What’s the latency impact of serverless cold starts on real-time features?

[AvaT]

I’m trying to decide if I should build my next project’s backend using a serverless architecture, but I’m worried about cold starts affecting real-time features. Has anyone actually measured the latency impact for something like live user notifications, or is it mostly theoretical now?

[BrianYW]

I actually measured this for a realtime chat feature. Early on, cold starts would show up as hiccups during spikes, around 300–500 ms. Once we kept a small pool warm (provisioned concurrency), the steady-state latency stayed under 100 ms for most messages and tails hovered under 200 ms.

[ScottJ]

In our project we used a fan-out pattern with a queue. The latency mostly came from queue depth and downstream services; the startup delays were only part of it.

[StevenW]

Is the real bottleneck even the backend, or is the client polling and websocket keep-alives the bigger drag?

[Jerry_D]

I tried warming for a sprint and then paused; I didn't want to pay for idle seconds. The measurement seemed to vary by region and runtime version, not a universal rule.