Elias, There's quite a few questions in here, but hopefully I'll answer the meat of what you're trying to find out. Using `if_not_modified` is _not_ equivalent to a distributed atomic compare-and-swap (CAS). As you've correctly identified, concurrent writers on different coordinators can all succeed in writing different values, while still meeting the prerequisite for `if_not_modified`.
Reid On Nov 6, 2012, at 2:06 PM, Elias Levy <fearsome.lucid...@gmail.com> wrote: > The description of if_not_modified is lacking in the documentation. What > guarantees if_not_modified provides is a question that has been asked in the > list a few times, but for which no satisfactory answer has been given. Its > something I've been wanted answered as well, so I did a bit of digging into > the code. > > I gather from reading the code that whatever node receives the INM request > perform an internal get to retrieve the object to get its vclock to compare > to the incoming object. The get appears to use the R value that is the > default for the bucket. The default value of R for a bucket is "quorum", > unless changed, although the default for PR is 0. Thus, INM can strongly > guarantee no stale writes, but only if no primary replica nodes have failed > or have become partitioned. That also assumes that you are using at least an > W of "quorum", and ideally a PR of the same. > > Correct me if I am wrong, but there appears no way to set the R and PR values > to be used by the internal get during a INM put request, except for changing > the bucket default R and PR values. > > The INM only guarantees no stale writes, and only does so when R, PR, W, and > WR are such to to guarantee strong consistency. It does not guard against > concurrent writes. Multiple writers can succeed even when using INM if they > write at the same time. > > If you have allow_mult set to true in the bucket, that means the client must > handle sibling resolution. If allow_mult is set to false and last_write_wins > is set to true, concurrent writes will succeed, but the value with the latest > timestamp will win. If last_write_wins is set to false while allow_multi is > also false, Riak will keep siblings internally, but use vector clocks to only > return a single value. > > Its unclear from the documentation what it does in the case of conflicts to > resolve the siblings into the single value. The docs say you'll receive the > "latest" value, but does not say how latest is determined in the case of a > conflict. > > Elias Levy > > _______________________________________________ > riak-users mailing list > riak-users@lists.basho.com > http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com _______________________________________________ riak-users mailing list riak-users@lists.basho.com http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
