Batch inference applies a trained model to a finite collection of inputs in grouped execution. In production, it is commonly coordinated as a reproducible pipeline rather than run manually. section 5.0 section 10.1
A production batch-inference workflow typically coordinates these stages:
- Fetch an entity dataset from object storage. section 5.1.2
- Retrieve the required features from a feature store. section 5.1.2
- Load a named model version and run inference over the feature data. section 5.1.2
- Write predictions to an output bucket for downstream use. section 5.1.2
Income-classifier pipeline
A Kubeflow pipeline connects data loading, feature retrieval, model inference, and output writing. Pipeline parameters include storage credentials, feature configuration, model name, model type, model stage, and output location; compilation produces a YAML workflow for execution. section 5.1.2
Listing 5.11 Income classifier Kubeflow pipeline
import kfp
from kfp import dsl, compiler
fetch_data_op = kfp.components.load_component_from_file(
"components/read_data/component.yaml"
)
retrieve_features_op = kfp.components.load_component_from_file(
"components/retrieve_features/component.yaml"
)
run_inference_op = kfp.components.load_component_from_file(
"components/run_inference/component.yaml"
)
write_data_op = kfp.components.load_component_from_file(
"components/write_data/component.yaml"
)
The component listing makes the workflow structure concrete by showing separate operations for reading data, retrieving features, running inference, and writing results. section 5.1.2
