Struct StateCleaning 

pub(super) struct StateCleaning {
    pub watermark_col_idx: usize,
    pub stale_rows_at_front: bool,
    pub n_retain: usize,
}

Watermark-driven state cleaning strategy of OverWindowExecutor.

This is only enabled when the optimizer decides it’s safe, i.e., when the input is append-only, all window frames are bounded ROWS frames, and the first order key column is a watermark column with NULLs ordered as the largest values.

Under these conditions, a row can only affect (and be affected by) a bounded number of neighboring rows in the partition. Once a watermark wm is received, no row with order key < wm will ever arrive, so rows with order key < wm (stale rows) can never get new neighbors on the “smaller” side. Therefore, among the stale rows, only the n_retain ones that are closest to the watermark boundary can still be involved in future computation (as frame members of, or as rows affected by, rows arriving in the future), and all the others can be safely deleted from the state table.

The cleaning happens at barriers for partitions touched since the last barrier. This keeps the work proportional to recently active partitions and avoids keeping per-partition cleanup metadata in memory. Active partitions retain only the required rows plus rows not yet behind the watermark. A partition that stops receiving rows is cleaned lazily when it’s touched again.

Fields

watermark_col_idx: usize

Index of the watermark column (the first order key column) in the input schema.

stale_rows_at_front: bool

Whether stale rows are at the front (ASC order key) or at the back (DESC order key) of the partition.

n_retain: usize

Number of stale rows to retain in each partition, which is the number of preceding rows plus the number of following rows of the union of all ROWS frames.

Trait Implementations

impl Debug for StateCleaning

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.