Flame graph

A flame graph turns many sampled call stacks into stacked rectangles. Each box is a function, callers sit below callees, and width represents how often that function appeared in the samples. Wide towers are where the program spent time. Brendan Gregg created the format in 2011 while chasing a MySQL performance problem, so a huge profiler dump could be read at a glance. The warm random palette that gives the graph its name carries no meaning at all: width is the only thing to read.

Reach for one when CPU is high or a request is slow and a trace only tells you which operation is guilty, not which function inside it is burning cycles. Search for a function, follow a wide stack upward, then compare profiles before and after a change.

Gotcha: the horizontal position usually is not a timeline. Flame graph frames are commonly sorted to merge identical stacks; width is sample frequency, not duration along an x-axis. A flame chart is the time-ordered cousin. Sampling can also miss brief work, and off-CPU waiting needs an off-CPU or wall-clock profile rather than a CPU profile.