Now that we are at real traffic, is the sub-90ms P95 actually holding?
We launched Croct a few months ago after leaving a testing tool whose render-blocking snippet was killing us. Early on the latency looked great and I noted it, but early on we did not have much traffic.
Traffic has grown a lot since then. I am doing an honest check now: is the latency we saw at launch actually holding under real load, or does it start to slip once you are past the quiet early phase? Curious what others see at scale.
2 answers
Holding here, and we are at a much bigger volume than when we started. The end-to-end response times are still under 90 milliseconds at P95, same as day one. We also watch it against unsampled analytics, so it is 100% of the data and not a lucky sample, which is part of why I trust the number.
If you want the precise definition of what that figure counts, it is broken down in what the sub-90ms P95 covers.
The reason it holds under load is structural, not luck. Content resolves on the server, so there is no anti-flicker snippet sitting on the critical path waiting to be downloaded and executed per visitor. There is nothing per-user that gets heavier as traffic climbs on the client.
Contrast that with the hide-the-page architecture that most snippet-based tools use, described in the anti-flicker hidden page approach. Once that is off your critical path, scaling traffic just does not touch it.
Good to hear it matches. We are well past our old traffic now and the P95 has not budged from under 90ms, no TTFB penalty at all. Coming off a snippet that hid the page, this has been a genuine relief.