Known issues

This page lists known issues with AutoML Tables, along with ways you can avoid or recover from these issues.

Prediction

  • Batch prediction requests with only one feature column fail.

    You must provide at least 2 feature columns with batch predictions.

BigQuery integration

  • Errors with your BigQuery tables or views shown as internal error

    When you use BigQuery as a data source or prediction result target, and there are problems with your BigQuery schema or configuration, the error might be returned in AutoML Tables as an internal error. If you get an internal error when working with BigQuery, check your BigQuery schema and configuration.

Cloud AutoML API

  • Unsupported API versions present in API endpoints and documentation

    The only version of the Cloud AutoML API supported for AutoML Tables is v1beta. Using the v1 REST or RPC endpoints to access or modify AutoML Tables objects is not supported.

Using Google Cloud Console with AutoML Tables

  • User experience with Microsoft Edge and Microsoft Internet Explorer browsers might be suboptimal.

    Microsoft Edge and Microsoft Internet Explorer do not support all features of AutoML Tables. If you are having problems, try Google Chrome, Safari, or Firefox.

Fixed issues

The following issues were listed on this page, but are no longer affecting AutoML Tables.

  • Local feature importance results are not supported with timestamp features

    Local feature importance is now fully supported for models with features of type Timestamp.

  • High latency for online prediction requests after deploying the model

    After you deploy your model, the first online prediction requests no longer show increased latency.

  • Training for longer than needed no longer degrades model quality.

    AutoML Tables automatically stops training the model when it detects that model quality is no longer improving.

  • Datasets with less than 100,000 rows no longer result in decreased model quality.

    Datasets with less than 100,000 rows can be used to train models without a significant drop in model quality. Keep in mind that more data typically results in better model quality. The minimum amount of training data is 1,000 rows.