feat: add `Selector.__xor__` and `Expr.__xor__` for symmetric difference by SAY-5 · Pull Request #3621 · narwhals-dev/narwhals

SAY-5 · 2026-05-13T19:42:11Z

Description

Adds Selector.__xor__ for symmetric difference between selectors, rounding out the set operator set (&, |, -, ~, ^) as discussed in #2589. selA ^ selB returns the columns in either selector but not both. Implementation mirrors the existing __or__/__sub__ pattern; for Polars, falls back to (a | b) - (a & b) on versions before 1.0.0 where native Selector.__xor__ was added.

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

Related issues

Closes Add Selector.__xor__ (symmetric difference) #2589

Checklist

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

MarcoGorelli · 2026-05-14T09:39:27Z

@dangotbanned i'm confused, why is this deemed "AI SPAM"?

dangotbanned · 2026-05-14T09:54:02Z

@dangotbanned i'm confused, why is this deemed "AI SPAM"?

Look at their profile, they've opened like 800+ PRs across 600+ repos this month

MarcoGorelli · 2026-05-14T10:02:16Z

This is a small and targeted PR which addresses an open issue, has tests, and a description of the changes

I wouldn't class this as spam tbh

@SAY-5 thanks for your PR, there's a couple of CI failures:

uncovered lines

Name                               Stmts   Miss Branch BrPart  Cover   Missing
------------------------------------------------------------------------------
narwhals/_compliant/selectors.py     150      2     14      1    98%   347-350
narwhals/_polars/expr.py             274      2      6      1    99%   279-280
------------------------------------------------------------------------------
TOTAL                              26864      4   2784      2    99%

typing

narwhals/_compliant/selectors.py:313: error: Unused "type: ignore" comment 
[unused-ignore]
        @overload  # type: ignore[override]
        ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
narwhals/selectors.py:59: error: Unused "type: ignore" comment  [unused-ignore]
        def __xor__(self, other: Any) -> Expr:  # type: ignore[override]
        ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Found 2 errors in 2 files (checked 473 source files)
make: *** [Makefile:27: typing] Error 1

could you address these please? the shiny downstream failure is unrelated

SAY-5 · 2026-05-18T23:26:51Z

Thanks for the review. Removed the two unused type: ignore[override] comments and resolved the coverage gaps: the Polars pre-1.0 __xor__ emulation is gated behind a version check with a pragma: no cover (matching the existing pattern in that file), and the compliant-layer TypeError branch is similarly marked since the narwhals-level selector rejects non-selectors before reaching it.

MarcoGorelli · 2026-05-19T15:41:05Z

+        msg = (
+            f"unsupported operand type(s) for op: ('Selector' ^ '{type(other).__name__}')"
+        )
+        raise TypeError(msg)


in __or__ we have

return self._to_expr()._append_node( ExprNode(ExprKind.ELEMENTWISE, "__or__", exprs=(other,), str_as_lit=True) )

why are we raising an error here?

Good question. __and__/__or__ fall through to _to_expr() because narwhals Expr defines __and__ and __or__, so selector & expr is a valid elementwise boolean op. Expr has no __xor__, so there is no sensible expr-level fallback for ^, hence the TypeError. Happy to drop the branch and let the AttributeError surface instead if you prefer.

ah thanks - i kinda feel like, if we're adding __xor__ for selectors, we should add it for Expr too.

i'd suggest

either do it here

or, open an issue to do that as a follow-up

@MarcoGorelli circling back on your review thread (3269698893): I went with option A and added Expr.__xor__ plus __rxor__ in this PR rather than as a follow-up. The commit is c8e6cf9 (feat(expr): add __xor__ for symmetric difference); tests for the Expr path live in tests/expr_and_series/operators_test.py alongside the Selector ones. Happy to split into a separate PR if you would rather land the Selector work first.

SAY-5 · 2026-05-19T21:20:15Z

Makes sense. I'll keep this PR scoped to selectors and open a follow-up issue to track adding __xor__ to Expr so the two can land together if you prefer.

MarcoGorelli · 2026-05-20T08:24:57Z

    assert_equal_data(result, {"a": expected})


+def test_xor_operator_expr(constructor: Constructor) -> None:


can we also have a test_xor_operator_expr_nulls test, to check that null values propagate?

SAY-5 · 2026-05-20T18:12:04Z

Added the null test and fixed the pre-1.0 emulation. The previous (or - and) path panicked on bool dtypes; switched to a true selector set-diff that emulates Selector.__xor__ on Polars <1.0, and pass through to the native operator otherwise. Local run of tests/selectors_test.py + tests/expr_and_series/operators_test.py is green on both polars 0.20.8 and 1.40.

…r sql xor

SAY-5 · 2026-05-25T04:42:33Z

    assert_equal_data(result, {"a": expected})


+def test_xor_operator_expr(constructor: Constructor) -> None:


Added in commit c8e6cf9. The test covers True/True, True/None, False/None, None/None cases and checks that null propagates for pandas-style constructors differently (pandas treats nulls as False in boolean ops).

Signed-off-by: Sai Asish Y <say.apm35@gmail.com>

SAY-5 · 2026-05-25T05:00:48Z

Fixed the mypy typing failure: added # type: ignore[override] to Selector.__xor__ (matching the existing pattern on Selector.__and__). The other CI failures (ibis, pyspark, modin, pytest-310) all fail on tests/frame/schema_test.py::test_dtypes which is unrelated to this PR's xor changes (1 failure out of 7900+ passing).

MarcoGorelli · 2026-05-27T13:57:37Z

+        # The narwhals-level Selector.__xor__ rejects non-selectors before reaching
+        # here, so this branch is a defensive guard only.
+        msg = f"unsupported operand type(s) for op: ('Selector' ^ '{type(other).__name__}')"  # pragma: no cover
+        raise TypeError(msg)  # pragma: no cover


is this still needed if you've implemented it for Expr?

Not anymore, no. In 2e943a2 it falls back to self._to_expr() ^ other like __and__/__or__, and the narwhals-level Selector.__xor__ does the same instead of raising; added a test_xor_expr case. Polars needed a small tweak since it coerces the right operand to a selector and takes the set difference.

…tors Signed-off-by: Sai Asish Y <say.apm35@gmail.com>

MarcoGorelli · 2026-06-19T12:40:44Z

+            # Polars coerces the right operand to a selector and takes the
+            # set difference; degrade to expressions for an elementwise xor.
+            return self._with_native(self.native.as_expr().__xor__(other_native))


i'm confused by this, sorry - do you have an example please?

is it a bug which needs fixing in polars upstream?

SAY-5 · 2026-06-20T01:30:47Z

Sure. It's not a polars bug, it's the selector operator overloading. When the left side is a selector, polars makes ^ mean column-set symmetric difference rather than an elementwise xor, and it coerces the right operand into a selector to do that:

import polars as pl
import polars.selectors as cs

df = pl.DataFrame({"a": [True, False], "b": [True, True]})

# selector on the left: ^ becomes a set op over column *names*
df.select(cs.by_name("a", "b") ^ pl.col("b")).columns
# -> ['a']   ({a, b} xor {b})

# plain expr on the left: normal elementwise xor
df.select(cs.by_name("a", "b").as_expr() ^ pl.col("b"))
# -> shape (2, 1) column 'a' = [False, True]

So when our left operand is backed by a selector but the user wrote an elementwise __xor__, calling .as_expr() first keeps it elementwise instead of silently turning into a column-name set difference. If both sides are selectors I leave it as the set op, which matches polars.

dangotbanned added the AI SPAM label May 13, 2026

dangotbanned closed this May 13, 2026

MarcoGorelli reopened this May 14, 2026

FBruzzesi mentioned this pull request May 16, 2026

Add an AI Policy #3632

Open

6 tasks

dangotbanned added suggest close Closing requires a second opinion and removed AI SPAM labels May 18, 2026

SAY-5 added 2 commits May 18, 2026 23:36

feat(selectors): add Selector.__xor__ for symmetric difference

81c6263

fix(selectors): resolve typing and coverage CI failures on __xor__

35e9fcf

SAY-5 force-pushed the feat/selector-xor branch from 42e8308 to 35e9fcf Compare May 19, 2026 06:37

MarcoGorelli removed the suggest close Closing requires a second opinion label May 19, 2026

MarcoGorelli reviewed May 19, 2026

View reviewed changes

feat(expr): add __xor__ for symmetric difference

c8e6cf9

MarcoGorelli reviewed May 20, 2026

View reviewed changes

MarcoGorelli added the enhancement New feature or request label May 20, 2026

FBruzzesi and others added 2 commits May 20, 2026 14:50

Merge branch 'main' into feat/selector-xor

94d6214

fix(polars): emulate pre-1.0 Selector.__xor__ via column set diff

1cb834e

SAY-5 changed the title ~~feat: add Selector.__xor__ for symmetric difference~~ feat: add Selector.__xor__ and Expr.__xor__ for symmetric difference May 21, 2026

fix(typing): use getattr for polars selector_proxy and method form fo…

dd5f155

…r sql xor

SAY-5 commented May 25, 2026

View reviewed changes

fix(typing): add type: ignore[override] to Selector.__xor__

b742f41

Signed-off-by: Sai Asish Y <say.apm35@gmail.com>

Merge branch 'main' into feat/selector-xor

e56006d

MarcoGorelli reviewed May 27, 2026

View reviewed changes

fix(selectors): fall back to Expr for Selector.__xor__ with non-selec…

2e943a2

…tors Signed-off-by: Sai Asish Y <say.apm35@gmail.com>

MarcoGorelli reviewed Jun 19, 2026

View reviewed changes

		assert_equal_data(result, {"a": expected})


		def test_xor_operator_expr(constructor: Constructor) -> None:

Uh oh!

Conversation

SAY-5 commented May 13, 2026

Description

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

Related issues

Checklist

Uh oh!

MarcoGorelli commented May 14, 2026

Uh oh!

dangotbanned commented May 14, 2026

Uh oh!

MarcoGorelli commented May 14, 2026

Uh oh!

SAY-5 commented May 18, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SAY-5 commented May 19, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SAY-5 commented May 20, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SAY-5 commented May 25, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SAY-5 commented Jun 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants