[ 
https://issues.apache.org/jira/browse/SAMZA-421?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14200301#comment-14200301
 ] 

Martin Kleppmann commented on SAMZA-421:
----------------------------------------

How many messages did you process during this profiling run? I have no doubt 
that the code is fast when it's only processed a small number of messages. My 
concern was that it would slow down or run out of memory when it has collected 
5 minutes worth of measurements (at 1M msgs/sec, that would be 300M 
measurements). Not sure how much memory a ConcurrentSkipListMap<Long, Long> 
with 300M entries will take, but probably a couple of gigs. I still feel that's 
kinda heavyweight for a metrics system.

The overhead wouldn't necessarily show up in RunLoop.process(), but also when 
the metrics system periodically takes a snapshot of the measurements, and in GC.

> Test the performance before and after adding the Timer metric
> -------------------------------------------------------------
>
>                 Key: SAMZA-421
>                 URL: https://issues.apache.org/jira/browse/SAMZA-421
>             Project: Samza
>          Issue Type: Test
>          Components: metrics
>            Reporter: Yan Fang
>            Assignee: Yan Fang
>              Labels: newbie
>         Attachments: testTimer1.png, update-timer-speed.png
>
>
> As 
> [recommended|https://issues.apache.org/jira/browse/SAMZA-349?focusedCommentId=14128365&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14128365]
>  by Martin in SAMZA-349, it's worth testing the performance before and after 
> implementing the Timer metric.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to