[
https://issues.apache.org/jira/browse/HDFS-14758?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16918839#comment-16918839
]
Kihwal Lee commented on HDFS-14758:
-----------------------------------
Unconditionally calling recoverLease() on a failure will work when it is
guaranteed that there is only one writer. That feature can be selectively
turned on by clients like hbase. This JIRA is about the automatic lease
recovery and commit block synchronization. There is no point in waiting for a
long period of time before recovering them. If a block stays a long time in
open (aka under-construction) state, there is increased chances of losing data,
since replication is not scheduled for such blocks.
Making it configurable and having a 20 minute default sound reasonable.
> Decrease lease hard limit
> -------------------------
>
> Key: HDFS-14758
> URL: https://issues.apache.org/jira/browse/HDFS-14758
> Project: Hadoop HDFS
> Issue Type: Improvement
> Reporter: Eric Payne
> Assignee: hemanthboyina
> Priority: Minor
>
> The hard limit is currently hard-coded to be 1 hour. This also determines the
> NN automatic lease recovery interval. Something like 20 min will make more
> sense.
> After the 5 min soft limit, other clients can recover the lease. If no one
> else takes the lease away, the original client still can renew the lease
> within the hard limit. So even after a NN full GC of 8 minutes, leases can be
> still valid.
> However, there is one risk in reducing the hard limit. E.g. Reduced to 20
> min. If the NN crashes and the manual failover takes more than 20 minutes,
> clients will abort.
--
This message was sent by Atlassian Jira
(v8.3.2#803003)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
