Thank you for your inputs Gwen and Michael. The original reason why I suggested a set based processing is because of the flexibility is provides. The JIRA had a comment by a user requesting a feature that could be achieved with this.
After reading Gwen and Michael's points, (I went through the documentation and the code in detail) and agree with what you have to say. Also, fewer guarantees make what I had in mind less certain and thus simplifying it to a single message based transformation would ensure that users who do require more flexibility with the transformations will automatically “turn to" Kafka Streams. The transformation logic on a message by message basis makes more sense. One usecase that Kafka Connect could consider is adding or removing a message completely. (This was trivially possible with the collection passing). Should users be pointed towards Kafka Streams even for this use case? I think this is a very useful feature for Connect too, and I’ll try to rethink on the API too. Removing a message is as easy as returning a null and having the next transformer skip it, but adding messages would involve say a queue between transformers and a method which says “pass message” to the next, which can be called multiple times from one “transform” function; a variation on the chain of responsibility design pattern. > On Jul 12, 2016, at 12:54 AM, Michael Noll <mich...@confluent.io> wrote: > > As Gwen said, my initial thought is that message transformations that are > "more than trivial" should rather be done by Kafka Streams, rather than by > Kafka Connect (for the reasons that Gwen mentioned). > > Transforming one message at a time would be a good fit for Kafka Connect. > An important use case is to remove sensitive data (such as PII) from an > incoming data stream before it hits Kafka's persistent storage -- this use > case can't be implemented well with Kafka Streams because, by design, Kafka > Streams is meant to read its input data from Kafka (i.e. at the point when > Kafka Streams could be used to removed sensitive data fields the data is > already stored persistently in Kafka, and this might be a no-go depending > on the use case). > > I'm of course interested to hear what other people think. > > > On Tue, Jul 12, 2016 at 6:06 AM, Gwen Shapira <g...@confluent.io> wrote: > >> I think we need to restrict the functionality to one-message-at-a-time. >> >> Basically, connect gives very little guarantees about the size of the set >> of the composition (you may get same messages over and over, mix of old and >> new, etc) >> >> In order to do useful things over a collection, you need better defined >> semantics of what's included. Kafka Streams is putting tons of effort into >> having good windowing semantics, and I think apps that require modification >> of collections are a better fit there. >> >> I'm willing to change my mind though (have been known to happen) - what are >> the comments about usage that point toward the collections approach? >> >> Gwen >> >> On Mon, Jul 11, 2016 at 3:32 PM, Nisarg Shah <snis...@gmail.com> wrote: >> >>> Thanks Jay, added that to the KIP. >>> >>> Besides reviewing the KIP as a whole, I wanted to know about what >> everyone >>> thinks about what data should be dealt at the Transformer level. >> Transform >>> the whole Collection of Records (giving the flexibility of modifying >>> messages across the set) OR >>> Transform messages one at a time, iteratively. This will restrict >>> modifications across messages. >>> >>> I’ll get a working sample ready soon, to have a look. There were some >>> comments about Transformer usage that pointed to the first approach, >> which >>> I prefer too given the flexibility. >>> >>>> On Jul 11, 2016, at 2:49 PM, Jay Kreps <j...@confluent.io> wrote: >>>> >>>> One minor thing, the Transformer interface probably needs a close() >>> method >>>> (i.e. the opposite of initialize). This would be used for any >> transformer >>>> that uses a resource like a file/socket/db connection/etc that needs to >>> be >>>> closed. You usually don't need this but when you do need it you really >>> need >>>> it. >>>> >>>> -Jay >>>> >>>> On Mon, Jul 11, 2016 at 1:47 PM, Nisarg Shah <snis...@gmail.com> >> wrote: >>>> >>>>> Hello, >>>>> >>>>> This KIP < >>>>> >>> >> https://cwiki.apache.org/confluence/display/KAFKA/KIP-66:+Add+Kafka+Connect+Transformers+to+allow+transformations+to+messages >>>> >>>>> is for KAFKA-3209 <https://issues.apache.org/jira/browse/KAFKA-3209>. >>>>> It’s about capabilities to transform messages in Kafka Connect. >>>>> >>>>> Some design decisions need to be taken, so please advise me on the >> same. >>>>> Feel free to express any thoughts or concerns as well. >>>>> >>>>> Many many thanks to Ewen Cheslack-Postava. >>>>> >>>>> -Nisarg >>> >>> >> > > > > -- > Best regards, > Michael Noll > > > > *Michael G. Noll | Product Manager | Confluent | +1 650.453.5860Download > Apache Kafka and Confluent Platform: www.confluent.io/download > <http://www.confluent.io/download>*
