Skip to content

Flame Graphs

"Once you can read a flame graph, you can debug performance in any language whose profiler emits stacks."

Flame graphs are the visualisation Brendan Gregg invented to make sampled stack profiles tractable to read. Width = time spent (or allocations, or any sampled metric); stack depth = call depth; colour is usually meaningless (just contrast). They work for CPU, allocations, off-CPU, lock contention — anything that emits stacks.

Scope

  • Anatomy: x-axis = time/share, y-axis = stack depth, ordering is alphabetical (no time axis)
  • Reading: wide plateaus = hot functions, deep narrow towers = hot call paths, "icicle" inverted flame graphs
  • Differential flame graphs — before/after comparison
  • Off-CPU flame graphs — what's blocking, not what's running (the under-used variant)
  • Generating: flamegraph.pl, cargo flamegraph, pprof -http, async-profiler, Speedscope, FlameScope
  • Common misreadings: "this function is 50% so I'll optimise it" (cumulative vs flat confusion)
  • CPU Profiling — the most common profile flame graphs visualise.
  • Allocation Profiling — same visualisation, different metric.
  • Brendan Gregg's reference — brendangregg.com/flamegraphs.html (the canonical resource).

Status

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