[ARROW-11802] [Rust][DataFusion] Mixing of crossbeam channel and async tasks can lead to deadlock

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.0.1, 4.0.0
• Component/s: Rust - DataFusion
• Labels: pull-request-available
• External issue URL: https://github.com/apache/arrow/issues/27653
• Assignee: Andrew Lamb
• Reporter: Andrew Lamb
• Created: 26/Feb/21 21:39
• Updated: 11/Jan/23 08:22
• Resolved: 03/Mar/21 11:57

Description

edrevo noticed, on https://github.com/apache/arrow/pull/9523#issuecomment-786237494, that the use of crossbeam channels can potentially deadlock datafusion

The use of crossbeam channel is left over from earlier, non `async` implementations and get been fingered in some hangs that MikeSeddonAU has observed in DataFusion ). Specifically the crossbeam channel can block a thread when the channel is full or empty, which can result in blocking all the tokio executor threads and deadlocking the system

The proposal is is to use tokio's mpsc channels instead of crossbeam which can properly yield back to tokio to run another task when the channel is either full or empty. .

Issue Links

• links to: GitHub Pull Request #9605