[ 
https://issues.apache.org/jira/browse/HBASE-10576?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13978998#comment-13978998
 ] 

Jeffrey Zhong commented on HBASE-10576:
---------------------------------------

Thanks for all the very good comments & feedbacks. I think the most doubts come 
from "no meta entries for shadow regions".

We can create meta entries for shadow regions but with "shadow" state(like 
disable state) so that AM & Balancer can skip them. With entries in META table, 
I think most questions are answered. As [~ram_krish] suggested above, there 
will be some tweaks in coprocessor hook & region open code in order to open a 
region from coprocessor.

{quote}
it conflicts with that we only want the master to handle region 
assignments/recovery
{quote}
There is no conflict because the shadow region location is pre-determined by 
its data region. 

My original thought without meta entries for shadow regions is a coprocessors 
intensive approach as following.

For creating local-index tables, we reopen data table(through enable/disable 
table). During data table region opening, we can create index regions under 
same data region directory if not present or open index regions if already 
exists. 

For index table access, we can use coprocessor preScannerOpen() to intercept 
scan if the passed in scan has attribute(or hint) using index OR tweak hbase 
client to be aware of shadowing regions in terms of location finding.


> Custom load balancer to co-locate the regions of two tables which are having 
> same split keys
> --------------------------------------------------------------------------------------------
>
>                 Key: HBASE-10576
>                 URL: https://issues.apache.org/jira/browse/HBASE-10576
>             Project: HBase
>          Issue Type: Sub-task
>          Components: Balancer
>            Reporter: rajeshbabu
>            Assignee: rajeshbabu
>         Attachments: HBASE-10536_v2.patch, HBASE-10576.patch
>
>
> To support local indexing both user table and index table should have same 
> split keys. This issue to provide custom balancer to colocate the regions of 
> two tables which are having same split keys. 
> This helps in Phoenix as well.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Reply via email to