[
https://issues.apache.org/jira/browse/YUNIKORN-2648?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17851407#comment-17851407
]
Craig Condit edited comment on YUNIKORN-2648 at 6/2/24 11:36 AM:
-----------------------------------------------------------------
I don’t think we can do this safely. I wrote the initial deadlock detector
implementation. Waiting until we load the configmap is far too late in the
configuration process to enable the deadlock detector. It is *absolutely vital*
that no locks whatsoever have been taken until the detector is initialized.
That is simply not possible if we wait until a configmap has been loaded,
parsed, and applied via K8s. It also creates an attractive nuisance for users,
who may be tempted to run this without understanding the rather drastic
consequences of its use. The implementation was purposely kept undocumented and
difficult to enable to prevent its use except by developers. It is completely
unsuitable for a production workload unless there is no other way to obtain the
deadlock information.
was (Author: ccondit):
I don’t think we can do this safely. I wrote the initial deadlock detector
implementation. Waiting until we load the configmap is far too late in the
configuration process to enable the deadlock detector. It is{*}absolutely
vital{*} that no locks whatsoever have been taken until the detector is
initialized. That is simply not possible if we wait until a configmap has been
loaded, parsed, and applied via K8s.
> Add deadlock detection config to the configmap
> ----------------------------------------------
>
> Key: YUNIKORN-2648
> URL: https://issues.apache.org/jira/browse/YUNIKORN-2648
> Project: Apache YuniKorn
> Issue Type: Improvement
> Components: core - common
> Reporter: Wilfred Spiegelenburg
> Assignee: Hsien-Cheng(Ryan) Huang
> Priority: Critical
>
> The current deadlock detection is configured using environment variables.
> That requires a change of the image and a restart of the scheduler to take
> effect and is not easy to maintain.
> We should be using yunikorn-defaults config map for the settings. We want a
> default set, turned off, for production use cases. However making the configs
> loadable from the config map makes turning it on easier.
> Update the configmap and restart the scheduler to turn the detection on or
> off.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
