4.9 AutoML Limitations

The following limitations apply to AutoML.

Text Handling Limitations

• AutoML only supports datasets in the English language.

• MySQL HeatWave AutoML does not support TEXT columns with NULL values.

• MySQL HeatWave AutoML does not support target columns (a column with ground truth values) with a TEXT data type.

• MySQL HeatWave AutoML does not support recommendation tasks with columns that have a TEXT data type.

• For the forecasting task, endogenous_variables cannot be in TEXT.

Account Name Limitations

• The ML_TRAIN routine does not support MySQL user names that contain a period. For example, a user named 'joe.smith'@'%' cannot run the ML_TRAIN routine. The model catalog schema created by the ML_TRAIN procedure incorporates the user name in the schema name (for example., ML_SCHEMA_joesmith), and a period is not a permitted schema name character.

Memory Limitations

• The table used to train a model cannot exceed 10 GB, 100 million rows, or 1017 columns.

Routine and Query Limitations

• ML_EXPLAIN_TABLE and ML_PREDICT_TABLE are compute intensive processes, with ML_EXPLAIN_TABLE being the most compute intensive. Limiting operations to batches of 10 to 100 rows by splitting large tables into smaller tables is recommended. Use batch processing with the batch_size option. See the following to learn more:

  • ML_PREDICT_TABLE

  • ML_EXPLAIN_TABLE

• ML_EXPLAIN, ML_EXPLAIN_ROW, and ML_EXPLAIN_TABLE routines limit explanations to the 100 most relevant features.

• The ML_PREDICT_TABLE ml_results column contains the prediction results and the data. This combination must be less than 65,532 characters.

• Concurrent MySQL AI analytics and AutoML queries are not supported. An AutoML query must wait for MySQL AI analytics queries to finish, and vice versa. MySQL AI analytics queries are given priority over AutoML queries.

• The ML_PREDICT_ROW, ML_MODEL_IMPORT, and ML_MODEL_EXPORT routines are not supported with the TwoTower recommendation model.

Other Limitations

• If you delete a recommendation model trained with the TwoTower model from the model catalog, you need to run a Delete Model API to manage the generated embedding tables.