[RFC] [python-package] remove h2o `datatable` support?

[jameslamb]

Summary

Support for the h2o's datatable library was added to LightGBM 5.5+ years ago, in #1970.

Proposing here that lightgbm:

• issue a deprecation warning for the next 2-3 releases whenever datatable is used
• permanently remove datatable support 2-3 releases from now

Motivation

That project seems to be abandoned:

• last commit and PyPI release was 10 months ago:
  • https://github.com/h2oai/datatable
  • https://pypi.org/project/datatable/
• seems that h2o has moved on to something new called H2OFrame in the h20-py package:
  • https://docs.h2o.ai/h2o/latest-stable/h2o-py/docs/frame.html#h2oframe
  • https://github.com/h2oai/h2o-3/blob/6191acc115df654c5964225480d7d4542c22bed0/h2o-py/h2o/frame.py#L39

In those 5.5 years since #1970, the only bug reports / feature requests received about datatable support have been from one person working for h2o... and the last of those was 4 years ago:

• (Sep 2020) Support h2o datatable and numpy types, including for categorical types #3386
• (Feb 2019) implement datatable ingest directly into lightgbm #2003
• (Jan 2019) memory leak with H2O's DataTable #1968

And in all that time, I don't think we have ever tested against datatable in CI.

Description

Doing this would simplify the Python package, making it easier for others to contribute.

It'd also make it more manageable to add support for newer, more popular input formats like polars (#6204).

See @trivialfis's summary of the current state of supporting data frame libraries at dmlc/xgboost#10554 (comment) ... I agree with it.

References

I am not proposing here that lightgbm should support H2OFrame... Dask doesn't, XGBoost doesn't, scikit-learn doesn't... and I think our limited time and attention here would be better spent on more widely-used input formats, like polars.

[jameslamb]

@guolinke @shiyu1994 @StrikerRUS @jmoralez @borchero @btrotta please let me know what you think whenever you have time

[StrikerRUS]

I'm +1 for dropping support of datatable. Especially given that so called "support" is simple .to_numpy() method call 🙃

[trivialfis]

Thank you for the ping. Sounds good to me considering that there's no new commit to the project now.

[jmoralez]

I'm +1 as well

[guolinke]

I am +1

[borchero]

I'm in favor of removing as well ✅

[jameslamb]

Thank you all for the quick responses! I'll put up a PR adding a deprecation warning.