[ 
https://issues.apache.org/jira/browse/CASSANDRA-2662?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13035597#comment-13035597
 ] 

Brandon Williams commented on CASSANDRA-2662:
---------------------------------------------

Thinking this through a bit more:

Given coordinator A, and replicas X, Y, and Z (in subsnitch order), on the 
first round X will be chosen, and let's say it receives a score of 1.  With the 
patch, at this point Y and Z will be initialized with zero.  On the next round, 
Y will be chosen, and let's say it receives a score of or near 1, depending on 
network latency.  On the third round, Z will be chosen, and let's say it also 
receives a score similar to Y.  Now the cache is hot on all nodes, and 
subsequent reads have the possibility to oscillate between all three based on 
network latency variance.  This can be mitigated though with the badness 
threshold.  With the badness threshold on, the first round will occur as 
before, but subsequent rounds will continue to use X until it degrades past the 
threshold, at which point they will use Y, until the dynamic snitch reset()s, 
at which point everything will repeat.  I don't think this is harmful to 
CASSANDRA-1314 after all.



> Nodes get ignored by dynamic snitch when read repair chance is zero
> -------------------------------------------------------------------
>
>                 Key: CASSANDRA-2662
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-2662
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Core
>    Affects Versions: 0.7.6
>            Reporter: Daniel Doubleday
>            Assignee: Brandon Williams
>            Priority: Trivial
>         Attachments: dynsnitch.patch
>
>
> DynamicEndpointSnitch falls back to subsnitch when one of the scores of the 
> endpoints being compared is missing.
> This leads to a stable order of hosts until reads will lead to recorded 
> scores. 
> If setting read repair chance to 0 and reads are performed with quorum then 
> (rf - # quorum nodes) will never get reads.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to