[Data] unify_schemas() not robust to empty Schemas when processing struct columns.

[anrooo]

What happened + What you expected to happen

With #53454 we now unify schemas in _bundles_to_block_list when materializing datasets. This is exposing a bug recently introduced into unify_schemas_with_validation. RefBundles can have empty schemas and in these cases the following line blows up:

ray/python/ray/data/_internal/arrow_ops/transform_pyarrow.py

Line 264 in 36cfce8

field_types = [s.field(col_name).type for s in schemas]

You'll notice that on line 234 in this file for TensorArray types we check if the name of the column is present in the schema we're iterating over before retrieving the type. A similar check is needed for structs, otherwise if the overall dataset schema contains structs this method blows up.

cc: @srinathk10, this was a regression cherry-picked into ray2.48 it seems.

Versions / Dependencies

ray2.48

Reproduction script

n/a

Issue Severity

High: It blocks me from completing my task.

[iamjustinhsu]

Hi @anrooo! Thanks for pointing this out! We are actually working on this rn, and should have a fix very soon. We'll keep you posted as soon as we can

[iamjustinhsu]

The problem should be mitigated by #55926, which was just merged. Can you verify that this works for you? The nightly builds can be found https://docs.ray.io/en/latest/ray-overview/installation.html#daily-releases-nightlies, so u might need to wait a day