[
https://issues.apache.org/jira/browse/HDFS-9426?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15004321#comment-15004321
]
Kihwal Lee commented on HDFS-9426:
----------------------------------
bq. But is that expected scenario, where cluster upgrade can be finalized
without complete upgrade of all datanodes?
Yes. For larger clusters a rolling upgrade might take a while. If they don't
have enough free space or there is huge churn of blocks, the remaining space
will shrink quickly. In that case users have to finalize the upgrade after
enough number of nodes roll and no anomaly is observed.
> Rollingupgrade finalization is not backward compatible
> ------------------------------------------------------
>
> Key: HDFS-9426
> URL: https://issues.apache.org/jira/browse/HDFS-9426
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Kihwal Lee
> Priority: Blocker
>
> After HDFS-7645, the namenode can return non-null {{rollingUpgradeInfo}} in
> heatbeat reponses. 2.7.1 or 2.6.x datanodes won't finalize the upgrade
> because it's not null.
> NN might have to check the DN version and return different
> {{rollingUpgradeInfo}}.
> HDFS-8656 recognized the compatibility issue of the changed semantics, but
> unfortunately did not address the semantics of the heartbeat response.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
