[
https://issues.apache.org/jira/browse/CASSANDRA-8686?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17726700#comment-17726700
]
Stefan Miklosovic commented on CASSANDRA-8686:
----------------------------------------------
I implemented it here (1) (the last commit there).
I introduced a new option called "-target". Currently, it has "latency" target
and "saturation" type.
"saturation" type preserves same behavior as before - it will stop to increase
the workload if we do not see any meaningful improvements for some number of
rounds.
"latency" target type takes percentiles and their latencies like this
{code:java}
cassandra-stress \
write n=1M
cl=QUORUM \
no-warmup \
-mode native cql3 \
-rate threads\>=100 threads\<=1500 threadIncrease=200 auto \
-target type=latency percentiles=0.95=25,0.99=50,0.999=100{code}
"threadIncrease" is my invention too. I do not like that it doubled the number
of threads and it was raising "too quickly".
So the outcome here is that after StressAction is done, it will look into what
the percentiles were and if one of them violated the target, the whole stress
test will end.
Keep in mind that percentiles can be anything, like:
{code:java}
percentiles=0.80=25,0.90=50,0.95=100,0.999=150{code}
[~brandon.williams] would you mind to take a look at this one too, please?
(1) https://github.com/apache/cassandra/pull/2373
> Introduce Latency Target for Stress
> -----------------------------------
>
> Key: CASSANDRA-8686
> URL: https://issues.apache.org/jira/browse/CASSANDRA-8686
> Project: Cassandra
> Issue Type: Improvement
> Components: Tool/stress
> Reporter: jonathan lacefield
> Assignee: Stefan Miklosovic
> Priority: Low
> Labels: stress
> Time Spent: 10m
> Remaining Estimate: 0h
>
> This item is a request to add a latency target to the rate option for the new
> stress tool. The goal of the latency target would be to provide a guideline
> for SLAs to the stress tool so the stress tool can determine threads and
> throughputs that can be sustained while meeting the SLA targets.
> For example:
> cassandra-stress [options/commands] -rate latency p90=5 p95=10 p99=100
> The outcome of this command would be a stress execution that would gradually
> increase threads, and hence throughput (trans/sec), until the latency profile
> can no longer be satisfied with the current workload (yaml file definition)
> and/or cluster. This would provide a ceiling for throughput and connections
> for the given cluster, workload, and SLA profile.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
