[
https://issues.apache.org/jira/browse/HDFS-6184?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13958273#comment-13958273
]
Ming Ma commented on HDFS-6184:
-------------------------------
Thanks Jing. To clarify the service handler suggestion, if we can set
servicerpc handler thread count larger than the # of DNs, even though all DNs'
servicerpc calls are blocked by SBN write lock, HAServiceProtocol.monitorHealth
rpc can still be processed by another idle service handler thread. However, in
a big production cluster, it will be expensive to set servicerpc handler thread
count larger than the # of DNs.
> Better health check from ZKFC
> -----------------------------
>
> Key: HDFS-6184
> URL: https://issues.apache.org/jira/browse/HDFS-6184
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: namenode
> Reporter: Ming Ma
>
> We have seen several false positives in terms of when ZKFC considers NN to be
> unhealthy. Some of these triggers unnecessary failover. Examples,
> 1. SBN checkpoint caused ZKFC's RPC call into NN timeout. The consequence
> isn't bad; just that SBN will quit ZK membership and rejoin it later. But it
> is unnecessary. The reason is checkpoint acquires NN global write lock and
> all rpc requests are blocked. Even though HAServiceProtocol.monitorHealth
> doesn't need to acquire NN lock; it still needs to user service rpc queue.
> 2. When ANN is busy, sometimes the global lock can block other requests.
> ZKFC's RPC call timeout. This will trigger failover. The question is even if
> after the failover, the new ANN might run into similar issue.
> We can increase ZKFC to NN timeout value to mitigate this to some degree. If
> ZKFC can be more accurate in judgment if NN is health or not and can predict
> the failover will help, that will be useful. For example, we can,
> 1. Have ZKFC made decision based on NN thread dump.
> 2. Have a dedicated rpc pool for ZKFC > NN. Given health check doesn't need
> to acquire NN global lock; so it can go through even if NN is doing
> checkpointing or very busy.
> Any comments?
--
This message was sent by Atlassian JIRA
(v6.2#6252)
