3.14.1 The Model Catalog

3.14.1.1 The Model Catalog Table

3.14.1.2 The Model Object Catalog Table

3.14.1.3 Model Metadata

HeatWave AutoML stores machine learning models in a model catalog in HeatWave MySQL. A model catalog is an InnoDB table named MODEL_CATALOG. HeatWave AutoML creates a model catalog for any user that creates a machine learning model.

The MODEL_CATALOG table is created in a schema named ML_SCHEMA_user_name, where the user_name is the name of the owning user.

When a user creates a model, the ML_TRAIN routine creates the model catalog schema and table if they do not exist. ML_TRAIN inserts the model as a row in the MODEL_CATALOG table at the end of training.

A model catalog is accessible only to the owning user unless the user grants privileges on the model catalog to another user. This means that HeatWave AutoML routines can only use models that are accessible to the user running the routines. For information about granting model catalog privileges, see Section 3.14.10, “Sharing Models”.

A database administrator can manage a model catalog table as they would a regular MySQL table.

As of MySQL 9.0.0, HeatWave AutoML can support large models that are only limited by the amount of memory defined by the shape. The changes to the model catalog and model metadata are as follows:

• Models are not stored in the model_object column in the MODEL_CATALOG table. This column is set to NULL.

• Models are chunked and stored uncompressed in a new model_object_catalog table. Each chunk is saved with the same model_handle.

• model_metadata includes the number of chunks.

A call to one of the following routines will upgrade the model catalog, and store the model in the model_object_catalog table:

• ML_TRAIN

• ML_MODEL_LOAD

• ML_EXPLAIN

• ML_MODEL_IMPORT

• ML_MODEL_EXPORT

If the call to one of these routines is not successful or is aborted, then the previous model catalog will still be available.

As of MySQL 9.0.0, the model catalog can support these model formats:

• HWMLv1.0

  A model trained by HeatWave AutoML and stored compressed.

• HWMLv2.0

  A model trained by HeatWave AutoML and stored uncompressed.

• ONNXv1.0

  An ONNX model verified by HeatWave AutoML and stored compressed.

• ONNXv2.0

  An ONNX model verified by HeatWave AutoML and stored uncompressed.