A 98 percent hit ratio can still hide your slowest requests. Here is what to measure instead. Performance work goes wrong when it starts with a guess. Every change in this post began with a measurement on real hardware and ended with a second measurement to prove it helped.

Measure before you touch anything
Field data from real users tells you where the problem is; lab tools tell you why. We start with the Core Web Vitals report, pick the worst page template, and only then open the profiler.
- Start from field data, not a lab score.
- Fix the largest resource first.
- Compare percentiles, not single runs.
- Re-measure a week later.
The change that mattered
The biggest single win was rarely clever. Right-sized images, fewer third-party scripts, and a cache header on static assets did more than any amount of micro-optimisation in our own code.
Any fool can write code that a computer can understand. Good programmers write code that humans can understand.
Martin Fowler, Refactoring, 1999
Quick takeaway
Measure, change one thing, measure again.
Verifying the win
Field data
Real users on real devices.
Right-sized media
Serve the pixels the layout needs.
Caching
Long cache lifetimes for fingerprinted assets.
After each change we compare the median and the 75th percentile, not the best run. A fix that only helps fast devices is not a fix for the people who were actually suffering.
If you try this on your own project, start with one page or one service rather than the whole codebase. A small experiment gives you real numbers to argue with, and it is easy to roll back if the idea does not fit.
We have been running this approach for several months now. The initial setup took about a day, and the ongoing cost is close to zero once the team knows where to look.
None of this is new, and that is the point. The techniques that last are usually the boring ones that are documented well and supported by the platform itself.
As always, the examples are simplified to fit on a page. The production version has more error handling and logging, but the shape of the code is the same.
Wrapping up
As always, the examples are simplified to fit on a page. The production version has more error handling and logging, but the shape of the code is the same.
Thanks for including the numbers. So many posts skip the measurement part.
Thanks for catching that, it is fixed now.
+1, this matches what we saw after the migration.
Thanks for catching that, it is fixed now.
Bookmarked. The code sample alone saved me an afternoon.