For a long time I treated profilers as something other people used. The output looked like a mountain range and I had no idea which peak to care about.

Two rules and you can read one

• Width is time. A wide bar spent a lot of the sample in that function, and that's the only axis that matters.
• Height is call depth, not cost. A tall tower of thin bars is deep code, not slow code.
• Look for a wide plateau near the top: that's where the time actually goes, not the frames below it that merely called it.

The first real one I read had a wide plateau in JSON serialisation, on a route I had assumed was database-bound. I had spent two days tuning the wrong half of the request.

Profile first. Intuition is a hypothesis, not evidence.