[
https://issues.apache.org/jira/browse/YARN-6592?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16297841#comment-16297841
]
Konstantinos Karanasos commented on YARN-6592:
----------------------------------------------
Hi [~cheersyang]. Here is a high level description:
* Node labels are currently part of the capacity scheduler only and are coupled
with the cluster resources. And, unless something has changed, you can have
only one node label per node. For instance, let's say that you have blue and
red nodes. Then, when you are defining your queues in the capacity scheduler,
you can specify what percentage of the queue can go on red nodes and which
percentage on blue nodes. Similarly, you can specify what percentage of the
overall red nodes can be used by a specific queue. Overall, the important bit
is that node labels are coupled with cluster capacity.
* Node attributes are, like you say, key-value pairs, and as such, they are
strictly more expressive than node labels. But apart from that, they are not
related to the clusters capacity. A request can simply say that it wants to be
placed on a node with a specific java version or a specific generation of GPU.
* The current umbrella JIRA (YARN-6592) goes a step further by also assigning
attributes to containers and not just nodes. So, you can say that you want your
HBase master to be at a different node from your HBase region servers. It also
supports more involved constraints. If you check YARN-6593, you will see that
in the API we introduced, you are allowed to specify node attributes in
placement constraints. This way we want to unify the node attribute and the
container label constraints.
Hope this clarifies a bit the situation.
> Rich placement constraints in YARN
> ----------------------------------
>
> Key: YARN-6592
> URL: https://issues.apache.org/jira/browse/YARN-6592
> Project: Hadoop YARN
> Issue Type: New Feature
> Reporter: Konstantinos Karanasos
> Attachments: YARN-6592-Rich-Placement-Constraints-Design-V1.pdf
>
>
> This JIRA consolidates the efforts of YARN-5468 and YARN-4902.
> It adds support for rich placement constraints to YARN, such as affinity and
> anti-affinity between allocations within the same or across applications.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
