maxframe.liteframe.groupby.LiteFrameGroupBy.rolling#

LiteFrameGroupBy.rolling(window, min_periods=None, center=False, columns=None, order_by=None, order_by_ascending=True, window_shift=0)[source]#

Provide rolling-window calculations within each group.

The data is partitioned by the group key (hash shuffle) so that every group resides in a single partition; the rolling aggregation is then computed per-group. Only integer window sizes are supported.

Parameters:
  • window (int) – Size of the moving window (number of rows). Must be positive.

  • min_periods (int or None, default None) – Minimum number of observations in a window required to produce a value (otherwise null). Defaults to window (matching pandas). For count the value is the number of non-null observations, but min_periods still gates on the number of window rows (observations), not the non-null count — consistent with pandas.

  • center (bool, default False) – Center the window on the current row.

  • columns (str, list of str, or None, default None) – Column(s) to aggregate. If None, all non-group-key columns are aggregated. Non-selected columns are passed through unchanged.

Returns:

Intermediate object; call an aggregation method (e.g. .sum()) to produce a LiteFrame.

Return type:

LiteFrameRolling

Notes

Warning

Grouped transforms hash-shuffle rows by group key, so the output row order follows reducer/hash-partition order and is not aligned with the input frame (unlike pandas, which preserves the index). Order restoration via a place-back operator (LiteFrameIlocSetItem) is planned; see docs/designs/liteframe/operators/window.md.

Null-key rows are preserved with a null result (matching the other grouped transforms and the row-count-preserving transform contract); they are not grouped together. (This diverges from pandas grouped-rolling, which drops null-key rows under dropna=True.)

Raises: