My previous reply messed up the options numbering, It may confuse some people.
I am resending it again by correcting that. Appologies for confusions.
>>>1) rely on external HA filer
If you can invest on NFS filers, it is one option in HA. If you want to
try this, you can use released Hadoop-2 version and try. but below #3 and #4
will avoid this external hardware dependency.
>>>2) multiple edit directories
HDFS already supports, multiple edit directory configuration. I am not
exactly sure what option you are referring here.
>>>3) book keeper
Please take a look at https://issues.apache.org/jira/browse/HDFS-3399
We are testing with BookeKeeper from last 2/3 months and going well. BK is
progressing on autorecovery and security parts. Almost auto recoverry done
(BOOKKEEPER-237) and will be released in BK 4.2 version very soon. BK already
started work on security part as well. Also this integration part will come out
with next hadoop-2 release as well. Also attached tested scenarios in HDFS-3399
for your reference if you want to take a look. Also there is one subTask in
that umbrella JIRA for user manual information. BookKeeper Wiki
https://cwiki.apache.org/confluence/display/BOOKKEEPER/BookKeeper
>>>4) keep edits in HDFS / quorum based
you can take a look at HDFS-3077. In this umbrella JIRA, work is going
actively.
https://issues.apache.org/jira/browse/HDFS-3077 you can get more design
details in that JIRA.
Regards,
Uma
________________________________________
From: Sujee Maniyam [su...@sujee.net]
Sent: Wednesday, September 05, 2012 4:07 AM
To: hdfs-dev@hadoop.apache.org
Subject: current direction in namenode HA
Hello devs,
I am trying to understand the current state / direction of namenode
HA implementation.
For using shared directory, I see the following options
(from
http://www.cloudera.com/blog/2012/03/high-availability-for-the-hadoop-distributed-file-system-hdfs/
and https://issues.apache.org/jira/browse/HDFS-3278)
1) rely on external HA filer
2) multiple edit directories
3) book keeper
4) keep edits in HDFS / quorum based
is there going to be an 'official / supported' method, or it is going
to be a configurable choice when setting up a cluster?
thanks
Sujee
http://sujee.net
