[zahariash]

This work was produced with the help of language models.

Part of #1432.

On the issue's 50-row table, a prepared MATCH (t:T) RETURN t.id, t.age took 94 µs with the default 16 threads and 11 µs with CALL threads=1. The table is one morsel, but its scan ran as a parallel task, so every execution woke the worker pool; ORDER BY woke it twice. TopK also allocated and zeroed two 2048-value vectors per payload column in every buffer, which glibc faulted back in on each execution (36 page faults for LIMIT 10 on one thread).

Fix. A node table scan with at most one morsel, counting uncommitted node groups, and an ORDER BY merge with at most one sorted run become one-thread tasks, which run on the calling thread unless a query timeout is set (#1297). Only one thread had work there, so no parallelism is lost. The scan now sets up its shared state, where it counts morsels, just before scheduling. TopKBuffer allocates its scratch vectors on first use in reduce() or merge().

Median µs, Ryzen 9 6900HS, 16 threads, C++ API:

query on 50 rows	text, main → PR	prepared, main → PR	allocations per prepared execution
scan	229 → 134	94 → 12	349 → 103
WHERE t.age > 45	301 → 184	102 → 12	503 → 137
+ ORDER BY t.age DESC	506 → 253	244 → 40	1,204 → 387
+ LIMIT 10	472 → 296	196 → 75	1,478 → 479

Most of the remaining text-query time is compiling; the comment on #1432 covers it.

Tests. SingleMorselTaskTest records each scanned row's thread with a UDF: a one-morsel scan runs on the caller, while three morsels, an uncommitted row, or two tables under MATCH (n) run on workers. A merge task accepts one thread for one sorted run and two for two. top_k_reduce.test reaches TopK's reduce(), which no test did. Each check fails when its condition is removed. Arrow and ice-disk tables with several row groups are not covered. Full ctest: no failures.

Cost. Scans of more than one morsel are unchanged. A 2M-node table (sum, sort, TopK) and a 1,000-node source with 2M edges stay within 0.4% of main's instructions, TopK measured on one thread. On LDBC SNB SF1 (the 30-query suite and complex IC1–IC14, 4 and 16 cores), no query is slower in every round; geomeans are 0.97–1.02 of main.

[adsharma]

I was concerned about sub-morsel parallelism. But it's not an issue. In the cases where the node groups are split, this optimization is bypassed because getNumMorsels() > 1.