[ https://issues.apache.org/jira/browse/CASSANDRA-3868?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13214313#comment-13214313 ]
Peter Schuller commented on CASSANDRA-3868: ------------------------------------------- We have been saved by this in production at least once, as it allows a cluster to absorb counters without doing reads in the write path. It's a very useful feature in that it "turns off" the one behavior of counters which is contrary to the usual write-centric design of Cassandra. Not replicating for a while can easily be a much better option than not being able to absorb the write traffic at all, since you're still eventually consistent even if you're significantly increasing the risk of dropping increments. In particular, I would note that whenever CL.ONE is used, it can correctly be satisfied without replication. We even discussed briefly maybe introducing an automatic skipping of the replicate-on-write step, if the stage is backed up with pending, in order to get more graceful degradation - that's how useful it was for us, and is expected to be for use-cases where you have large amounts of data and high write throughput (the issue being that the read-on-write is so *massively* more expensive than the write-only path, if you end up going down to platters). I'm strongly in dis-favor of removing this option for the shear practical applicability of it, and I don't believe in protecting the user from mistakes in cases like this. It seems unlikely that people accidentally turn off replicate on write off (unless documentation somehow directs them to). Given that the removal is basically just the handling of the option and such, rather than removing large chunks of code that someone has to maintain, I don't see a major disadvantage to keeping it. > Remove or nullify replicate_on_write option > ------------------------------------------- > > Key: CASSANDRA-3868 > URL: https://issues.apache.org/jira/browse/CASSANDRA-3868 > Project: Cassandra > Issue Type: Improvement > Components: Core > Affects Versions: 0.8.0 > Reporter: Brandon Williams > Fix For: 1.1.0 > > Attachments: 3868.txt > > > My understanding from Sylvain is that setting this option to false is rather > dangerous/stupid, and you should basically never do it. So 1.1 is a good > time to get rid of it, or make it a no-op. -- 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