hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "chunhui shen (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (HBASE-7403) Online Merge
Date Wed, 13 Mar 2013 06:02:18 GMT

     [ https://issues.apache.org/jira/browse/HBASE-7403?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

chunhui shen updated HBASE-7403:
--------------------------------

    Description: 
Support executing region merge transaction on Regionserver, similar with split transaction

Process of merging two regions:
a.client send RPC(dispacth merging regions) to master
b.master move the regions together (on the same regionserver)
c.master send RPC(merge regions) to regionserver
d.Regionserver execute the regions merge transaction in the thread pool
e.the above b,c,d run asynchronously

Process of region merge transaction:
a.Construct a new region merge transaction.

b.prepare for the merge transaction, the transaction will be canceled if it is unavailable,

e.g. two regions don't belong to same table; two regions are not adjacent in a non-compulsory
merge; region is closed or has reference

c.execute the transaction as the following:
    /**
     * Set region as in transition, set it into MERGING state.
     */
    SET_MERGING_IN_ZK,
    /**
     * We created the temporary merge data directory.
     */
    CREATED_MERGE_DIR,
    /**
     * Closed the merging region A.
     */
    CLOSED_REGION_A,
    /**
     * The merging region A has been taken out of the server's online regions list.
     */
    OFFLINED_REGION_A,
    /**
     * Closed the merging region B.
     */
    CLOSED_REGION_B,
    /**
     * The merging region B has been taken out of the server's online regions list.
     */
    OFFLINED_REGION_B,
    /**
     * Started in on creation of the merged region.
     */
    STARTED_MERGED_REGION_CREATION,
    /**
     * Point of no return. If we got here, then transaction is not recoverable
     * other than by crashing out the regionserver.
     */
    PONR

d.roll back if step c throws exception


Usage:

HBaseAdmin#mergeRegions


  was:
The feature of this online merge:
1.Online,no necessary to disable table
2.Less change for current code, could applied in trunk,0.94 or 0.92,0.90
3.Easy to call merege request, no need to input a long region name, only encoded name enough
4.No limit when operation, you don't need to tabke care the events like Server Dead, Balance,
Split, Disabing/Enabing table, no need to take care whether you send a wrong merge request,
it has alread done for you
5.Only little offline time for two merging regions

Usage:
1.Tool:  
bin/hbase org.apache.hadoop.hbase.util.OnlineMerge [-force] [-async] [-show] <table-name>
<region-encodedname-1> <region-encodedname-2>

2.API: static void MergeManager#createMergeRequest


We need merge in the following cases:
1.Region hole or region overlap, can’t be fix by hbck
2.Region become empty because of TTL and not reasonable Rowkey design
3.Region is always empty or very small because of presplit when create table
4.Too many empty or small regions would reduce the system performance(e.g. mslab)

Current merge tools only support offline and are not able to redo if exception is thrown in
the process of merging, causing a dirty data

For online system, we need a online merge.


This implement logic of this patch for  Online Merge is :
For example, merge regionA and regionB into regionC
1.Offline the two regions A and B
2.Merge the two regions in the HDFS(Create regionC’s directory, move regionA’s and regionB’s
file to regionC’s directory, delete regionA’s and regionB’s directory)
3.Add the merged regionC to .META.
4.Assign the merged regionC

As design of this patch , once we do the merge work in the HDFS,we could redo it until successful
if it throws exception or abort or server restart, but couldn’t be rolled back. 

It depends on
Use zookeeper to record the transaction journal state, make redo easier
Use zookeeper to send/receive merge request
Merge transaction is executed on the master
Support calling merge request through API or shell tool

About the merge process, please see the attachment and patch

    
> Online Merge
> ------------
>
>                 Key: HBASE-7403
>                 URL: https://issues.apache.org/jira/browse/HBASE-7403
>             Project: HBase
>          Issue Type: New Feature
>    Affects Versions: 0.95.0
>            Reporter: chunhui shen
>            Assignee: chunhui shen
>            Priority: Critical
>             Fix For: 0.95.0, 0.98.0
>
>         Attachments: 7403-trunkv5.patch, 7403-trunkv6.patch, 7403v5.diff, 7403-v5.txt,
7403v5.txt, hbase-7403-94v1.patch, hbase-7403-trunkv10.patch, hbase-7403-trunkv11.patch, hbase-7403-trunkv12.patch,
hbase-7403-trunkv13.patch, hbase-7403-trunkv14.patch, hbase-7403-trunkv15.patch, hbase-7403-trunkv16.patch,
hbase-7403-trunkv19.patch, hbase-7403-trunkv1.patch, hbase-7403-trunkv20.patch, hbase-7403-trunkv5.patch,
hbase-7403-trunkv6.patch, hbase-7403-trunkv7.patch, hbase-7403-trunkv8.patch, hbase-7403-trunkv9.patch,
merge region.pdf
>
>
> Support executing region merge transaction on Regionserver, similar with split transaction
> Process of merging two regions:
> a.client send RPC(dispacth merging regions) to master
> b.master move the regions together (on the same regionserver)
> c.master send RPC(merge regions) to regionserver
> d.Regionserver execute the regions merge transaction in the thread pool
> e.the above b,c,d run asynchronously
> Process of region merge transaction:
> a.Construct a new region merge transaction.
> b.prepare for the merge transaction, the transaction will be canceled if it is unavailable,

> e.g. two regions don't belong to same table; two regions are not adjacent in a non-compulsory
merge; region is closed or has reference
> c.execute the transaction as the following:
>     /**
>      * Set region as in transition, set it into MERGING state.
>      */
>     SET_MERGING_IN_ZK,
>     /**
>      * We created the temporary merge data directory.
>      */
>     CREATED_MERGE_DIR,
>     /**
>      * Closed the merging region A.
>      */
>     CLOSED_REGION_A,
>     /**
>      * The merging region A has been taken out of the server's online regions list.
>      */
>     OFFLINED_REGION_A,
>     /**
>      * Closed the merging region B.
>      */
>     CLOSED_REGION_B,
>     /**
>      * The merging region B has been taken out of the server's online regions list.
>      */
>     OFFLINED_REGION_B,
>     /**
>      * Started in on creation of the merged region.
>      */
>     STARTED_MERGED_REGION_CREATION,
>     /**
>      * Point of no return. If we got here, then transaction is not recoverable
>      * other than by crashing out the regionserver.
>      */
>     PONR
> d.roll back if step c throws exception
> Usage:
> HBaseAdmin#mergeRegions

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Mime
View raw message