A Model catalog is a metadata catalog that provides the unified interface to manage the metadata of machine learning models in a centralized way. It follows the typical Gravitino 3-level namespace (catalog, schema, and model) to manage the ML models metadata. In addition, it supports managing the versions for each model.
The advantages of using model catalog are:
The key concept of model management is to manage the path (URI) of the model. Instead of managing the model storage path physically and separately, model metadata defines the mapping relation between the model name and the storage path. In the meantime, with the support of extensible properties of model metadata, users can define the model metadata with more detailed information rather than just the storage path.
A Model catalog doesn't have specific properties. It uses the common catalog properties.
Refer to Catalog operations for more details.
Schema is the second level of the model catalog namespace, the model catalog supports creating, updating, deleting, and listing schemas.
Schema in the model catalog doesn't have predefined properties. Users can define the properties for each schema.
Refer to Schema operation for more details.
The Model catalog supports registering, listing and deleting models and model versions.
Property name | Description | Default value | Required | Immutable | Since Version |
---|---|---|---|---|---|
default-uri-name | The default URI name for the versions of the model. | (none) | No | No | 1.0.0 |
Refer to Model operation for more details.
The Model catalog supports linking, listing and deleting model versions.
Property name | Description | Default value | Required | Immutable | Since Version |
---|---|---|---|---|---|
default-uri-name | The default URI name for the model version. If set, it will override the default-uri-name property at model level. | (none) | No | No | 1.0.0 |
Refer to ModelVersion operation for more details.