[
https://issues.apache.org/jira/browse/HDFS-12282?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16137891#comment-16137891
]
Anu Engineer commented on HDFS-12282:
-------------------------------------
bq. SCMNodeManager#handleHeartbeat is already a worker thread to process HB
queue, running in certain interval at background, even there is heavy I/O, it
won't affect HB performance.
I am afraid that might not be true. Imagine this scenario, the Disk I/O is
incredibly slow. If this thread does not finish execution, then SCM's view of
the cluster will be stale. That is machines which are *really* dead will not be
marked as dead for example.
Now, these dead nodes would be assumed to be alive and clients will be told to
do I/O to these nodes. Hence over all error rates in the cluster would shoot up.
Let us take a step back, I wanted to understand something more high level. Why
do we need this in the path of anything heart beat related? May be I am missing
something here.
> Ozone: DeleteKey-4: Block delete via HB between SCM and DN
> ----------------------------------------------------------
>
> Key: HDFS-12282
> URL: https://issues.apache.org/jira/browse/HDFS-12282
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: datanode, ozone, scm
> Reporter: Weiwei Yang
> Assignee: Weiwei Yang
> Attachments: Block delete via HB between SCM and DN.png,
> HDFS-12282-HDFS-7240.001.patch, HDFS-12282-HDFS-7240.002.patch,
> HDFS-12282-HDFS-7240.003.patch, HDFS-12282.WIP.patch
>
>
> This is the task 3 in the design doc, implements the SCM to datanode
> interactions. Including
> # SCM sends block deletion message via HB to datanode
> # datanode changes block state to deleting when processes the HB response
> # datanode sends deletion ACKs back to SCM
> # SCM handles ACKs and removes blocks in DB
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
