[Question] How to store arbitrary values in MLMD?

[gabrielwen]

Users need to pass configurations between training and serving, e.g. feature set used for training should also be used for serving. How do we do it? Should we use values or is there a place for us to store free form JSON?

[zhenghuiwang]

The type system of underlying MLMD server is extensible. You can

1. register the custom type for the configurations via POST /api/v1alpha1/artifact_types
2. Read and write the config as an artifact: Read GET /api/v1alpha1/artifact_types/{name}/artifacts/{id} Write POST /api/v1alpha1/artifact_types/{parent}/artifacts

The values is for metrics only not for any custom data.

[jtfogarty]

/kind feature /priority p2

[jtfogarty]

/area engprod