Batch Prediction and HPO

Many ML workloads are naturally parallel: score these files, try these parameters, process these images. Coiled Functions and Dask clusters both fit this shape.

Batch prediction

@coiled.function(region="us-west-2")
def predict(uri):
    batch = load_batch(uri)
    return score_batch(batch)

results = predict.map(input_uris)

Hyperparameter search

Use simple mapped functions for basic sweeps. Use Dask-integrated tools when trials need shared coordination or richer study state.

Operational note

Parallelism should not hide evaluation. Log trial parameters, datasets, metrics, and failure modes.