[Data] Cap pandas to <3 for ray[data] and make SettingWithCopyWarning handling pandas-3-safe

[daiping8]

Description

Summary

This PR prevents pip install "ray[data]" from pulling in pandas==3.* (which currently breaks Ray Data at runtime), and hardens Ray's warning-handling code to avoid crashing if SettingWithCopyWarning is missing.

Changes

• Dependency fix: Set pandas requirement to >=1.3,<3 for ray[data] (and align ray[tune] to use the same bound) in python/setup.py.
• Runtime robustness: Update SettingWithCopyWarning lookup in:python/ray/air/util/data_batch_conversion.py python/ray/data/util/data_batch_conversion.py to gracefully degrade when the warning isn’t available.
• Dev requirements sync: Align python/requirements.txt to pandas>=1.3,<3.

How to reproduce

pip install "ray[data]==2.53.0" "pandas==3.0.0"

Execute the following code before and after applying the patch of this PR.

import numpy as np
import ray


def main() -> None:
    import pandas as pd

    ray.init()
    dfs = []
    for _ in range(50):
        dfs.append(
            pd.DataFrame(
                {
                    "tensor": [np.zeros((2, 2), dtype=np.float32) for _ in range(50)],
                    "x": np.arange(50, dtype=np.int64),
                }
            )
        )
    ds = ray.data.from_pandas(dfs)
    ds = ds.map_batches(lambda df: df, batch_format="pandas")
    ds = ds.repartition(8)
    ds = ds.materialize()
    print("count =", ds.count())

    ray.shutdown()


if __name__ == "__main__":
    main()

Related issues

Closes #60402"

[gemini-code-assist]

Code Review

This pull request correctly caps the pandas dependency to <3 to avoid runtime breakages in Ray Data and makes the SettingWithCopyWarning handling more robust. The changes to setup.py and requirements.txt are appropriate.

My review includes suggestions to improve the maintainability of the warning handling code in data_batch_conversion.py by simplifying the logic and addressing code duplication. I've also pointed out a duplicated dependency in requirements.txt that could be cleaned up.

[aslonnie]

need @ray-project/ray-data comments.

[bveeramani]

@daiping8 after fix to SettingWithCopyWarning, does Ray Data still crash with pandas==3.0.0? If so, what's the traceback?

[daiping8]

@daiping8 after fix to SettingWithCopyWarning, does Ray Data still crash with pandas==3.0.0? If so, what's the traceback?

After fixing the warning, it no longer crashes. You can run the test cases in the PR Description.

[aslonnie]

it no longer crashes. You can run the test cases in the PR Description.

nice, so we can close this PR now? feel like ray data team is trying to support running with pandas>=3

[daiping8]

it no longer crashes. You can run the test cases in the PR Description.

nice, so we can close this PR now? feel like ray data team is trying to support running with pandas>=3

I think it can be merged, but we'd better check with others for their opinions.

Pandas 3.0.0 was released on January 21, 2026. I think it will take some time to identify bugs and formulate a support plan for Ray Data.

[aslonnie]

we do not upper bound version if it is not broken. that is library user's choice and freedom.

I think it will take some time to identify bugs and formulate a support plan for Ray Data.

user can cap with additional requirement constraints themselves if desired. or use our images / tested compiled lock files if stability is the priority.

check with others for their opinions.

@bveeramani what do you think?

it can be merged

just in case that balaji/data team accepts this, if this PR wants to be merged, it needs to update the requirement / lock files in the repo. no rush though.

[bveeramani]

@daiping8 could you help me understand -- other than the breaking change to SettingWithCopyWarning, what are the other changes in pandas 3 that might affect Ray Data?

[aslonnie]

(waiting for @ray-project/ray-data to understand the context and make a decision)

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.