[
https://issues.apache.org/jira/browse/KAFKA-1597?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14102480#comment-14102480
]
Alexis Midon commented on KAFKA-1597:
-------------------------------------
# NetworkProcessorAvgIdlePercent and RequestHandlerAvgIdlePercent don't give
information regarding response processing
# could you explain what RequestBeingProcessed is? how different is it from the
number of I/O threads? My understanding of the code is that requests are
processed by the I/O threads process requests.
https://github.com/apache/kafka/blob/0.8.1/core/src/main/scala/kafka/server/KafkaServer.scala#L90
https://dchtm6r471mui.cloudfront.net/hackpad.com_SIX7UAFo7hB_p.186494_1405370457117_KafkaThreadingModel.png
# Regarding InflightRequests. A request flows through a sequence of
transformations/states. To me, measuring every component of that pipeline is
the most basic and reliable way of understanding the behavior of the system. It
also gives users the option to use the data in any way they might want. So i'd
that exposing these root metrics would be great. Considering that only
BeingSentResponses is missing, the implementation cost is pretty low.
> New metrics: ResponseQueueSize and BeingSentResponses
> -----------------------------------------------------
>
> Key: KAFKA-1597
> URL: https://issues.apache.org/jira/browse/KAFKA-1597
> Project: Kafka
> Issue Type: New Feature
> Components: core
> Reporter: Alexis Midon
> Priority: Minor
> Labels: features
> Attachments: KAFKA-1594_BeingSentResponses.patch,
> KAFKA-1594_ResponseQueueSize.patch
>
>
> This patch adds two metrics:
> h3. ResponseQueueSize
> As of 0.8.1, the sizes of the response queues are [reported as different
> metrics|https://github.com/apache/kafka/blob/0.8.1/core/src/main/scala/kafka/network/RequestChannel.scala#L127-L134]
> - one per processor thread. This is not very ideal for different reasons:
> * charts have to sum the different metrics
> * the metrics collection system might not support 'wild card queries' like
> {{sum:kafka.network.RequestChannel.Processor_*_ResponseQueueSize}} in which
> case monitoring now depends on the number of configured network threads
> * monitoring the response by thread is not very valuable. However the global
> number of responses is useful.
> * proposal*
> So this patch exposes the total number of queued responses as a metric
> {{ResponseQueueSize}}
> *implementation*
> In {{RequestChannel}}, create a Gauge that adds up the size of the response
> queues.
> h3. BeingSentResponses
> As of 0.8.1, the processor threads will poll responses from the queues and
> attach them to the SelectionKey as fast as possible. The consequence of that
> is that the response queues are not a good indicator of the number of
> "in-flight" responses. The {{ServerSocketChannel}} acting as another queue of
> response to be sent.
> The current metrics don't reflect the size of this "buffer", which is an
> issue.
> *proposal*
> This patch adds a gauge that keeps track of the number of responses being
> handled by the {{ServerSocketChannel}}.
> That new metric is named "BeingSentResponses" (who said naming was hard?)
> *implementation*
> To calculate that metric, the patch adds up the number of SelectionKeys
> interested in writing, across processor threads.
> Another approach could be to keep all in-flight responses in a data structure
> (let's say a map) shared by the processor threads. A response will be added
> to that map when dequeued from the response queue, and removed when the write
> is complete. The gauge will simply report the size of that map. I decided
> against that second approach as it is more intrusive and requires some
> additional bookkeeping to gather information already available through the
> {{SelectionKey}}'s
--
This message was sent by Atlassian JIRA
(v6.2#6252)
