flyteplugins-pandera 2.3.3

pandera plugin for flyte

Flyte Pandera Plugin

flyteplugins-pandera adds support for pandera.typing.pandas.DataFrame, pandera.typing.polars.DataFrame / LazyFrame, and pandera.typing.pyspark_sql.DataFrame in Flyte v2.

Install:

pip install flyteplugins-pandera 'pandera[pandas]'            # pandas only
pip install flyteplugins-pandera 'pandera[polars]' flyteplugins-polars   # Polars + structured dataset I/O
pip install flyteplugins-pandera 'pandera[pyspark]' flyteplugins-spark     # PySpark SQL + parquet I/O

For PySpark, structured dataset serialization uses Flyte’s DataFrameTransformerEngine parquet handlers from flyteplugins-spark (register Spark encoders/decoders alongside this plugin).

At runtime, the plugin:

1. delegates dataframe IO to Flyte's DataFrameTransformerEngine,
2. validates data with pandera schemas, and
3. writes a validation report to flyte.report.

Validation always runs on every encode/decode. Report tabs are suppressed automatically when Flyte is only moving literals across a nested-task boundary (parent task encoding child inputs, or materializing a child’s outputs inside the parent). The SDK sets TaskContext.in_driver_literal_conversion on the active task (check with flyte.ctx() when non-None, then .in_driver_literal_conversion) so you see one report per dataframe on the task that actually produced or consumed it as task body I/O, not extra tabs on the orchestrating “driver” task.

Troubleshooting

If logs show “Unsupported Type pandera.typing… Flyte will default to use PickleFile”, the pandera transformer was not registered:

• Install the plugin in every environment (local runner and task image): pip install flyteplugins-pandera.
• Flyte loads flyte.plugins.types during flyte.initialize() and on first TypeEngine use; confirm the distribution is installed (import importlib.metadata as m; print(list(m.entry_points(group="flyte.plugins.types")))).
• Import order: import your pandera.typing.* modules before plugin registration runs in files that run early (tests, __init__.py). Loading the plugin before pandera can leave two different pandera.typing.pandas.DataFrame (or polars container) class objects in the process; TypeEngine would only know about one of them, so annotations on the other fall through to pickle / the generic handler.