[
https://issues.apache.org/jira/browse/YARN-8298?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16586458#comment-16586458
]
Chandni Singh commented on YARN-8298:
-------------------------------------
{quote}
We have built sufficient number of knobs for individual container to upgrade in
a rolling fashion. However, it will depend on external orchestrator to perform
the rolling upgrade. Express upgrade is design to be atomic, therefore, it
simplifies upgrade process by doing all instances of a component in parallel.
Docker container takes only a few second to stop and start, therefore, the
interruption time is minimized to few seconds
{quote}
[~eyang] When an express upgrade is performed, I am of the opinion that the
upgrade of a single component should be done in a rolling fashion otherwise if
there is a failure the service is disrupted. If we provide express upgrade that
should be the default behavior. If upgrade of an instance fails, other
instances of the component should not be tried to be upgraded. Docker
container may take a few second to stop and start but the other instances of
the component will be active.
Besides that, with the 2nd approach, I meant that the scheduler should not do
any sort of orchestration including upgrading instances of a particular
component before another. This is blocked by YARN-8665 as it needs support for
cancelling an upgrade in case of failure.
Given that, if you want to go with the 2nd approach, then patch 4 contains all
the changes.
> Yarn Service Upgrade: Support express upgrade of a service
> ----------------------------------------------------------
>
> Key: YARN-8298
> URL: https://issues.apache.org/jira/browse/YARN-8298
> Project: Hadoop YARN
> Issue Type: Sub-task
> Affects Versions: 3.1.1
> Reporter: Chandni Singh
> Assignee: Chandni Singh
> Priority: Major
> Attachments: YARN-8298.001.patch, YARN-8298.002.patch,
> YARN-8298.003.patch, YARN-8298.004.patch
>
>
> Currently service upgrade involves 2 steps
> * initiate upgrade by providing new spec
> * trigger upgrade of each instance/component
>
> We need to add the ability to upgrade the service in one shot:
> # Aborting the upgrade will not be supported
> # Upgrade finalization will be done automatically.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
