Github user jinxing64 commented on the issue:
https://github.com/apache/spark/pull/16867
@squito
Thanks a lot for your comments : )
>When check speculatable tasks in TaskSetManager, current code scan all
task infos and sort durations of successful tasks in O(NlogN) time complexity.
`checkSpeculatableTasks` is scheduled every 100ms by
`scheduleAtFixedRate`(not `scheduleWithFixedDelay `), thus the interval can be
less than 100ms. In my cluster(yarn-cluster mode), if size of the task set is
over 300000 and the driver is running on some machine with poor cpu
performance, the `Arrays.sort` can take over than 100ms easily. Since
`checkSpeculatableTasks` will synchronize `TaskSchedulerImpl`, I suspect that's
why my driver hang.
I get median duration by `TreeSet.slice`, which comes from `IterableLike`
and cannot jump to the mid position unluckily. The time complexity is O(n) in
this pr.
I can get the mid position by reflection, but I don't want to do that, I
think that is harmful for code clarity.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
