Robert Kanter commented on YARN-2131:

Are you sure that the NameNode uses a lock file to do this?  I took a look at 
NameNode.java, and from what I can tell, it doesn't actually do anything to 
prevent you from formatting a running NameNode (there's a config property it 
checks where you can disable formatting, but that's not really the same thing). 
 Just to make sure I wasn't missing something, I also just tried it on a 
pseudo-cluster and it didn't stop me.  

I agree it would be nice to do that; I'm just not sure of a good way.  I don't 
think a lock file will work because the command could be run from any node 
where the RM is installed, not necessarily the currently running or active one; 
in which case, a lock file isn't accessible unless it's in HDFS (and I don't 
think we want an HDFS dependency for this).  I suppose we could use a lock on 
whatever is backing the store (e.g. Znode for ZKRMStateStore, file for 
FSRMStateStore, etc), but that gets tricky: what happens if the RM crashes 
without releasing the lock?  Then it's stuck.  Doing that would also mean we 
can't simply wipe out the entire store space because we need to keep the lock 
there.  I'll keep thinking about it, but I'm open to suggestions if you have 
any other ideas.  Unless you think we should make this another JIRA?

> Add a way to format the RMStateStore
> ------------------------------------
>                 Key: YARN-2131
>                 URL: https://issues.apache.org/jira/browse/YARN-2131
>             Project: Hadoop YARN
>          Issue Type: New Feature
>          Components: resourcemanager
>    Affects Versions: 2.4.0
>            Reporter: Karthik Kambatla
>            Assignee: Robert Kanter
>             Fix For: 2.6.0
>         Attachments: YARN-2131.patch, YARN-2131.patch, 
> YARN-2131_addendum.patch
> There are cases when we don't want to recover past applications, but recover 
> applications going forward. To do this, one has to clear the store. Today, 
> there is no easy way to do this and users should understand how each store 
> works.

This message was sent by Atlassian JIRA

Reply via email to