Performance profiling

Programmer intuition about performance is famously unreliable. The bottleneck is almost never where it feels like it should be, and the time spent optimising the wrong thing is doubly wasted — it costs the work and it adds complexity that makes the real fix harder.

Measure, change one thing, measure again. Every step below serves that loop.

1. Define slow, numerically

"Slow" is not actionable. Establish:

• What operation, measured where — server time, wall clock, time to first byte, time to interactive? These give different answers and different fixes.
• The current number, at a percentile. Use p95 or p99, not the mean. Averages hide the users who are actually suffering, and a mean that looks fine routinely conceals a tail that is ten times worse.
• The target, and why that number. "Under 200ms because it blocks a keystroke" is a target; "faster" is not.