hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Kihwal Lee (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-4495) Allow client-side lease renewal to be retried beyond soft-limit
Date Wed, 13 Feb 2013 16:46:13 GMT

    [ https://issues.apache.org/jira/browse/HDFS-4495?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13577700#comment-13577700

Kihwal Lee commented on HDFS-4495:

Allowing clients to renew leases past the soft limit is acceptable. The lease soft limit is
used by lease recovery and the way it works guarantees that no two writers are writing to
the same file or block.  When lease recovery is triggered by create() the file is recreated,
so any activity on the old RBW block won't interfere with the new file. If append() causes
a lease recovery, a block recovery is forced against the last RBW block and until it is done
client gets RecoveryInProgressException.  As a part of block recovery, existing writers are
terminated and a new generation number is stamped on the block, so any existing client won't
be able to write any longer.  A client explicitly calling recoverLease() also works similarly.

> Allow client-side lease renewal to be retried beyond soft-limit
> ---------------------------------------------------------------
>                 Key: HDFS-4495
>                 URL: https://issues.apache.org/jira/browse/HDFS-4495
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: hdfs-client
>    Affects Versions: 2.0.3-alpha, 0.23.5
>            Reporter: Kihwal Lee
> The current lease renewal code in DFSClient gives up after several retries, if the soft
limit exceeds. This causes the client to abort.  Without this self destruction behavior, lease
renewal can be retried infinitely due to unrelated low-level issue and prevent DFSClient and
associated Connection from getting garbage collected.  
> However, giving up at passage of the soft limit has been shown to be too fragile; A long
GC on namenode or transient network outage can make clients fail. We need the self-destruct
behavior, but clients should be allowed to retry for longer period of time.

This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

View raw message