Each /api/wind refresh decoded two or three 0.25 degree GRIB2 messages with
ecCodes (WASM), validated ~1.04 M doubles per message and resampled them, all
synchronously on the server thread. Measured with cached upstream bytes: 43
to 164 ms of event-loop block per warm refresh, 230 to 381 ms for the first
one in a process, 316 to 974 ms when all six keys refresh together.
Decode, validation and resampling now run on one lazily started worker
thread, which transfers back only the display-grid samples. The server keeps
the sampled U/V of the last forecast object per model (1.04 MB each), so an
overlay change, or a refresh within the same forecast step, fetches and
decodes only the scalar. Manifests and grid bodies are byte-identical to
before for the same upstream bytes.
Co-Authored-By: Claude Fable 5.2 <[email protected]>
