[
https://issues.apache.org/jira/browse/KAFKA-656?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13587423#comment-13587423
]
Jay Kreps commented on KAFKA-656:
---------------------------------
Yes, that is basic idea. The concern with ewma is just the opacity of it as an
SLA. I kind of prefer picking a hard window of (say) 1-5 seconds and caping the
requests/messages/etc in that window to a hard, configurable limit (e.g. 1000k)
per client or topic. This is somewhat more clear about why you are in
violation. EWMA also means a weighting over an infinite history of prior values
which is not intuitive. For example a long history of zero messages/second does
not justify a 5 second burst of exceptional load. The value of ewma is to
smooth the estimate when data is sparse (otherwise you end up measuring "sample
variance" and for low traffic things get a line that jumps around erratically).
However this is okay for SLA violation because anyone approaching their SLA
doesn't have a data sparsity problem--they are making tons of requests.
> Add Quotas to Kafka
> -------------------
>
> Key: KAFKA-656
> URL: https://issues.apache.org/jira/browse/KAFKA-656
> Project: Kafka
> Issue Type: New Feature
> Components: core
> Affects Versions: 0.8.1
> Reporter: Jay Kreps
> Labels: project
>
> It would be nice to implement a quota system in Kafka to improve our support
> for highly multi-tenant usage. The goal of this system would be to prevent
> one naughty user from accidently overloading the whole cluster.
> There are several quantities we would want to track:
> 1. Requests pers second
> 2. Bytes written per second
> 3. Bytes read per second
> There are two reasonable groupings we would want to aggregate and enforce
> these thresholds at:
> 1. Topic level
> 2. Client level (e.g. by client id from the request)
> When a request hits one of these limits we will simply reject it with a
> QUOTA_EXCEEDED exception.
> To avoid suddenly breaking things without warning, we should ideally support
> two thresholds: a soft threshold at which we produce some kind of warning and
> a hard threshold at which we give the error. The soft threshold could just be
> defined as 80% (or whatever) of the hard threshold.
> There are nuances to getting this right. If you measure second-by-second a
> single burst may exceed the threshold, so we need a sustained measurement
> over a period of time.
> Likewise when do we stop giving this error? To make this work right we likely
> need to charge against the quota for request *attempts* not just successful
> requests. Otherwise a client that is overloading the server will just flap on
> and off--i.e. we would disable them for a period of time but when we
> re-enabled them they would likely still be abusing us.
> It would be good to a wiki design on how this would all work as a starting
> point for discussion.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
