[
https://issues.apache.org/jira/browse/HDFS-11708?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16027652#comment-16027652
]
Konstantin Shvachko commented on HDFS-11708:
--------------------------------------------
Hey guys, I reproduced failure of
{{TestPread#testPreadFailureWithChangedBlockLocations}} without the change to
{{DFSInputStream}}.
But only if I run the test case individually. If I run the entire {{TestPread}}
everything passes. Could you fix this please.
I see now that {{refreshLocatedBlock()}} uses cached locations. It seems like a
rather fancy way to update a block variable that got stale. Would've preferred
something more explicit, but oh well. Two observations:
# Should you call {{refreshLocatedBlock()}} after {{chooseDataNode()}}, not
before? {{chooseDataNode()}} can update the block, but we will still work with
the old locations outside.
# So in this case we will call {{refreshLocatedBlock()}} twice on each
iteration, one inside {{chooseDataNode()}} and then one outside. Is it possible
to avoid this overhead? Something like updating {{LocatedBlock block}}
parameter inside {{chooseDataNode()}} by setting new values rather than
re-assigning new reference, which will make {{block}} input/output parameter
rather than input-only.
> positional read will fail if replicas moved to different DNs after stream is
> opened
> -----------------------------------------------------------------------------------
>
> Key: HDFS-11708
> URL: https://issues.apache.org/jira/browse/HDFS-11708
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs-client
> Affects Versions: 2.7.3
> Reporter: Vinayakumar B
> Assignee: Vinayakumar B
> Priority: Critical
> Labels: release-blocker
> Attachments: HDFS-11708-01.patch, HDFS-11708-02.patch,
> HDFS-11708-03.patch, HDFS-11708-04.patch, HDFS-11708-05.patch
>
>
> Scenario:
> 1. File was written to DN1, DN2 with RF=2
> 2. File stream opened to read and kept. Block Locations are [DN1,DN2]
> 3. One of the replica (DN2) moved to another datanode (DN3) due to datanode
> dead/balancing/etc.
> 4. Latest block locations in NameNode will be DN1 and DN3 in the 'same order'
> 5. DN1 went down, but not yet detected as dead in NameNode.
> 6. Client start reading using positional read api "read(pos, buf[], offset,
> length)"
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
