cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chris Goffinet (Commented) (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-3518) Back pressure users by request/s instead of concurrent reads/writes
Date Tue, 22 Nov 2011 07:08:39 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-3518?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13154920#comment-13154920
] 

Chris Goffinet commented on CASSANDRA-3518:
-------------------------------------------

Radim,

I know what the Request Scheduler is.. please read my ticket more carefully. 

# Scheduler Options vary based on the type of scheduler
# NoScheduler - Has no options
# RoundRobin
#  - throttle_limit -- The throttle_limit is the number of in-flight
#                      requests per client.  Requests beyond 
#                      that limit are queued up until
#                      running requests can complete.
#                      The value of 80 here is twice the number of
#                      concurrent_reads + concurrent_writes.

number of in-flight requests per client. I want to a) do this per user b) break it down so
it's request/s not how many concurrent reads/writes in flight.

                
> Back pressure users by request/s instead of concurrent reads/writes
> -------------------------------------------------------------------
>
>                 Key: CASSANDRA-3518
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-3518
>             Project: Cassandra
>          Issue Type: Bug
>    Affects Versions: 1.0.2
>            Reporter: Chris Goffinet
>
> We are running into use cases where it makes a lot of sense to have QoS at the request
level per user. Imagine this case:
> I have a cluster that can do 100,000 req/s. But I want to limit the user to only being
able to do either 50,000 read or write/s per second in the cluster. I rather give back pressure
to the user then make the cluster fall down because the user tried to take down my cluster.

> Also another case we have is where you have experimental features and want to give access
to certain group of customers and let them run experiments on data. You *dont* want them taking
down the cluster, you rather make them fail fast, or slow them down. If I could limit a user
to N req/s for reads or writes, instead of adding back pressure based on # of concurrent requests
in each stage, this would go a long way for us.
> We have had a few incidents where spinning up new features caused unexpected load and
we couldn't stop them without turning the feature off. 

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Mime
View raw message