[
https://issues.apache.org/jira/browse/ARROW-9707?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17201762#comment-17201762
]
Adam Lippai commented on ARROW-9707:
------------------------------------
Speaking about the thread pools, my last suggestion wasn't accepted (or
declined):
[https://docs.google.com/document/d/1_wc6diy3YrRgEIhVIGzrO5AK8yhwfjWlmKtGnvbsrrY/edit#heading=h.4f9v08zbs2z3]
The use-case might be different, but recently Bevy created a rust task
scheduler using context-specific threadpools:
[https://bevyengine.org/news/bevy-0-2/]
My understanding is that Bevy and the linked GDocs (DAG vs 2-3 staged pipeline
vs generic threadpool) is similar approach and I find it the most generic (but
acceptable) solution.
> [Rust] [DataFusion] Re-implement threading model
> ------------------------------------------------
>
> Key: ARROW-9707
> URL: https://issues.apache.org/jira/browse/ARROW-9707
> Project: Apache Arrow
> Issue Type: Sub-task
> Components: Rust, Rust - DataFusion
> Reporter: Andy Grove
> Assignee: Andy Grove
> Priority: Major
> Fix For: 2.0.0
>
> Attachments: image-2020-09-24-22-46-46-959.png
>
>
> The current threading model is very simple and does not scale. We currently
> use 1-2 dedicated threads per partition and they all run simultaneously,
> which is a huge problem if you have more partitions than logical or physical
> cores.
> This task is to re-implement the threading model so that query execution uses
> a fixed (configurable) number of threads. Work will be broken down into
> stages and tasks and each in-process executor (running on a dedicated thread)
> will process its queue of tasks.
> This process will be driven by a scheduler.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
