[core][tune] fix RayTaskError (de)serialization logic

[matthewdeng]

Summary

Fixes a (de)serialization issue in RayTaskError that happens when the original cause has its own __reduce__ method defined.

Closes #54379.

Details

The existing implementation relies on the definition of BaseException.__reduce__:

# BaseException implements a __reduce__ method that returns
# a tuple with the type and the value of self.args.
# https://stackoverflow.com/a/49715949/2213289

However, since it's also subclassing the cause_cls, if __reduce__ is also defined there then that will override the original definition from BaseException.

This surfaced in #54379 because Ray Train V2 implements exception classes with custom __reduce__ definitions (e.g. TrainingFailedError).

The fix is to explicitly define __reduce__. Also removed the comment since it's no longer an assumption that's needed to understand the logic.

Testing

Added test_errors() test case in ray/tune/tests/test_train_v2_integration.py.

[Copilot]

Pull Request Overview

Fix serialization of RayTaskError when the cause exception defines its own __reduce__ and add an integration test for error capture.

• Explicitly define __reduce__ in dynamically generated dual-inheritance exceptions
• Remove outdated comment about relying on BaseException.__reduce__
• Add test_errors in test_train_v2_integration.py and a new Ray fixture

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
python/ray/exceptions.py	Add __reduce__ methods in both dual-exception closures to fix de/serialization
python/ray/tune/tests/test_train_v2_integration.py	Introduce ray_start_4_cpus fixture and test_errors to validate error capture

Comments suppressed due to low confidence (2)

python/ray/tune/tests/test_train_v2_integration.py:114

• The test only exercises a standard RuntimeError, but doesn't verify serialization for exceptions with custom __reduce__ methods (e.g., TrainingFailedError). Add a test case for a custom exception subclass to ensure the fix covers that scenario.

def test_errors(ray_start_4_cpus):

python/ray/tune/tests/test_train_v2_integration.py:17

• Using a function-scoped fixture for ray_start_4_cpus causes Ray to init/shutdown for each test, which may slow down the test suite. Consider using scope="module" to reduce overhead.

@pytest.fixture()

[Copilot]

This __reduce__ implementation is duplicated in a second closure at line 201. Consider extracting it into a shared helper or mixin to adhere to DRY principles.

Suggested change

	return (cls, self.args)
	return _reduce_helper(cls, self.args)