[ARROW-11300] [Rust][DataFusion] Improve hash aggregate performance with large number of groups in

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.0.0
• Component/s: Rust - DataFusion
• Labels: pull-request-available
• External issue URL: https://github.com/apache/arrow/issues/27200
• Assignee: Daniël Heres
• Reporter: Daniël Heres
• Created: 18/Jan/21 12:03
• Updated: 11/Jan/23 08:18
• Resolved: 28/Jan/21 03:04

Description

Currently, hash aggregates are performing well when having a small number of output groups, but the results on db-benchmark https://github.com/h2oai/db-benchmark/pull/182 test on data with a high number of output groups.
https://github.com/apache/arrow/pull/9234 improved the situation a bit, but DataFusion is still much slower than even the slowest result when comparing to the published results.

This seems mostly having to do with the way we use individual key/groups.
For each new key, we take the indices of the group, resulting in lots of small allocations and cache unfriendliness and other overhead if we have many keys with only a small (just 1-2) number of rows per group in a batch. Also the indices are converted from a Vec to an Array, making the situation worse (accounts for ~22% of the instructions on the master branch!), other profiling results seem to be from related allocations too.

To make it efficient for tiny groups, we should probably change the hash aggregate algorithm to take based on all indices from the batch in one go, and "slice" into the resulting array for the individual accumulators.
 
Here is some profiling info of the db-benchmark questions 1-5 against master:

Issue Links

• links to: GitHub Pull Request #9271