[JeonDaehong]

Purpose

Linked issue: part of #1336

last_value, first_value, last_non_null_value, first_non_null_value and primary-key hold the current winner as a one-row slice and take a new slice for every input row. Columns without an aggregate function default to last_non_null_value, so almost every aggregation table pays this per row.

Brief change log

• FieldAggregator::agg_array(&ArrayRef, row) defaults to agg. The pick aggregators override it to keep the input array and row index, and result() slices once.
• merge_ordered and the partial-update aggregator path pass the column as an ArrayRef. Columns without an aggregator keep the existing path.

PK table, 2M keys written in 3 overlapping rounds, sum on one column, full read (release, in-memory FileIO): 2,105–2,135 → 1,827–1,856 ms, measured in one binary with the old and new call paths switched by an env var, 3 alternating rounds. Allocations per merged key go from 10.2 to 8.2. Deduplicate and partial-update are unchanged.

Tests

• agg_array_matches_agg_for_every_policy: each pick policy gives the same result through agg_array and agg, with NULLs among the inputs.
• agg_array_winner_feeds_reversed_and_delete_paths: agg_reversed and replace_with_delete after agg_array.

Ran the ci.yml steps locally (fmt, clippy, msrv, unit, rest-server, integration, datafusion, python, go).

API and Format

None. agg_array is crate-internal.

Documentation

None.