[ https://issues.apache.org/jira/browse/HDFS-7725?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14302345#comment-14302345
]
Ming Ma commented on HDFS-7725:
-------------------------------
Thanks, [~zhz]. Yes, the logic of "don't modify nn stats if the node is dead" can be moved
to HeartbeatManager.
For the trunk version, we can wait for HDFS-7411. If HDFS-7411 isn't going to be in branch-2
anytime soon, then we will need some quick fix. Overall, can anyone find any correctness issue
with the current patch?
> Incorrect "nodes in service" metrics caused all writes to fail
> --------------------------------------------------------------
>
> Key: HDFS-7725
> URL: https://issues.apache.org/jira/browse/HDFS-7725
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Ming Ma
> Assignee: Ming Ma
> Attachments: HDFS-7725.patch
>
>
> One of our clusters sometimes couldn't allocate blocks from any DNs. BlockPlacementPolicyDefault
complains with the following messages for all DNs.
> {noformat}
> the node is too busy (load:x > y)
> {noformat}
> It turns out the {{HeartbeatManager}}'s {{nodesInService}} was computed incorrectly when
admins decomm or recomm dead nodes. Here are two scenarios.
> * Decomm dead nodes. It turns out HDFS-7374 has fixed it; not sure if it is intentional.
cc / [~zhz], [~andrew.wang], [~atm] Here is the sequence of event without HDFS-7374.
> ** Cluster has one live node. nodesInService == 1
> ** The node becomes dead. nodesInService == 0
> ** Decomm the node. nodesInService == -1
> * However, HDFS-7374 introduces another inconsistency when recomm is involved.
> ** Cluster has one live node. nodesInService == 1
> ** The node becomes dead. nodesInService == 0
> ** Decomm the node. nodesInService == 0
> ** Recomm the node. nodesInService == 1
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
|