Riak IRC

danlarkin # I asked about this last night, but I'll try again now in a slightly different way,
I've got many threads (on different machines) between which I want to coordinate state
I thought about using riak, but I can't find any user-facing atomic operations
like a CaS or forcing creation of siblings, even if the writes happen discretely, or something

seancribbs # danlarkin: no, Riak is eventually consistent. If you want to consider an operation atomic, it is only within the context of a single request (and not within the storage mechanism)
requests do not lock anything either, so you may still end up with race-conditions

danlarkin # I would be okay with commutativity
so eventually consistent is ok

seancribbs # commutativity?

danlarkin # the insertion order doesn't matter, just that nothing is overwritten

seancribbs # ah yes. you can bubble up update conflicts to the application
which can decide how to resolve/merge them

danlarkin # but conflicts only occur when writes happen concurrently, not discretely, right?

seancribbs # or when there is a partition, i.e. network failure

danlarkin # so one thread can still end up overwriting what another thread wrote

seancribbs # you can simulate a conflict by storing an object multiple times with the same parent vclock
no, we won't overwrite. siblings will be created whether you choose to see those siblings is up to you

danlarkin # so.. two threads, first one reads the document, makes a change, and writes it back -- but inbetween the read and write the other thread reads _and_ writes... would those changes be overwritten by the first thread's last-write-winning?

seancribbs # that's what I'm saying, it won't be. if you set "allow_mult" to true on the bucket, both versions will be written. then in your next read, you will be shown both versions

danlarkin # I see, I misunderstood then exxxcellent
seancribbs # if allow_mult is false (the default), only the latest version will be shown (even if siblings exist)

danlarkin # so how does riak know there's a "conflict", the client sends the vclock from its read back when it writes?

seancribbs # yes

danlarkin # perfect thanks sean