Hi Everyone,

As you are aware, We have been doing Preemption Hardening exercise to
address the gaps discovered in earlier releases. It is being tracked under
this <https://issues.apache.org/jira/browse/YUNIKORN-2493> umbrella jira.

It contains a number of sub tasks. While working on this, we came across a use
case <https://issues.apache.org/jira/browse/YUNIKORN-2736> from a member of
the community that seemed to be valid at first sight. On further
investigation, We realized that taking care of the same by making changes
in code would lead to Preemption Storm causing instability of the overall
functioning of the queues, which is not the desirable behaviour. Preemption
storm has been covered in the usage guide
<https://yunikorn.apache.org/docs/next/user_guide/preemption_cases#redistribution-of-quota-and-preemption-storm>
doc. Along the way, we realized that the same use case can be addressed
without causing Preemption Storm and also preventing other cases from
causing storms. Solution has been discussed in this
<https://docs.google.com/document/d/1vfv8XJJsIqlZN3ecyhP2EFeiUSQnY-9GHEi_UJFB-BU/edit?usp=sharing>
document. Please refer to the 2A case in the document in detail as it is
the only case that would be addressed. Not only does this proposed solution
increase chances of freeing up resources for a potential candidate waiting
for resources in a starving queue for a particular situation, it also
follows the same principles of already working cases discussed in 3A & 3B.
3A & 3B are the cases where one can realize how preemption laws have been
followed strictly. Proposed solution for case 2A follows the same path.

Please share your thoughts on this.

Thanks,
Mani

Reply via email to