com.linkedin.feathr.common

Class PegasusFeatureTypeResolver

java.lang.Object

com.linkedin.feathr.common.PegasusFeatureTypeResolver

public class PegasusFeatureTypeResolver
extends java.lang.Object

This class maps from the pegasus models for feature types to Frame's common domain models for feature types and vice versa. This creates a layer of indirection from the feature definition models expressed in Pegasus to the domain models used by the frame's runtime engine (e.g. frame-online and frame-offline)

Method Summary

Modifier and Type	Method and Description
static PegasusFeatureTypeResolver	getInstance()
java.util.Optional<java.lang.Integer>	resolveEmbeddingSize(com.linkedin.feathr.compute.FeatureVersion featureVersion) Deprecated.
FeatureTypeConfig	resolveFeatureType(com.linkedin.feathr.compute.FeatureVersion featureVersion) Resolves the FeatureTypeConfig from the the pegasus FeatureVersion model.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Detail

getInstance

public static PegasusFeatureTypeResolver getInstance()

resolveFeatureType

public FeatureTypeConfig resolveFeatureType(com.linkedin.feathr.compute.FeatureVersion featureVersion)

Resolves the FeatureTypeConfig from the the pegasus FeatureVersion model. It's based on the following mapping rules: - if `type` is TENSOR without `format` field, it is a FML tensor type - if `type` is TENSOR with `format`, it is a Tensor feature type with FeatureTypeConfig in the feature definition - if `type` is non-TENSOR without `format`, it is a legacy type - if `type` is non-TENSOR with `format`, it is a legacy type with the format storing other info like embedding size that can be resolved using resolveEmbeddingSize(FeatureVersion)

resolveEmbeddingSize

@Deprecated
public java.util.Optional<java.lang.Integer> resolveEmbeddingSize(com.linkedin.feathr.compute.FeatureVersion featureVersion)

Deprecated.

Resolves the possible SWA embedding size from the pegasus FeatureVersion model. The embedding size is valid only when the feature is a possible embedding feature (1-d vector), which means the feature type can only be DENSE_VECTOR, or TENSOR, or UNSPECIFIED. Meanwhile, the input FeatureVersion should have valid format information: 1) the format filed exists and is not null, 2) the shape size is 1. The API is scheduled to be deprecated after dropping legacy feature type support in Frame, after which the embedding size information will always be inside the FeatureTypeConfig built from resolveFeatureType(com.linkedin.feathr.compute.FeatureVersion). Warning: this should be only used when you know the feature is an embedding feature.