Feature Store · View 12 of 21 · Runtime

Critical Flow — An ETA Vector Read

Inside a user-facing request with a 250 ms total budget, of which the feature read gets 25 ms.

Editable source SVG draw.io All views

The compiled plan is resolved from a local cache, never from the registry, which is what keeps the read inside its budget and lets serving survive registry loss.
One batch read per request across groups rather than one read per group; the fan-out is the dominant term in tail latency.
An unavailable group returns a reason code and the declared default, and the response says which values are which. Only a consumer-declared critical group fails the call.

Server-side p50 3 ms, p99 15 ms, p99.9 40 ms for ≤ 200 features across ≤ 5 groups; on-demand transformation adds ≤ 2 ms p99.
Client-observed p99 ≤ 25 ms including network.

On-demand transformation runs consumer-authored code inside the serving path. It buys parity on exactly the transformations most likely to skew, and it puts that code in the latency budget and the failure domain.