[Enh]: Support for top-level function `pl.repeat()`

[MariusMerkleQC]

We would like to learn about your use case. For example, if this feature is needed to adopt Narwhals in an open source project, could you please enter the link to it below?

Are you planning to support pl.repeat()? I have a function that returns a pl.Expr: return pl.repeat(pl.lit(False), pl.len()) which I'd like to narwhalify. However, I cannot find a nw.repeat() function among the top-level functions specified in the API documentation.

Please describe the purpose of the new feature or describe the problem to solve.

See above

Suggest a solution if possible.

No response

If you have tried alternatives, please describe them below.

I tried returning a single pl.lit(False) or a pl.Series() but this is not equivalent.

Additional information that may help us understand your needs.

No response

[FBruzzesi]

Hey @MariusMerkleQC thanks for raising the issue. This is definitely in scope and @dangotbanned brought it up in the past (see #2839 (comment)), but we never had an issue for tracking it.

---

Unrelated to the feature itself but related to your use case: how does pl.repeat(pl.lit(False), pl.len()) differ from pl.lit(False, dtype=bool)? They are both lazy and using pl.len() means that you don't know the length a priori. To me this sounds like the case mentioned in the polars API reference:

If you want to construct a column in lazy mode and do not need a pre-determined length, use lit() instead.

[MariusMerkleQC]

Right, but when I want to apply pl.repeat() in a .select() statement, and then pl.lit(False) returns length 1, while I'd like to keep the length of the data frame. This work in a .with_columns() statement, though. See below (the code passes for polars=1.32.0):

import polars as pl
from polars.testing import assert_frame_not_equal, assert_frame_equal


def f1():
    return pl.repeat(pl.lit(False), pl.len()).alias("new_col")

def f2():
    return pl.lit(False).alias("new_col")


df = pl.DataFrame(data={"col": [1, 2, 3]})

assert_frame_equal(df.with_columns(f1()), df.with_columns(f2()))
assert_frame_not_equal(df.select(f1()), df.select(f2()))

[MarcoGorelli]

Agree that this should be in scope, but until then I think it'll work to do nw.lit(False).over(nw.lit(1))

[dangotbanned]

Thanks for summoning me @FBruzzesi! 😄
Agreed we should have this, just have two notes to add.

nw.when should we do this?

In some cases, we've already got it implemented internally/identified the native API:

narwhals/narwhals/_arrow/utils.py

Lines 101 to 102 in 5309dd5

	def zeros(n: int, /) -> pa.Int64Array:
	return pa.repeat(0, n)

My suggestion would be to wait for #2895 to land (optionally #2917) as there's some crossover on both the internal + public APIs.
So we could have a nice unified thing for eager=... (#2895 (comment)) and consistent rules around where we can use these kinds of expressions 🙂

Other use-cases?

For downstream that still directly depends on numpy, I think the pl.ones and pl.zeros sugar provide a nice transitional API towards using expressions
code search r"np.(repeat|ones|zeros)\("

Those could be implemented in CompliantNamespace in a similar way to how (#2895) gave us access to nw.Expr-only features at the compliant-level; but at a low cost 😏