Cutting P95 page load for staff on constrained networks

The application was built and tested on office connections, and it showed. At the branches, where the work actually happened, the 95th-percentile page load was 7.1 seconds — long enough to change how staff worked around the tool.

The network was not mine to fix and the hardware was not going to be replaced. Whatever I changed had to reduce bytes on the wire and round trips, without a rewrite and without disrupting people mid-shift.

• Measured at the 95th percentile rather than the average, because the average hid exactly the users who were suffering
• Enabled Brotli compression to cut transferred bytes
• Moved the tool onto HTTP/2 to remove head-of-line blocking across the many small requests the page made

• P95 page load fell from 7.1s to 4.3s, a 40% reduction
• ~1,500 branch staff affected daily

I would have set a performance budget tied to the branch network profile and enforced it in CI. The wins here were real but one-off; without a budget, the same regression creeps back in over the following year.

• Brotli
• HTTP/2
• Node.js
• MongoDB