[ 
https://issues.apache.org/jira/browse/CASSANDRA-14310?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16431350#comment-16431350
 ] 

Jordan West edited comment on CASSANDRA-14310 at 4/9/18 11:23 PM:
------------------------------------------------------------------

[~krummas], do you have a test run you can share? 

I'm +1 on this, if this is the fix we want to go with. It prevents the deadlock 
caused by startup.

I do think its worth considering how to prevent future incarnations of this 
deadlock from being introduced. To that end, is it necessary for 
{{ColumnFamilyStore.loadNewSSTables}} to be {{synchronized}} on 
{{ColumnFamilyStore.class}}? Clearly its instance method counterpart does but 
why the static one? I don't see why and if we remove it, or replace it with a 
dedicated lock if locking is indeed necessary, then the lock ordering problem 
it introduces is gone as well. 


was (Author: jrwest):
[~krummas], do you have a test run you can share? 

I'm +1 on this, if this is the fix we want to go with. It prevents the deadlock 
caused by startup.

I do think its worth considering how to prevent future incarnations of this 
deadlock from being introduced. To that end, is it necessary for 
{{ColumnFamilyStore.loadNewSSTables}} to be {{synchronized}} on 
{{ColumnFamilyStore.class}}? Clearly its instance method counterpart does but 
why the static class? I don't see why and if we remove it, or replace it with a 
dedicated lock if locking is indeed necessary, then the lock ordering problem 
it introduces is gone as well. 

> Don't allow nodetool refresh before cfs is opened
> -------------------------------------------------
>
>                 Key: CASSANDRA-14310
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-14310
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Marcus Eriksson
>            Assignee: Marcus Eriksson
>            Priority: Major
>             Fix For: 3.0.x, 3.11.x, 4.x
>
>
> There is a potential deadlock in during startup if nodetool refresh is called 
> while sstables are being opened. We should not allow refresh to be called 
> before everything is initialized.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org

Reply via email to