[SPARK-48791] Perf regression due to accumulator registration overhead using CopyOnWriteArrayList

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 4.0.0, 3.5.1, 3.3.4, 3.4.3
• Fix Version/s: 4.0.0, 3.5.2, 3.4.4
• Component/s: Spark Core
• Labels: pull-request-available
• Assignee: wuyi
• Reporter: wuyi
• Created: 03/Jul/24 06:15
• Updated: 09/Aug/24 18:34
• Resolved: 10/Jul/24 12:42

Description

We noticed query perf regression and locate the root cause is the overhead introuduced when registering accumulators using CopyOnWriteArrayList.

Issue Links

• is caused by: SPARK-39696 Uncaught exception in thread executor-heartbeater java.util.ConcurrentModificationException: mutation occurred during iteration
• links to: GitHub Pull Request #692
• links to: GitHub Pull Request #693
• links to: GitHub Pull Request #47197
• links to: GitHub Pull Request #47297
• links to: GitHub Pull Request #47348
• links to: GitHub Pull Request #47578
• links to: GitHub Pull Request #47612
• links to: GitHub Pull Request #47613