Hi Jonathan and Jun, Transactional replication using Kafka between stores at either end is an interesting topic. I have some experience with this problem in database replication products.
To understand how to implement it properly in Kafka it would help to define Jonathan's use case more formally. As I read the description there are three parts: a source DBMS, Kafka, and an analytics store. These can be arranged as follows: Producer Store -> Kafka -> Consumer Store e.g.: MySQL -> Kafka -> Spark over HDFS This is like the usual producer/consumer model except that the semantics are as follows. I added some details to the description to accommodate a number of practical problems that occur in replication topologies of this kind. 1.) The producer and consumer in the topology are stores with state and some notion of a transaction that changes the state of the store to which they are applied. Kafka is in the middle and also has transactions, namely to produce and consume messages. 2.) If a transaction executes on the producer store, you would like to execute a corresponding transaction on the consumer store. The transaction might not have the same effect downstream but the point is that transactions are linked one-to-one between producer and consumer. 3.) All of the stores or Kafka can fail independently and at any time. It must be possible to recover and continue once a failed component restarts. 4.) It is possible to have failures where a store or Kafka itself loses committed state and reverts to an earlier state. This happens in MySQL for example, when a host crashes before data are properly committed to InnoDB and/or the MySQL binlog. It can also happen if the upstream DBMS is restored from a backup or as a result of cluster failover with data loss. In this case you either want to regenerate lost transactions or (if it is hopeless) fail cleanly. 5.) Producer transactions might be larger than a single Kafka message (e.g. a KeyedMessage). They may not even fit into a single JVM's memory. This can occur for example if you do a bulk load or an administrative operation on a large table in the producer store. You might not have this problem now but given your requirement to work with a range of stores it seems likely to occur sooner rather than later. Such transactions must be broken into a stream of smaller messages with a protocol to identify that they belong to a single transaction. If there are failures such fragmented transactions must not result in partial transactions being applied to the consumer. 6.) All of the preceding requirements should be met with minimal impact on message throughput or transaction rates within stores at either end. Let me know if this is more than what you (Jonathan) intended. Usually if you really want #2, requirements #3-6 follow automatically. #5 is potentially a source of much pain if not addressed early on. Pending a response, I would just say solutions that require a transactional commit across two stores are difficult to write, often have performance downsides, and handle failures poorly because they cannot cover all the corner cases. The last point means they tend to drop data, generate unmatched transactions (orphans), or send things multiple times depending on the failure. It's generally better to design systems that use a sliding window protocol where a commit in the producer triggers a commit to Kafka triggers a commit to the consumer. Assuming your requirements are as stated above the question is how to design a transactional sliding window protocol that works on Kafka. Cheers, Robert Hodges On Thu, Jun 5, 2014 at 7:48 AM, Jun Rao <[email protected]> wrote: > It sounds like that you want to write to a data store and a data pipe > atomically. Since both the data store and the data pipe that you want to > use are highly available, the only case that you want to protect is the > client failing btw the two writes. One way to do that is to let the client > publish to Kafka first with the strongest ack. Then, run a few consumers to > read data from Kafka and then write the data to the data store. Any one of > those consumers can die and the work will be automatically picked up by the > remaining ones. You can use partition id and the offset of each message as > its UUID if needed. > > Thanks, > > Jun > > > On Wed, Jun 4, 2014 at 10:56 AM, Jonathan Hodges <[email protected]> > wrote: > > > Sorry didn't realize the mailing list wasn't copied... > > > > > > ---------- Forwarded message ---------- > > From: Jonathan Hodges <[email protected]> > > Date: Wed, Jun 4, 2014 at 10:56 AM > > Subject: Re: Hadoop Summit Meetups > > To: Neha Narkhede <[email protected]> > > > > > > We have a number of customer facing online learning applications. These > > applications are using heterogeneous technologies with different data > > models in underlying data stores such as RDBMS, Cassandra, MongoDB, etc. > > We would like to run offline analysis on the data contained in these > > learning applications with tools like Hadoop and Spark. > > > > One thought is to use Kafka as a way for these learning applications to > > emit data in near real-time for analytics. We developed a common model > > represented as Avro records in HDFS that spans these learning > applications > > so that we can accept the same structured message from them. This allows > > for comparing apples to apples across these apps as opposed to messy > > transformations. > > > > So this all sounds good until you dig into the details. One pattern is > for > > these applications to update state locally in their data stores first and > > then publish to Kafka. The problem with this is these two operations > > aren't atomic so the local persist can succeed and the publish to Kafka > > fail leaving the application and HDFS out of sync. You can try to add > some > > retry logic to the clients, but this quickly becomes very complicated and > > still doesn't solve the underlying problem. > > > > Another pattern is to publish to Kafka first with -1 and wait for the ack > > from leader and replicas before persisting locally. This is probably > > better than the other pattern but does add some complexity to the client. > > The clients must now generate unique entity IDs/UUID for persistence > when > > they typically rely on the data store for creating these. Also the > publish > > to Kafka can succeed and persist locally can fail leaving the stores out > of > > sync. In this case the learning application needs to determine how to > get > > itself in sync. It can rely on getting this back from Kafka, but it is > > possible the local store failure can't be fixed in a timely manner e.g. > > hardware failure, constraint, etc. In this case the application needs to > > show an error to the user and likely need to do something like send a > > delete message to Kafka to remove the earlier published message. > > > > A third last resort pattern might be go the CDC route with something like > > Databus. This would require implementing additional fetchers and relays > to > > support Cassandra and MongoDB. Also the data will need to be transformed > > on the Hadoop/Spark side for virtually every learning application since > > they have different data models. > > > > I hope this gives enough detail to start discussing transactional > messaging > > in Kafka. We are willing to help in this effort if it makes sense for > our > > use cases. > > > > Thanks > > Jonathan > > > > > > > > On Wed, Jun 4, 2014 at 9:44 AM, Neha Narkhede <[email protected]> > > wrote: > > > > > If you are comfortable, share it on the mailing list. If not, I'm happy > > to > > > have this discussion privately. > > > > > > Thanks, > > > Neha > > > On Jun 4, 2014 9:42 AM, "Neha Narkhede" <[email protected]> > wrote: > > > > > >> Glad it was useful. It will be great if you can share your > requirements > > >> on atomicity. A couple of us are very interested in thinking about > > >> transactional messaging in Kafka. > > >> > > >> Thanks, > > >> Neha > > >> On Jun 4, 2014 6:57 AM, "Jonathan Hodges" <[email protected]> wrote: > > >> > > >>> Hi Neha, > > >>> > > >>> Thanks so much to you and the Kafka team for putting together the > > meetup. > > >>> It was very nice and gave people from out of town like us the > ability > > to > > >>> join in person. > > >>> > > >>> We are the guys from Pearson Education and we talked a little about > > >>> supplying some details on some of our use cases with respect to > > atomicity > > >>> of source systems eventing data and persisting locally. Should we > just > > >>> post to the list or is there somewhere else we should send these > > details? > > >>> > > >>> Thanks again! > > >>> Jonathan > > >>> > > >>> > > >>> > > >>> On Fri, Apr 11, 2014 at 9:31 AM, Neha Narkhede < > > [email protected]> > > >>> wrote: > > >>> > > >>> > Yes, that's a great idea. I can help organize the meetup at > LinkedIn. > > >>> > > > >>> > Thanks, > > >>> > Neha > > >>> > > > >>> > > > >>> > On Fri, Apr 11, 2014 at 8:44 AM, Saurabh Agarwal (BLOOMBERG/ 731 > > >>> LEXIN) < > > >>> > [email protected]> wrote: > > >>> > > > >>> > > great idea. I am interested in attending as well.... > > >>> > > > > >>> > > ----- Original Message ----- > > >>> > > From: [email protected] > > >>> > > To: [email protected] > > >>> > > At: Apr 11 2014 11:40:56 > > >>> > > > > >>> > > With the Hadoop Summit in San Jose 6/3 - 6/5 I wondered if any of > > the > > >>> > > LinkedIn geniuses were thinking of putting together a meet-up on > > any > > >>> of > > >>> > the > > >>> > > associated technologies like Kafka, Samza, Databus, etc. For us > > poor > > >>> > souls > > >>> > > that don't live on the West Coast it was a great experience > > >>> attending the > > >>> > > Kafka meetup last year. > > >>> > > > > >>> > > Jonathan > > >>> > > > > >>> > > > > >>> > > > > >>> > > > > >>> > > > > >>> > > > >>> > > > ------------------------------------------------------------------------------- > > >>> > > > > >>> > > > >>> > > >> > > >
