chore: Remove `level` from main and v2 classes init by FBruzzesi · Pull Request #3558 · narwhals-dev/narwhals

FBruzzesi · 2026-04-19T22:07:25Z

Description

Not really user facing but I had it started some time ago, so why not finish it 😇

What type of PR is this? (check all applicable)

Related issues

Closes Remove level, DataFrameLike from main, v2 #2932
Parent of chore: Isolate interchange-level support #3721

Checklist

Code follows style guide (ruff)
Tests added
Documented the changes

… typing

FBruzzesi · 2026-04-19T22:35:08Z

@@ -1922,7 +1922,7 @@ def convert_str_slice_to_int_slice(


 def inherit_doc(
-    tp_parent: Callable[P, R1], /
+    tp_parent: Callable[..., R1], /


@dangotbanned you might have a better way of doing this but here what happened - by changing the main namespace __init__ signatures, now the V1 is overriding them and the arguments are different, to so the child ones are not bind to the parent ones - hence removing the P from the parent here.

This meant that return DataFrame(self, level="interchange") was resulting in a typing issue

So I fixed that now 🎉

But it made me realize we don't need it for any Series, DataFrame, LazyFrame if we just remove the constructor override?

narwhals/src/narwhals/stable/v1/__init__.py

Lines 133 to 137 in cd256d6

@inherit_doc(NwDataFrame)

def __init__(self, df: Any) -> None:

assert df._version is Version.V1 # noqa: S101

super().__init__(df)

That assertion is pointless - as it checks a ClassVar, and that's it AFAICT?

partially reverts (a645201) See #3558 (comment)

dangotbanned · 2026-06-22T14:54:59Z

I'm not sure how exactly, but I think the marimo failures mean that they're (probably indirectly) depending on ._level? 🤔

The error output isn't helping narrow it down though

FAILED tests/_data/test_preview_column.py::test_get_column_preview_for_duckdb - AssertionError: assert None is not None
 +  where None = DataColumnPreviewNotification(chart_spec=None, chart_code=None, error=None, missing_packages=None, stats=ColumnStats(t...=0.502518907629606, true=None, false=None, p5=0.0, p25=0.0, p75=1.0, p95=1.0), table_name='tbl', column_name='outcome').chart_spec
FAILED tests/_data/test_preview_column.py::test_get_column_preview_for_duckdb_categorical - AssertionError: assert None is not None
 +  where None = DataColumnPreviewNotification(chart_spec=None, chart_code=None, error=None, missing_packages=None, stats=ColumnStats(t...one, std=None, true=None, false=None, p5=None, p25=None, p75=None, p95=None), table_name='tbl', column_name='category').chart_spec
FAILED tests/_data/test_preview_column.py::test_get_column_preview_for_duckdb_date - AssertionError: assert None is not None
 +  where None = DataColumnPreviewNotification(chart_spec=None, chart_code=None, error=None, missing_packages=None, stats=ColumnStats(t...std=None, true=None, false=None, p5=None, p25=None, p75=None, p95=None), table_name='date_tbl', column_name='date_col').chart_spec
FAILED tests/_data/test_preview_column.py::test_get_column_preview_for_duckdb_datetime - AssertionError: assert None is not None
 +  where None = DataColumnPreviewNotification(chart_spec=None, chart_code=None, error=None, missing_packages=None, stats=ColumnStats(t..., true=None, false=None, p5=None, p25=None, p75=None, p95=None), table_name='datetime_tbl', column_name='datetime_col').chart_spec
FAILED tests/_data/test_preview_column.py::test_get_column_preview_for_duckdb_bool - AssertionError: assert None is not None
 +  where None = DataColumnPreviewNotification(chart_spec=None, chart_code=None, error=None, missing_packages=None, stats=ColumnStats(t...ne, std=None, true=50, false=50, p5=None, p25=None, p75=None, p95=None), table_name='bool_tbl', column_name='bool_col').chart_spec
= 5 failed, 2402 passed, 10 skipped, 26 xfailed, 1 xpassed, 49 warnings in 65.09s (0:01:05) =

dangotbanned · 2026-06-22T17:23:52Z


-            return DataFrameV1(self, level="interchange")  # type: ignore[no-any-return]
-        return self._version.lazyframe(self, level="lazy")
+            return DataFrameV1(self)  # type: ignore[no-any-return]


Right I think I've found the issue for (#3558 (comment))

The repro is this:

import narwhals as nw import narwhals.stable.v1 as nw_v1 # noqa: F811 native_duck = nw.from_dict({"a": [1, 2, 3]}, backend="polars").lazy("duckdb").to_native() v1_duck = nw_v1.from_native(native_duck) nw_v1.get_level(v1_duck)

"full"

The issue is that I've misunderstood what these guys are for:

narwhals/src/narwhals/stable/v1/__init__.py

Lines 926 to 933 in cd256d6

from narwhals._interchange.dataframe import InterchangeFrame

from narwhals._interchange.series import InterchangeSeries

ensure_type(obj, DataFrame, Series, LazyFrame)

if isinstance(obj, DataFrame):

return "interchange" if isinstance(obj._compliant, InterchangeFrame) else "full"

if isinstance(obj, Series):

return "interchange" if isinstance(obj._compliant, InterchangeSeries) else "full"

I thought that we'd have used InterchangeFrame, e.g. here:

narwhals/src/narwhals/translate.py

Lines 456 to 468 in cd256d6

# Interchange protocol

if version is Version.V1 and supports_dataframe_interchange(native_object):

from narwhals._interchange.dataframe import InterchangeFrame

if eager_only or series_only:

if not pass_through:

msg = (

"Cannot only use `series_only=True` or `eager_only=False` "

"with object which only implements __dataframe__"

)

raise TypeError(msg)

return native_object

return Version.V1.dataframe(InterchangeFrame(native_object))

I'm still pretty confused why we aren't doing that, and were instead passing around a string 🤔

Okay more digging...

The tests added from #1266 appear to have been invalid since #1725.

On main (71a0204)

import duckdb import narwhals as nw import narwhals.stable.v1 as nw_v1 # noqa: F811 native_duck = nw.from_dict({"a": [1, 2, 3]}, backend="polars").lazy("duckdb").to_native() v1_duck:nw_v1.DataFrame[duckdb.DuckDBPyRelation] = nw_v1.from_native(native_duck) print(v1_duck._compliant.__class__.__name__, repr(nw_v1.get_level(v1_duck))) v1_duck.select(nw_v1.col("a").max())

DuckDBLazyFrame 'interchange' ┌──────────────────┐ |Narwhals DataFrame| |------------------| | ┌───────┐ | | │ a │ | | │ int64 │ | | ├───────┤ | | │ 3 │ | | └───────┘ | └──────────────────┘

This looks like more than interchange-only support?

@MarcoGorelli 👋

Just wanna be sure you see this, since it looks like a 🐛 that's been hiding in interchange only support for a long time

is this issue that it's allowing select(nw_v1.col("a").max()), whereas it should be erroring?

if so, that doesn't really look like an issue? i don't think interchange level was ever meant to mean "only interchange"

That said, it's still not clear to me what marimo is relying on that this PR is breaking leading to their CI to fail

The CI failure is because of how I tried to factor out level:

if isinstance(obj, DataFrame): return "interchange" if isinstance(obj._compliant, InterchangeFrame) else "full" # ^^^^^^^^^^^^^^^^^ (me not realising that this was a `DuckDBPyRelation`)

That can be fixed by changing the condition to use both Version and Implementation.

But fixing that doesn't change (#3558 (comment)) - which I thought was a big issue

That can be fixed

As promised (fix: try a different level refactor?) does fix (marimo)

As an experiment

I'm gonna branch off here (#3721) and see how bad downstream breaks if we enforce (#3558 (comment))

I think the answer will be a lot, but who knows 😅

but I am confident that a DuckDBPyRelation is not a DataFrame.

not sure what this means tbh, could you explain please?

Show thread

@MarcoGorelli
I don't think interchange level was ever meant to mean "only interchange"

@dangotbanned
Maybe I've misunderstood, and this never meant that:

link to docs

but I am confident that a DuckDBPyRelation is not a DataFrame.

@MarcoGorelli
not sure what this means tbh, could you explain please?

not sure what this means tbh, could you explain please?

Sure.

If you asked me a week ago what interchange level meant; I would've described a subset of this.
If I did my homework; I'd have taken a look at InterchangeFrame and describe that.

But both of those are wrong.

So now, if I want to describe what it means - what am I left with?

To a type checker it is v1.DataFrame[DuckDBPyRelation]

At runtime, it's anything that doesn't raise an error?

Summary

I don't understand what is being promised by interchange-level

That concerns me because we are supporting it indefinitely

Our test suite does not cover anything beyond IO, selecting columns by name, and dtypes

Updated

See (#3721), (#3721 (comment))

... but this reveals another bug 😭 Mentioned in (#3558 (comment))

Going scorched-earth to find out who's using it downstream Part 1 of (#3558 (comment))

I'm dismissing my review, because I left it before finding (#3558 (comment)), (#3558 (comment))

Just want to avoid this merging before finishing (#3721)

Really sorry, I shouldn't have approved while we still had downstream CI running

FBruzzesi added 2 commits April 19, 2026 23:22

Move DataFrameLike protocol into _interchange/dataframe.py, and check…

e9b912e

… typing

remove level from main and v2 classes __init__

de815a7

FBruzzesi requested a review from dangotbanned April 19, 2026 22:07

FBruzzesi added the internal label Apr 19, 2026

FBruzzesi added 2 commits April 20, 2026 00:23

inherit_doc typing

a645201

coverage

415d241

FBruzzesi commented Apr 19, 2026

View reviewed changes

FBruzzesi and others added 4 commits April 20, 2026 00:37

make mypy happy

0791344

merge main

1e0f9b3

merge main

fc10c5e

Merge branch 'main' into chore/issue-2932-rm-level

b0f2586

dangotbanned self-assigned this Jun 22, 2026

dangotbanned added 4 commits June 22, 2026 13:37

test: cover the desired behavior

8347fb6

chore: remove _level from main

dadf498

refactor: remove _level & level from everything

4577fa5

revert(typing): "inherit_doc typing"

cd256d6

partially reverts (a645201) See #3558 (comment)

This comment was marked as outdated.

Sign in to view

dangotbanned removed their assignment Jun 22, 2026

dangotbanned marked this pull request as draft June 22, 2026 14:55

dangotbanned reviewed Jun 22, 2026

View reviewed changes

dangotbanned reviewed Jun 23, 2026

View reviewed changes

Comment thread src/narwhals/typing.py Outdated

chore: avoid diff on narwhals.typing

4015454

dangotbanned added needs investigation accepted labels Jun 27, 2026

dangotbanned added 2 commits June 27, 2026 16:35

fix: try a different level refactor?

52b490a

... but this reveals another bug 😭 Mentioned in (#3558 (comment))

Merge branch 'main' into chore/issue-2932-rm-level

39a320d

dangotbanned added a commit that referenced this pull request Jun 27, 2026

chore!: Remove interchange (temporary)

e745e37

Going scorched-earth to find out who's using it downstream Part 1 of (#3558 (comment))

dangotbanned mentioned this pull request Jun 27, 2026

chore: Isolate interchange-level support #3721

Open

9 tasks

Merge branch 'main' into chore/issue-2932-rm-level

53a4bbf

Merge branch 'main' into chore/issue-2932-rm-level

f649edf

	@inherit_doc(NwDataFrame)
	def __init__(self, df: Any) -> None:
	assert df._version is Version.V1 # noqa: S101
	super().__init__(df)

	from narwhals._interchange.dataframe import InterchangeFrame
	from narwhals._interchange.series import InterchangeSeries

	ensure_type(obj, DataFrame, Series, LazyFrame)
	if isinstance(obj, DataFrame):
	return "interchange" if isinstance(obj._compliant, InterchangeFrame) else "full"
	if isinstance(obj, Series):
	return "interchange" if isinstance(obj._compliant, InterchangeSeries) else "full"

	# Interchange protocol
	if version is Version.V1 and supports_dataframe_interchange(native_object):
	from narwhals._interchange.dataframe import InterchangeFrame

	if eager_only or series_only:
	if not pass_through:
	msg = (
	"Cannot only use `series_only=True` or `eager_only=False` "
	"with object which only implements __dataframe__"
	)
	raise TypeError(msg)
	return native_object
	return Version.V1.dataframe(InterchangeFrame(native_object))

Uh oh!

Conversation

FBruzzesi commented Apr 19, 2026 • edited by dangotbanned Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

What type of PR is this? (check all applicable)

Related issues

Checklist

Uh oh!

FBruzzesi Apr 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dangotbanned Jun 22, 2026

Choose a reason for hiding this comment

Uh oh!

This comment was marked as outdated.

Uh oh!

dangotbanned commented Jun 22, 2026

Uh oh!

dangotbanned Jun 22, 2026

Choose a reason for hiding this comment

Uh oh!

dangotbanned Jun 22, 2026

Choose a reason for hiding this comment

Uh oh!

dangotbanned Jun 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

On main (71a0204)

Uh oh!

dangotbanned Jun 23, 2026

Choose a reason for hiding this comment

Uh oh!

MarcoGorelli Jun 25, 2026

Choose a reason for hiding this comment

Uh oh!

dangotbanned Jun 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dangotbanned Jun 27, 2026

Choose a reason for hiding this comment

Uh oh!

dangotbanned Jun 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

As an experiment

Uh oh!

MarcoGorelli Jun 27, 2026

Choose a reason for hiding this comment

Uh oh!

dangotbanned Jun 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Summary

Updated

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

FBruzzesi commented Apr 19, 2026 •

edited by dangotbanned

Loading

FBruzzesi Apr 19, 2026 •

edited

Loading

dangotbanned Jun 22, 2026 •

edited

Loading

On `main` (`71a0204`)

dangotbanned Jun 25, 2026 •

edited

Loading

dangotbanned Jun 27, 2026 •

edited

Loading

dangotbanned Jun 27, 2026 •

edited

Loading