Driverless AI Migration Guide

This guide outlines recommendations and known migration issues when upgrading Driverless AI to a newer version.

Note: This guide is based on the Driverless AI migration test suites. Complete coverage in production is not guaranteed. You may encounter issues during migration that are not covered here. If that happens, contact support@h2o.ai for further diagnostics.

Migrating to Driverless AI 1.11.x

Driverless AI 1.11.1 upgraded its internal Python runtime from Python 3.8 to Python 3.11, which may cause migration issues. These issues arise from compatibility differences between the two Python versions.

Known Migration Issues

  1. TensorFlowModel Issues: If the final model is a TensorFlowModel, it is recommended to retrain the model from scratch, as the prediction functionality is not performing as expected in version 1.11.x.

    • Known Errors: - AttributeError: 'Functional' object has no attribute '_self_tracked_trackables' - AttributeError: 'Model' object has no attribute '_dynamic'

  2. Migrated Experiments: If old experiments used custom components (such as transformers, models, or other recipes added during training), the migrated experiments may break due to conflicts between dependencies in the user-defined recipes and those used in Driverless AI.

  3. Recipes: Recipes that depend on specific package versions might fail if those package versions are no longer included in Driverless AI. Update the package versions as indicated by the error messages.