[
https://issues.apache.org/jira/browse/HADOOP-10794?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14060928#comment-14060928
]
Colin Patrick McCabe commented on HADOOP-10794:
-----------------------------------------------
I agree with Andrew and Steve here. It probably makes sense to have somebody
monitor the clock skew between nodes, and warn if it gets too high.
It's worth pointing out that we have very carefully avoided depending on
synchronized clocks in HDFS and MR. If YARN wants to use local clocks to give
an approximation of task runtime, that's fine, but we should not depend on too
much accuracy there. NTP has its limits.
I think it makes sense to make YARN have its NodeManagers ping back
periodically, and complain if their local clocks are too far off (probably we
want a granularity of minutes here...) It fits in well with the other
resources YARN is managing, and would allow people to easily diagnose incorrect
task runtimes.
> A hadoop cluster needs clock synchronization
> --------------------------------------------
>
> Key: HADOOP-10794
> URL: https://issues.apache.org/jira/browse/HADOOP-10794
> Project: Hadoop Common
> Issue Type: Bug
> Reporter: Zhijie Shen
>
> As a distributed system, a hadoop cluster wants the clock on all the
> participating hosts synchronized. Otherwise, some problems might happen. For
> example, in YARN-2251, due to the clock on the host for the task container
> falls behind that on the host of the AM container, the computed elapsed time
> (the diff between the timestamps produced on two hosts) becomes negative.
> In YARN-2251, we tried to mask the negative elapsed time. However, we should
> seek for a decent long term solution, such as providing mechanism to do and
> check clock synchronization.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
