[
https://issues.apache.org/jira/browse/HDFS-16040?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17352032#comment-17352032
]
Simbarashe Dzinamarira commented on HDFS-16040:
-----------------------------------------------
Took a second look to check is the rpcQueueTime is being calculated correctly.
I believe it is.
{noformat}
details.set(Timing.QUEUE,
startTime - timestampNanos - details.get(Timing.ENQUEUE));{noformat}
startTime is when the call is removed from the call queue for processing. (The
final time it is actually removed, not intermediate pops withs re-queuing)
timestampNanos is when the call was received.
details.get(Timing.ENQUEUE) is the time it took to initially place the call on
the call queue. (Doesn't count the re-queues in the rpc server).
The logic above will correctly count the times between pops and re-queues as
part of the rpcQueueTime.
> RpcQueueTime metric counts requeued calls as unique events.
> -----------------------------------------------------------
>
> Key: HDFS-16040
> URL: https://issues.apache.org/jira/browse/HDFS-16040
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs
> Affects Versions: 2.10.0, 3.3.0
> Reporter: Simbarashe Dzinamarira
> Assignee: Simbarashe Dzinamarira
> Priority: Major
> Attachments: HDFS-16040.001.patch, HDFS-16040.002.patch,
> HDFS-16040.003.patch
>
>
> The RpcQueueTime metric is updated every time a call is re-queued while
> waiting for the server state to reach the call's client's state ID. This is
> in contrast to RpcProcessingTime which is only updated when the call if
> finally processed.
> On the Observer NameNode this can result in RpcQueueTimeNumOps being much
> larger than RpcProcessingTimeNumOps. The re-queueing is an internal
> optimization to avoid blocking and shouldn't result in an inflated metric.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
