[
https://issues.apache.org/jira/browse/HADOOP-1928?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12529553
]
stack commented on HADOOP-1928:
-------------------------------
Here is one possible approach:
+ On startup, a regionserver reports in for duty and usually, what happens next
is that the master sends the regionserver a MSG_REGION_OPEN message with a
payload of an HRegionInfo describing the particular region the master wants the
regionserver to open. Other messages sent by the master include
MSG_REGION_CLOSE, MSG_REGIONSERVER_STOP, etc.
+ Messages are comprised of a byte indicating the message type and optionally
an HRegionInfo payload for a particular region if the message from the master
addresses a particular region. Currently messages either have an empty payload
or they carry an HRegionInfo payload (See HMsg).
+ We could add a new INIT message. The INIT message would be sent by the master
the first time a regionserver reports in for duty. Its payload would be a
MapWritable laden with properties the regionserver needs going about its work:
e.g. filesystem to use and hbase.root to use, etc.
+ We'd change the HMsg so rather than an explicit HRegionInfo, instead it would
carry a Writable whose type would be deteremined by the command carried in the
message.
+ On the regionserver side, it would read the passed properties and set them
into its local config. No attempts at touching the filesystem or startup of
service threads would happen until after it had received the INIT message from
the master.
> [hbase] Have master pass the regionserver the filesystem to use
> ---------------------------------------------------------------
>
> Key: HADOOP-1928
> URL: https://issues.apache.org/jira/browse/HADOOP-1928
> Project: Hadoop
> Issue Type: Improvement
> Components: contrib/hbase
> Reporter: stack
> Assignee: stack
> Priority: Minor
>
> The less configuration needed starting a regionserver the better. Currently
> the regionserver needs to read hadoop-site mainly so it knows which
> filesystem to go against. If it could get this info from the master instead
> when it gets the response to its reporting-for-duty message, then the
> dependency on hadoop-site.xml could mostly go away.
> Looking at regionserver, it looks like it would just take a little
> reorganization so that the hlog opening and the startup of worker threads
> doesn't happen until after we get our first response from the master.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
