[
https://issues.apache.org/jira/browse/HADOOP-940?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12476080
]
Hairong Kuang commented on HADOOP-940:
--------------------------------------
This patch sets the time out period to be 10 minutes and the pending
replication monitor thread scans the pending replication queue for time out
blocks once every 1 minute. Because the chance that a block does not get
replicated is rare, I feel that the monitor thread scans too frequently. Does
it make more sense to set the time out period to be 5 minutes and the scanning
frequency to be once every 5 minutes?
Each entry in the pendingReplication queue is a block with only one timer. When
a replica is placed, the timer associated with the block is reset. This may
cause that the re-replication of a replica gets delayed. For example, if a
replica's replication request gets lost but another replica of the same block
is placed right before its timer gets expired, the lost request won't be
detected until twice of the timeout period is passed.
> pendingReplications of FSNamesystem is not informative
> ------------------------------------------------------
>
> Key: HADOOP-940
> URL: https://issues.apache.org/jira/browse/HADOOP-940
> Project: Hadoop
> Issue Type: Improvement
> Components: dfs
> Affects Versions: 0.10.1
> Reporter: Hairong Kuang
> Attachments: pendingReplication.patch
>
>
> Currently when a neededReplication block is scheduled to be replicated, it is
> put to the pendingReplications queue. When it is no longer under replicated,
> it is pulled out of the pendingReplications queue. But the queue does not
> provide any information like how many targets have been choosen or who those
> targets are. PendingReplications are not used when deciding if a block is
> under replication. This may cause a block to be over replications or
> inaccurate estimate of its replication priority.
> For example, when a block has 1 replicas but it's replication factor is 2, a
> data node is choosen to replicate this block and the block is put in the
> pendingReplications queue. If the block's replication factor is changed to be
> 3 before the block replication notification, which is the next block report,
> comes in, the block will be put into neededReplictions queue again under the
> assumption that it needs to choose 2 targets instead of 1. So the block will
> end up with 4 replicas.
> I propose that we change pendingReplications to be a map from a block to the
> choosen data nodes. Data nodes in both pendingReplications and blockMap are
> used when deciding the total number of replicas that a block has. When the
> name node is notified that the block is replicated in a choosen data node,
> the data node is moved from pendingReplications to blockMap.
> Each choosen target is also associated with a timer indicating how long it
> expects to receive the block replication notification. PendingReplications
> queue needs to be periodically scanned to remove those data nodes whose timer
> is expired.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
