[
https://issues.apache.org/jira/browse/HDFS-6353?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jing Zhao updated HDFS-6353:
----------------------------
Attachment: HDFS-6353.001.patch
Had some further discussion with [~sureshms]. It should be more complete to
make the decision (about whether to do an extra checkpoint before stopping NN)
based on not only the time of the last checkpoint, but also the tx gap between
the current status and the ckpt. Update the patch accordingly.
> Handle checkpoint failure more gracefully
> -----------------------------------------
>
> Key: HDFS-6353
> URL: https://issues.apache.org/jira/browse/HDFS-6353
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: namenode
> Reporter: Suresh Srinivas
> Assignee: Jing Zhao
> Attachments: HDFS-6353.000.patch, HDFS-6353.001.patch
>
>
> One of the failure patterns I have seen is, in some rare circumstances, due
> to some inconsistency the secondary or standby fails to consume editlog. The
> only solution when this happens is to save the namespace at the current
> active namenode. But sometimes when this happens, unsuspecting admin might
> end up restarting the namenode, requiring more complicated solution to the
> problem (such as ignore editlog record that cannot be consumed etc.).
> How about adding the following functionality:
> When checkpointer (standby or secondary) fails to consume editlog, based on a
> configurable flag (on/off) to let the active namenode know about this
> failure. Active namenode can enters safemode and saves namespace. When in
> this type of safemode, namenode UI also shows information about checkpoint
> failure and that it is saving namespace. Once the namespace is saved,
> namenode can come out of safemode.
> This means service unavailability (even in HA cluster). But it might be worth
> it to avoid long startup times or need for other manual fixes. Thoughts?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
