Description

ArrowSeries.shift(n) padded the result with n nulls without capping n at the series length, so shifting by more than the length grew the series instead of returning an all-null series of the same length. For example, shift(3) on a two-row pyarrow series returned three rows.

The same series is used by the rolling functions — rolling_sum, rolling_mean, rolling_var and rolling_std all shift by window_size internally — so a window larger than the series desynced the chunk lengths and raised ArrowInvalid: Array arguments must all be the same length instead of returning nulls:

import narwhals as nw
import pyarrow as pa

s = nw.from_native(pa.chunked_array([[1]]), series_only=True)

s.rolling_sum(window_size=2)   # main: ArrowInvalid  ->  this PR: [None]
s.shift(3).to_list()           # main: [None, None, None]  ->  this PR: [None]

polars and pandas both return an all-null series of the original length here, so this brings the pyarrow backend in line with them.

The fix caps the null padding at the series length (min(n, length)), which keeps the existing behaviour for every in-range shift untouched. Added a regression test for shift beyond the length across the eager backends, and one for rolling_sum with a window longer than the series.

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

• 💾 Refactor
• ✨ Feature
• 🐛 Bug Fix
• 🔧 Optimization
• 📝 Documentation
• ✅ Test
• 🐳 Other

Related issues

• Found while testing the pyarrow backend; no existing issue.

AI assistance

• No AI tools were used for this PR.
• AI tools were used.

Checklist

• Code follows style guide (ruff)
• Tests added
• Documented the changes (no public API change; behaviour now matches polars/pandas)