You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@hbase.apache.org by "Todd Lipcon (JIRA)" <ji...@apache.org> on 2010/11/23 10:36:13 UTC

[jira] Created: (HBASE-3268) Auto-tune balance frequency based on cluster size

Auto-tune balance frequency based on cluster size
-------------------------------------------------

                 Key: HBASE-3268
                 URL: https://issues.apache.org/jira/browse/HBASE-3268
             Project: HBase
          Issue Type: Improvement
          Components: master
            Reporter: Todd Lipcon


Right now we only balance the cluster once every 5 minutes by default. This is likely to confuse new users. When you start a new region server, you expect it to pick up some load very quickly, but right now you have to wait 5 minutes for it to start doing anything in the worst case.

We could/should also add a button/shell command to "trigger balance now"

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


[jira] Commented: (HBASE-3268) Auto-tune balance frequency based on cluster size

Posted by "Jonathan Gray (JIRA)" <ji...@apache.org>.
    [ https://issues.apache.org/jira/browse/HBASE-3268?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12934918#action_12934918 ] 

Jonathan Gray commented on HBASE-3268:
--------------------------------------

Also, when the master gets a new regionserver on an already running cluster, it should automatically trigger a balance.

> Auto-tune balance frequency based on cluster size
> -------------------------------------------------
>
>                 Key: HBASE-3268
>                 URL: https://issues.apache.org/jira/browse/HBASE-3268
>             Project: HBase
>          Issue Type: Improvement
>          Components: master
>            Reporter: Todd Lipcon
>
> Right now we only balance the cluster once every 5 minutes by default. This is likely to confuse new users. When you start a new region server, you expect it to pick up some load very quickly, but right now you have to wait 5 minutes for it to start doing anything in the worst case.
> We could/should also add a button/shell command to "trigger balance now"

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


[jira] Commented: (HBASE-3268) Auto-tune balance frequency based on cluster size

Posted by "Andrew Purtell (JIRA)" <ji...@apache.org>.
    [ https://issues.apache.org/jira/browse/HBASE-3268?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12934795#action_12934795 ] 

Andrew Purtell commented on HBASE-3268:
---------------------------------------

+1

Actually I've been considering filing against this as a bug. I have been testing recently some heavy write scenarios that on current 0.90 pile regions on a single RS and can cause it to OOME before balancing happens. 

Perhaps at least the default should be 1 minute instead of 5?


> Auto-tune balance frequency based on cluster size
> -------------------------------------------------
>
>                 Key: HBASE-3268
>                 URL: https://issues.apache.org/jira/browse/HBASE-3268
>             Project: HBase
>          Issue Type: Improvement
>          Components: master
>            Reporter: Todd Lipcon
>
> Right now we only balance the cluster once every 5 minutes by default. This is likely to confuse new users. When you start a new region server, you expect it to pick up some load very quickly, but right now you have to wait 5 minutes for it to start doing anything in the worst case.
> We could/should also add a button/shell command to "trigger balance now"

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


[jira] Commented: (HBASE-3268) Auto-tune balance frequency based on cluster size

Posted by "Todd Lipcon (JIRA)" <ji...@apache.org>.
    [ https://issues.apache.org/jira/browse/HBASE-3268?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12934927#action_12934927 ] 

Todd Lipcon commented on HBASE-3268:
------------------------------------

I think the idea of triggering balance when we get a new server is a good one.

One thing we want to be a little careful of is the situation when someone flips on 10 new servers "at the same time". Rather than triggering a rebalance for each (and thus lots of churn), we want a little bit of lag before the rebalance.

Maybe when a new server is added, we trigger the rebalance in 5-10 seconds?

> Auto-tune balance frequency based on cluster size
> -------------------------------------------------
>
>                 Key: HBASE-3268
>                 URL: https://issues.apache.org/jira/browse/HBASE-3268
>             Project: HBase
>          Issue Type: Improvement
>          Components: master
>            Reporter: Todd Lipcon
>
> Right now we only balance the cluster once every 5 minutes by default. This is likely to confuse new users. When you start a new region server, you expect it to pick up some load very quickly, but right now you have to wait 5 minutes for it to start doing anything in the worst case.
> We could/should also add a button/shell command to "trigger balance now"

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


[jira] Commented: (HBASE-3268) Auto-tune balance frequency based on cluster size

Posted by "Jonathan Gray (JIRA)" <ji...@apache.org>.
    [ https://issues.apache.org/jira/browse/HBASE-3268?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12934938#action_12934938 ] 

Jonathan Gray commented on HBASE-3268:
--------------------------------------

Something like that sounds reasonable.  I'm trying to figure some of these other issues now so this is up for grabs.

> Auto-tune balance frequency based on cluster size
> -------------------------------------------------
>
>                 Key: HBASE-3268
>                 URL: https://issues.apache.org/jira/browse/HBASE-3268
>             Project: HBase
>          Issue Type: Improvement
>          Components: master
>            Reporter: Todd Lipcon
>
> Right now we only balance the cluster once every 5 minutes by default. This is likely to confuse new users. When you start a new region server, you expect it to pick up some load very quickly, but right now you have to wait 5 minutes for it to start doing anything in the worst case.
> We could/should also add a button/shell command to "trigger balance now"

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


[jira] Commented: (HBASE-3268) Auto-tune balance frequency based on cluster size

Posted by "Jonathan Gray (JIRA)" <ji...@apache.org>.
    [ https://issues.apache.org/jira/browse/HBASE-3268?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12934915#action_12934915 ] 

Jonathan Gray commented on HBASE-3268:
--------------------------------------

I think rather than needing to do much more frequent load balances or just in addition to them, we should add more intelligence into non-balancing region assignment (all random now).  And we could also lazily move splits off their original server.

But for now making it more aggressive at one minute or so should be fine.

> Auto-tune balance frequency based on cluster size
> -------------------------------------------------
>
>                 Key: HBASE-3268
>                 URL: https://issues.apache.org/jira/browse/HBASE-3268
>             Project: HBase
>          Issue Type: Improvement
>          Components: master
>            Reporter: Todd Lipcon
>
> Right now we only balance the cluster once every 5 minutes by default. This is likely to confuse new users. When you start a new region server, you expect it to pick up some load very quickly, but right now you have to wait 5 minutes for it to start doing anything in the worst case.
> We could/should also add a button/shell command to "trigger balance now"

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.