[PATCH] Use optimized single-datum tuplesort in ExecSort

Ronan Dunklau <ronan.dunklau@aiven.io>

From: Ronan Dunklau <ronan.dunklau@aiven.io>
To: pgsql-hackers@lists.postgresql.org
Cc: ranier.vf@gmail.com, James Coleman <jtc331@gmail.com>
Date: 2021-07-06T06:15:41Z
Lists: pgsql-hackers

Attachments

Hello,

While testing the patch "Add proper planner support for ORDER BY / DISTINCT 
aggregates" [0] I discovered the performance penalty from adding a sort node 
essentially came from not using the single-datum tuplesort optimization in 
ExecSort (contrary to the sorting done in ExecAgg).

I originally proposed this patch as a companion in the same thread [1], but 
following James suggestion I'm making a separate thread just for this as the 
optimization is worthwhile independently of David's patch: it looks like we 
can expect a 2x speedup on a "select a single ordered column" case.

The patch aimed to be as simple as possible: we only turn this optimization on 
when the tuple being sorted has only one attribute, it is "byval" (so as not 
to incur copies which would be hard to track in the execution tree) and 
unbound (again, not having to deal with copying borrowed datum anywhere).

The attached patch is originally by me, with some cleanup by Ranier Vilela. 
I'm sending Ranier's version here.


[0]: https://commitfest.postgresql.org/33/3164/
[1]: https://www.postgresql.org/message-id/4480689.ObhdGn8bVM%40aivenronan

-- 
Ronan Dunklau

Commits

  1. Make nodeSort.c use Datum sorts for single column sorts