wall-clock vs on-CPU time — and why the two profiles disagree
• pprof (Go), async-profiler / JFR (Java), perf (Linux), Instruments (macOS), py-spy (Python), cargo flamegraph (Rust)
• Symbolication: why a frame says ?? and how to fix it
• Reading: top, callgraph, cumulative vs flat; the difference matters

Related¶

• Flame Graphs — the dominant visualisation for CPU profiles.
• Benchmarking — comparison after profile-guided change.
• Diagnostics → Diagnostic Endpoints — /debug/pprof for live capture.

Status¶

⏳ Per-tier files (junior / middle / senior / professional / interview) pending.