hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jean-Daniel Cryans (JIRA)" <j...@apache.org>
Subject [jira] Updated: (HBASE-2129) Simple Master/Slave replication
Date Fri, 15 Jan 2010 04:44:54 GMT

     [ https://issues.apache.org/jira/browse/HBASE-2129?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Jean-Daniel Cryans updated HBASE-2129:
--------------------------------------

    Attachment: HBASE-2129.patch

This first patch adds a new contrib called "mdc_replication". It is master/slave and only
supports 1 slave. The modifications done to the core project are only there to enable the
sub-classing and extended functionality (like ZKW which has more friendly methods).

The package.html file gives an overview of the features and describes how to deploy it on
2 clusters. Every user table is replicated, there's no scoping. I also uses Hadoop RPC as
it was already integrated, in the future the plan is to use a non-versioned RPC mechanism.

Replication works by first collecting the HLog.Entry in a queue on the master cluster and
are sent in a batch every 10 seconds by default. On the slave cluster those edits are stored
in a log file and then replayed by a separate thread. It is currently blocking so while the
edits are played no other batch of edits can come in.

A region server decides on which RS to replicate to by scanning the /rs directory in the slaves
ZK folder and then randomly chooses a subset of nodes according to a default ration of 10%.
This is to make sure that we can handle clusters of different sizes. When the RS actually
replicates a batch of edits, it first chooses a random node from the subset. This is a cheap
way to do try to even the load on the slave cluster.

If a region server dies on the slave side, the RS on the master gets a new subset and retries
with a (hopefully) new node.

If a region server dies on the master side, the RS getting a region from it will first replicate
every edit found in the oldlogfile.

Apart from the new unit tests, I tested this patch on a single machine by putting two fully
distributed clusters of 1 region server that used the same HDFS and ZK quorum (the machine
is a i7, 12GB RAM, 2x500GB in RAID1). I was able to run PE sequentialWrite 1 on the first
cluster and read all the edits on the slave cluster.

There is a big potential for refactoring (ReplicationSink works a bit like a HLog) in this
patch and some parts are done with something bigger in mind (multiple slaves for example).
So, at this point, comments on the general layout are very welcomed but nitpicks won't be
very useful since it this code is meant to change a lot.

> Simple Master/Slave replication
> -------------------------------
>
>                 Key: HBASE-2129
>                 URL: https://issues.apache.org/jira/browse/HBASE-2129
>             Project: Hadoop HBase
>          Issue Type: Sub-task
>            Reporter: Jean-Daniel Cryans
>            Assignee: Jean-Daniel Cryans
>             Fix For: 0.21.0
>
>         Attachments: HBASE-2129.patch
>
>
> We should first build a very simple replication mechanism to validate our assumptions
and get a feel of what replication is in this very distributed context. 

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


Mime
View raw message