Drop support for Python `3.9`

[dangotbanned]

Description

3.9 EOL is scheduled for 2025-10-31.

There's some select features mentioned in (#2084 (comment)), and the full What’s New In Python 3.10 for more info.

Probably will make sense to time this around/shortly after (#3184).

Related

Meta

• 3.8
  • Tracking: Drop support for python 3.8 #2084
  • Add Python version support policy #2638
• 3.14
  • Tracking: Python 3.14 support #3184
• fix(deps): cudf to cudf-cu12 in pyproject.toml #3405

Critical Bugs

Fixes have been available since v2.13.0

• bug: join discrepancies when null values are in the joining keys #3307
• bug: when-then.over incorrect results for pandas/pyarrow #3311
• Incorrect results/errors for partitioned over() with nulls #3300

Backends

• DEPS: Drop Python 3.10 pandas-dev/pandas#62066
• feat(python): Drop Python 3.9 support pola-rs/polars#25984
• GH-47443: [Python][Packaging] Drop Python 3.9 support apache/arrow#47478
• Drop Python 3.10 support rapidsai/cudf#21225
• [SPARK-52561][DOCS][FOLLOW-UP] Update migration guide about Python 3.9 dropped apache/spark#51631
• [SPARK-54213][PYTHON][CONNECT] Remove Python 3.9 from Spark Connect apache/spark#52911
• Drop support for Python 3.9 dask/dask#11245
• ibis-project/ibis@aa16d08
• duckdb
• sqlframe
• modin ¹

Ecosystem

As a project aimed at the maintainers of other projects - tooling matters too!

• Remove support for Python 3.9 pytest-dev/pytest#13724
• Drop support for running with Python 3.9 python/mypy#20156
• Drop support for Python 3.9 (May 2026) python/typeshed#13782
• https://github.com/pypa/hatch/releases/tag/hatch-v1.16.0

Meta TODO

• Figure out new upstream minimums (see 3.8 drop table)

Footnotes

1. modin appears to have only been receiving bug-fixes by Snowflake for a while now modin-project/modin/releases. ↩

[FBruzzesi]

Started a branch: deps/bump-py310

Mostly stayed away from using match statements. There might be a few cases, but I couldn't find it so more advantageous compared to if-elif-else blocks we have, since there are no cases of pattern matching.

[dangotbanned]

@FBruzzesi !!!!!

I've been super keen to start using match in #2572 - so I'm very excited about this 😄

[MarcoGorelli]

i realise i'll make my unpopular by saying this, but i'd suggest delaying this for another 6 months - in practice it's not always easy for people to upgrade, and critical bugs like #3307, #3311, #3300 wouldn't reach people who can't (yet) upgrade

[FBruzzesi]

i realise i'll make my unpopular by saying this, but i'd suggest delaying this for another 6 months - in practice it's not always easy for people to upgrade, and critical bugs like #3307, #3311, #3300 wouldn't reach people who can't (yet) upgrade

I think that's ok - in my opinion the most relevant feature in 3.10 is match statement and pattern matching, which we don't really have much of as of now.

The other two py 3.10 features I could find use of are:

• no need to implement our version of zip_strict
• remove Union[...], import TypeAlias directly from typing module

---

Regarding critical bugs, those can always happen and it's hard to anticipate them, at a certain point we will need to make the cut

[dangotbanned]

@MarcoGorelli 2/3 of those issues (#3204 (comment)) are resolved, are you expecting the other to take 6 months?

[MarcoGorelli]

at a certain point we will need to make the cut

totally agree 👍 6 months is somewhat arbitrary, but it kinda feels right for a relatively young project like ours

[dangotbanned]

The other two py 3.10 features I could find use of are:

The other part of this issue (which I didn't follow up on)

Meta TODO

• Figure out new upstream minimums ([see 3.8 drop table](https://github.com/

Another motivating factor is we (may) be able to remove some workarounds/NotImplementedError paths for old versions

[FBruzzesi]

@dangotbanned you can already see the bumps in the branch. The TL;DR is:

• pandas -> 1.3.4 -> removed a few test failure for old versions and is_dtype_numpy_nullable definition block
• modin -> I arbitrarily bumped it to 0.22.0, also based on the "analysis" in build: determine and test minimum version for Modin #817

For indirect dependencies in GHA: numpy==1.21.4 and scipy==1.7.3, everything else stays the same

[dangotbanned]

pandas -> 1.3.4 ->

ooh nice that also removes this

narwhals/narwhals/_pandas_like/expr.py

Lines 330 to 338 in cd8abc8

	elif function_name.startswith("ewm"):
	if self._implementation.is_pandas() and (
	self._implementation._backend_version()
	) < (1, 2): # pragma: no cover
	msg = (
	"Exponentially weighted calculation is not available in over "
	f"context for pandas versions older than 1.2.0, found {self._implementation._backend_version()}."
	)
	raise NotImplementedError(msg)

But this one you already found is pretty big - it just only shows up in tests, rather than being in our code.
IIRC that was an issue in an earlier version of #2528

narwhals/tests/frame/with_row_index_test.py

Lines 38 to 45 in cd8abc8

	def test_with_row_index_lazy(
	constructor: Constructor, order_by: str | Sequence[str], expected_index: list[int]
	) -> None:
	if (
	"pandas" in str(constructor) and PANDAS_VERSION < (1, 3) and order_by == "abc"
	): # pragma: no cover
	reason = "ValueError: first not supported for non-numeric data."
	pytest.skip(reason=reason)

[raisadz]

Related to this discussion #3332 (comment).