Marcel Kornacker

Currently, images are materialized as PIL.Image.Image, which has performance problems. Switch to materializing images as cv2/numpy arrays and implement cv2 image functions.

At the moment, everything is written out as jpeg, which isn't appropriate for runtime-intensive image-generating models.

Use pytest-xdist to run tests in parallel. This requires some work, because the tests won't be able to share the store.

Adds draw_bounding_boxes() udf

1

Also: renamed functions.eval to functions.vision

Metadata for media_validation flag

1

This only contains the metadata changes, not the implementation of the on_read validation strategy.

Converting aggregate FunctionCalls to SQL

1

Also: - new class ExprDict - fixes a bug in TableVersion.insert(df), which didn't generate correct rowids - changes limit=0 to mean 'return no rows' instead of 'return all rows'