[ARROW-13997] [C++] restore exec node based query performance

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 6.0.0
• Fix Version/s: 6.0.0
• Component/s: C++
• External issue URL: https://github.com/apache/arrow/issues/29597
• Assignee: Unassigned
• Reporter: Jonathan Keane
• Created: 14/Sep/21 19:04
• Updated: 11/Jan/23 08:36
• Resolved: 05/Oct/21 19:58

Description

The R bindings transitioned to using exec nodes for all queries. Since ARROW-13740 was merged, the performance for our queries has been slower. ARROW-13940 corrected the accidental disabling of multi threading, which recovered some of the performance, but not all of it.

The three benchmarks showing the degradation:
https://conbench.ursa.dev/benchmarks/c6d37c1517af405dbb8ebad83eec010c/
https://conbench.ursa.dev/benchmarks/cec58c9c21234ffbb1bab28b7fe409ef/
https://conbench.ursa.dev/benchmarks/99c04f5ef2fe4caf868904699c2f377c/

The queries for these are defined (by name) at: https://github.com/ursacomputing/arrowbench/blob/19cb0e5dac316ae5e3c9cac4d7534fc202459fe0/R/bm-dataset-taxi-parquet.R#L27-L90

For 2 of the three, there was a large increase when ARROW-13740 was merged and plateau, and then a smaller step down when ARROW-13940 was merged, but we still haven't reverted to the pre-change mean.

Issue Links

• is caused by: ARROW-14009 [C++] Ensure SourceNode truly feeds batches to plan in parallel