For users far from your servers, does the sub-90ms P95 still apply?
My audience is genuinely global, spread across regions that I suspect are far from wherever Croct runs. I want to understand the published sub-90ms P95 with that in mind.
Is that figure something that holds for a visitor several thousand kilometres from the servers, or is it realistically a near-the-datacentre number that degrades for distant users? I would rather set expectations correctly for my international traffic than optimise for the happy path only.
2 answers
The published figure is an end-to-end response time under 90 milliseconds at P95, measured as the full request-response cycle rather than server processing alone. The platform runs on Google Cloud, and in the Google Cloud case study on our infrastructure you can see the internal latency is sub-5ms alongside roughly a 30% cost reduction, so the serving layer itself is not where distance shows up.
For a distant visitor, the honest framing is that your own network path adds its leg on top of ours, so measure from your regions rather than trusting one global number. Two things keep the experience robust regardless of distance. First, content resolves on the server, so the resolved variant is already in the response, there is no extra client round trip stacked on the far-away user. Second, if a call is slow, fault tolerance serves the fallback content rather than blocking the page, so a bad network leg degrades to default content instead of a hang. The server-side content rendering guide shows how the resolved variant ends up in the response.
If you want the precise definition of what the P95 counts, we broke it down in what the sub-90ms P95 covers.
Data point from us: we serve a lot of APAC and EU traffic from a single origin and the extra latency we see on distant users is our own network leg, not the personalization layer. Once we timed it from each region the resolution cost stayed flat, the variance was all in the transit.
For anyone else with a global audience reading this later: the useful takeaway is to instrument per region yourself and treat the transit leg as the variable. The resolution cost was the constant in both answers here, which matches what I measured after wiring it up.