hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andrew Purtell (JIRA)" <j...@apache.org>
Subject [jira] Resolved: (HBASE-1968) Give clients access to the write buffer
Date Tue, 10 Nov 2009 04:57:32 GMT

     [ https://issues.apache.org/jira/browse/HBASE-1968?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Andrew Purtell resolved HBASE-1968.
-----------------------------------

    Resolution: Fixed

Committed trivial patch to 0.20 branch. 

> Give clients access to the write buffer
> ---------------------------------------
>
>                 Key: HBASE-1968
>                 URL: https://issues.apache.org/jira/browse/HBASE-1968
>             Project: Hadoop HBase
>          Issue Type: Improvement
>            Reporter: Andrew Purtell
>            Assignee: Andrew Purtell
>            Priority: Trivial
>             Fix For: 0.20.2
>
>
> From a Trend dev team:
> {quote}
> When insert rows into one table by calling the method public synchronized void put(final
Put put), if the column family of one row does not exist, the insert operation will failed
and throw NoSuchColumnFamilyException.. We observed that all the following insert operation
will fails even though all of them have valid column family. That is one exception of insert
operation can cause failure of all the following insert operation.
> {quote}
> Their further analysis explains in detail the scenario, which I will summarize here:
> 1) An invalid put is added to the writeBuffer by put(Put put). It will trigger a NoSuchColumnFamilyException
once it goes to the region server.
> 2) At some point, the buffer is flushed.
> 3) When the invalid put is processed, an exception is thrown. The finally clause of flushCommits()
removes all successful puts from the writebuffer list but the failed put remains at the top.
This entry becomes an immovable blocker which prevents any subsequent entry from being processed.

> 4) Subsequent puts will add more entries to the write buffer until the buffer limit is
reached, compounding the problem by allowing more edits to be queued which can never be processed.

> A workaround could be for the client to call getWriteBuffer() -- on trunk -- and remove
the entry at the head of the list manually, but without the patch on this issue, the client
cannot get access to the list on branch. 

-- 
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