weights not yet supported on array with null values

Error message

weights not yet supported on array with null values

What it means

nulls::rolling_min, the null-handling rolling minimum in polars-compute, begins with a panic guard: if weights is Some(_) on an array with nulls it panics 'weights not yet supported on array with null values'. Weighted rolling min/max has no correct definition on the null path, so the guard refuses instead of guessing.

Source

Thrown at crates/polars-compute/src/rolling/nulls/min_max.rs:22

pub type MinWindow<'a, T> = MinMaxWindow<'a, T, MinPropagateNan>;
pub type MaxWindow<'a, T> = MinMaxWindow<'a, T, MaxPropagateNan>;

use super::*;

pub fn rolling_min<T>(
    arr: &PrimitiveArray<T>,
    window_size: usize,
    min_periods: usize,
    center: bool,
    weights: Option<&[f64]>,
    _params: Option<RollingFnParams>,
) -> ArrayRef
where
    T: NativeType + IsFloat,
{
    if weights.is_some() {
        panic!("weights not yet supported on array with null values")
    }
    if center {
        rolling_apply_agg_window::<MinMaxWindow<T, MinPropagateNan>, _, _, _>(
            arr.values().as_slice(),
            arr.validity().as_ref().unwrap(),
            window_size,
            min_periods,
            det_offsets_center,
            None,
        )
    } else {
        rolling_apply_agg_window::<MinMaxWindow<T, MinPropagateNan>, _, _, _>(
            arr.values().as_slice(),
            arr.validity().as_ref().unwrap(),
            window_size,
            min_periods,
            det_offsets,
            None,

View on GitHub (pinned to 9b5d73fd00)

Solutions

Drop weights for nullable input - unweighted rolling_min handles nulls correctly
Fill nulls first (s.fill_null(strategy)) and be explicit about how filling interacts with min
Compute on s.drop_nulls() and reindex afterwards if alignment permits
Request/implement weighted null-aware rolling min upstream

Example fix

# before
s.rolling_min(window_size=5, weights=w)  # panics if s has nulls
# after
out = s.drop_nulls().rolling_min(window_size=5, weights=w).reindex(s.arg_null... )
# or simply: s.rolling_min(window_size=5)  # unweighted, null-aware

Defensive patterns

Strategy: validation

Validate before calling

def safe_rolling_min(s: pl.Series, window_size: int, weights=None, min_periods=1):
    if weights is not None and s.null_count() > 0:
        raise ValueError("weights + nulls unsupported for rolling_min")
    return s.rolling_min(window_size, weights=weights, min_periods=min_periods)

Try / catch

try:
    out = s.rolling_min(window_size=5, weights=w)
except pl.exceptions.PanicException:
    out = s.rolling_min(window_size=5, min_periods=2)  # unweighted null-aware fallback

Prevention

Gate weights on null_count() == 0 for rolling min/max
Forward-fill or drop nulls before weighted rolling features
Keep weighting vectors optional with a null-safe default in shared code

When it happens

Trigger: s.rolling_min(window_size=k, weights=[...]) or the equivalent expression on a Series containing at least one null; group_by aggregations where some groups contain nulls and a weights vector is passed.

Common situations: Rolling-window feature engineering (min over trailing k periods) with a decay/exp weighting vector applied to nullable market or sensor data; the same query ran unweighted or on cleaned data without issue.

Related errors

AI-assisted analysis of pola-rs/polars@9b5d73fd00 (2026-08-19). Data as JSON: /api/errors/a9c54ecdd17eaed1. Report an issue: GitHub.