The per-key generalization is useful. As Jiangjie mentioned in KAFKA-2260, one thing that we need to sort out is what happens if a produce request has messages with different keys and some of the messages have expected offsets while some others don't. Currently, the produce response has an error code per partition, not per message. One way is to just define the semantics as: the produce request will only go through if all keys in the request pass the offset test.
Thanks, Jun On Sat, Jul 18, 2015 at 10:37 AM, Ben Kirwin <b...@kirw.in> wrote: > Just wanted to flag a little discussion that happened on the ticket: > > https://issues.apache.org/jira/browse/KAFKA-2260?focusedCommentId=14632259&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14632259 > > In particular, Yasuhiro Matsuda proposed an interesting variant on > this that performs the offset check on the message key (instead of > just the partition), with bounded space requirements, at the cost of > potentially some spurious failures. (ie. the produce request may fail > even if that particular key hasn't been updated recently.) This > addresses a couple of the drawbacks of the per-key approach mentioned > at the bottom of the KIP. > > On Fri, Jul 17, 2015 at 6:47 PM, Ben Kirwin <b...@kirw.in> wrote: > > Hi all, > > > > So, perhaps it's worth adding a couple specific examples of where this > > feature is useful, to make this a bit more concrete: > > > > - Suppose I'm using Kafka as a commit log for a partitioned KV store, > > like Samza or Pistachio (?) do. We bootstrap the process state by > > reading from that partition, and log all state updates to that > > partition when we're running. Now imagine that one of my processes > > locks up -- GC or similar -- and the system transitions that partition > > over to another node. When the GC is finished, the old 'owner' of that > > partition might still be trying to write to the commit log at the same > > as the new one is. A process might detect this by noticing that the > > offset of the published message is bigger than it thought the upcoming > > offset was, which implies someone else has been writing to the log... > > but by then it's too late, and the commit log is already corrupt. With > > a 'conditional produce', one of those processes will have it's publish > > request refused -- so we've avoided corrupting the state. > > > > - Envision some copycat-like system, where we have some sharded > > postgres setup and we're tailing each shard into its own partition. > > Normally, it's fairly easy to avoid duplicates here: we can track > > which offset in the WAL corresponds to which offset in Kafka, and we > > know how many messages we've written to Kafka already, so the state is > > very simple. However, it is possible that for a moment -- due to > > rebalancing or operator error or some other thing -- two different > > nodes are tailing the same postgres shard at once! Normally this would > > introduce duplicate messages, but by specifying the expected offset, > > we can avoid this. > > > > So perhaps it's better to say that this is useful when a single > > producer is *expected*, but multiple producers are *possible*? (In the > > same way that the high-level consumer normally has 1 consumer in a > > group reading from a partition, but there are small windows where more > > than one might be reading at the same time.) This is also the spirit > > of the 'runtime cost' comment -- in the common case, where there is > > little to no contention, there's no performance overhead either. I > > mentioned this a little in the Motivation section -- maybe I should > > flesh that out a little bit? > > > > For me, the motivation to work this up was that I kept running into > > cases, like the above, where the existing API was almost-but-not-quite > > enough to give the guarantees I was looking for -- and the extension > > needed to handle those cases too was pretty small and natural-feeling. > > > > On Fri, Jul 17, 2015 at 4:49 PM, Ashish Singh <asi...@cloudera.com> > wrote: > >> Good concept. I have a question though. > >> > >> Say there are two producers A and B. Both producers are producing to > same > >> partition. > >> - A sends a message with expected offset, x1 > >> - Broker accepts is and sends an Ack > >> - B sends a message with expected offset, x1 > >> - Broker rejects it, sends nack > >> - B sends message again with expected offset, x1+1 > >> - Broker accepts it and sends Ack > >> I guess this is what this KIP suggests, right? If yes, then how does > this > >> ensure that same message will not be written twice when two producers > are > >> producing to same partition? Producer on receiving a nack will try again > >> with next offset and will keep doing so till the message is accepted. > Am I > >> missing something? > >> > >> Also, you have mentioned on KIP, "it imposes little to no runtime cost > in > >> memory or time", I think that is not true for time. With this approach > >> producers' performance will reduce proportionally to number of producers > >> writing to same partition. Please correct me if I am missing out > something. > >> > >> > >> On Fri, Jul 17, 2015 at 11:32 AM, Mayuresh Gharat < > >> gharatmayures...@gmail.com> wrote: > >> > >>> If we have 2 producers producing to a partition, they can be out of > order, > >>> then how does one producer know what offset to expect as it does not > >>> interact with other producer? > >>> > >>> Can you give an example flow that explains how it works with single > >>> producer and with multiple producers? > >>> > >>> > >>> Thanks, > >>> > >>> Mayuresh > >>> > >>> On Fri, Jul 17, 2015 at 10:28 AM, Flavio Junqueira < > >>> fpjunque...@yahoo.com.invalid> wrote: > >>> > >>> > I like this feature, it reminds me of conditional updates in > zookeeper. > >>> > I'm not sure if it'd be best to have some mechanism for fencing > rather > >>> than > >>> > a conditional write like you're proposing. The reason I'm saying > this is > >>> > that the conditional write applies to requests individually, while it > >>> > sounds like you want to make sure that there is a single client > writing > >>> so > >>> > over multiple requests. > >>> > > >>> > -Flavio > >>> > > >>> > > On 17 Jul 2015, at 07:30, Ben Kirwin <b...@kirw.in> wrote: > >>> > > > >>> > > Hi there, > >>> > > > >>> > > I just added a KIP for a 'conditional publish' operation: a simple > >>> > > CAS-like mechanism for the Kafka producer. The wiki page is here: > >>> > > > >>> > > > >>> > > >>> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-27+-+Conditional+Publish > >>> > > > >>> > > And there's some previous discussion on the ticket and the users > list: > >>> > > > >>> > > https://issues.apache.org/jira/browse/KAFKA-2260 > >>> > > > >>> > > > >>> > > >>> > https://mail-archives.apache.org/mod_mbox/kafka-users/201506.mbox/%3CCAAeOB6ccyAA13YNPqVQv2o-mT5r=c9v7a+55sf2wp93qg7+...@mail.gmail.com%3E > >>> > > > >>> > > As always, comments and suggestions are very welcome. > >>> > > > >>> > > Thanks, > >>> > > Ben > >>> > > >>> > > >>> > >>> > >>> -- > >>> -Regards, > >>> Mayuresh R. Gharat > >>> (862) 250-7125 > >>> > >> > >> > >> > >> -- > >> > >> Regards, > >> Ashish >
