hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Raghu Angadi (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-4681) DFSClient block read failures cause open DFSInputStream to become unusable
Date Wed, 27 May 2009 19:15:45 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-4681?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12713701#action_12713701
] 

Raghu Angadi commented on HADOOP-4681:
--------------------------------------

> Isn't that what this patch effects (+/- the 2x dfs.client.max.block.acquire.failures
semantics, which are a little odd)?

sort of, I think. It is an improvement, but only on the same lines as the current structure
(no explicit policy, hard to see what happens and when).  For e.g. it looks like 'a failure'
is defined as not being able to connect/fetch from all the replicas. Also success is defined
as just being able to get 'ok' from datanode. Which implies it might again go into infinite
loop in some conditions (say all replicas are corrupt, thus only one is remaining, and it
fails checksum at one particular position).

The final fix could be as simiple as moving resetting 'failures' to readBuffer().  But I would
prefer to add things like 'failures' to be incremented for every failure with a replica and
limit the count to something like min(#replicas, large value like 10).

Finally, I am mainly proposing an explicit policy in a code comment. 0.21 is fine.. big users
like HBASE can have it backported.

> DFSClient block read failures cause open DFSInputStream to become unusable
> --------------------------------------------------------------------------
>
>                 Key: HADOOP-4681
>                 URL: https://issues.apache.org/jira/browse/HADOOP-4681
>             Project: Hadoop Core
>          Issue Type: Bug
>          Components: dfs
>    Affects Versions: 0.18.2, 0.19.0, 0.19.1, 0.20.0
>            Reporter: Igor Bolotin
>             Fix For: 0.19.2
>
>         Attachments: 4681.patch
>
>
> We are using some Lucene indexes directly from HDFS and for quite long time we were using
Hadoop version 0.15.3.
> When tried to upgrade to Hadoop 0.19 - index searches started to fail with exceptions
like:
> 2008-11-13 16:50:20,314 WARN [Listener-4] [] DFSClient : DFS Read: java.io.IOException:
Could not obtain block: blk_5604690829708125511_15489 file=/usr/collarity/data/urls-new/part-00000/20081110-163426/_0.tis
> at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.chooseDataNode(DFSClient.java:1708)
> at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.blockSeekTo(DFSClient.java:1536)
> at org.apache.hadoop.hdfs.DFSClient$DFSInputStream.read(DFSClient.java:1663)
> at java.io.DataInputStream.read(DataInputStream.java:132)
> at org.apache.nutch.indexer.FsDirectory$DfsIndexInput.readInternal(FsDirectory.java:174)
> at org.apache.lucene.store.BufferedIndexInput.refill(BufferedIndexInput.java:152)
> at org.apache.lucene.store.BufferedIndexInput.readByte(BufferedIndexInput.java:38)
> at org.apache.lucene.store.IndexInput.readVInt(IndexInput.java:76)
> at org.apache.lucene.index.TermBuffer.read(TermBuffer.java:63)
> at org.apache.lucene.index.SegmentTermEnum.next(SegmentTermEnum.java:131)
> at org.apache.lucene.index.SegmentTermEnum.scanTo(SegmentTermEnum.java:162)
> at org.apache.lucene.index.TermInfosReader.scanEnum(TermInfosReader.java:223)
> at org.apache.lucene.index.TermInfosReader.get(TermInfosReader.java:217)
> at org.apache.lucene.index.SegmentTermDocs.seek(SegmentTermDocs.java:54) 
> ...
> The investigation showed that the root of this issue is that we exceeded # of xcievers
in the data nodes and that was fixed by changing configuration settings to 2k.
> However - one thing that bothered me was that even after datanodes recovered from overload
and most of client servers had been shut down - we still observed errors in the logs of running
servers.
> Further investigation showed that fix for HADOOP-1911 introduced another problem - the
DFSInputStream instance might become unusable once number of failures over lifetime of this
instance exceeds configured threshold.
> The fix for this specific issue seems to be trivial - just reset failure counter before
reading next block (patch will be attached shortly).
> This seems to be also related to HADOOP-3185, but I'm not sure I really understand necessity
of keeping track of failed block accesses in the DFS client.

-- 
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