Details
-
Wish
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
Description
A few parts of Arrow have started spawning threads to parallelize CPU-bound tasks. They do so by launching a hard-coded number of threads, disregarding machine configuration or the fact that Arrow itself might be used from several threads at once (so hardcoding 8 threads might end up launch N * 8 threads if the user is calling Arrow from N threads at once...).
Instead we probably want a global thread pool policy, with a singleton thread pool for CPU-bound tasks (using, by default, a number of threads equal to the machine's capacity).
See discussion in https://github.com/apache/arrow/pull/1893
Attachments
Issue Links
- links to