[ 
https://issues.apache.org/jira/browse/FLINK-8043?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16331007#comment-16331007
 ] 

Steven Zhen Wu edited comment on FLINK-8043 at 1/19/18 3:24 PM:
----------------------------------------------------------------

[~till.rohrmann] 

one question: why is  _fullRestarts_ a _Gauge_ while _task___failures_ a 
_Counter_? Either both should be Gauge or both should be Counter. they capture 
similar info. also a minor nitpick on inconsistent naming pattern (camel case 
v.s. underscore)?

Regarding the usability problem: _fullRestarts_ as a _Gauge_ is  good for 
showing on UI. But it is probably not suitable for threshold based alerting. if 
we set up alert like "fullRestarts > 0 happens 10 times in last 15 minutes", it 
will always be true once a single full restart happens. Threshold based 
alerting only works with counter type metric. 


was (Author: stevenz3wu):
[~till.rohrmann] 

one question: why is  _fullRestarts_ a _Gauge_ while _task___failures_ a 
_Counter_? also naming pattern is inconsistent?

Regarding the usability problem: _fullRestarts_ as a _Gauge_ is  good for 
showing on UI. But it is probably not suitable for threshold based alerting. if 
we set up alert like "fullRestarts > 0 happens 10 times in last 15 minutes", it 
will always be true once a single full restart happens. Threshold based 
alerting only works with counter type metric. 

> change fullRestarts (for fine grained recovery) from guage to counter
> ---------------------------------------------------------------------
>
>                 Key: FLINK-8043
>                 URL: https://issues.apache.org/jira/browse/FLINK-8043
>             Project: Flink
>          Issue Type: Bug
>          Components: ResourceManager
>    Affects Versions: 1.3.2
>            Reporter: Steven Zhen Wu
>            Priority: Blocker
>             Fix For: 1.4.1
>
>
> Fine grained recovery publish fullRestarts as guage, which is not suitable 
> for threshold based alerting. Usually we would alert like "fullRestarts > 0 
> happens 10 times in last 15 minutes".
> In comparison, "task_failures" is published as counter.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to