I think there are a couple of factors that make transformations and connectors different.
First, NiFi's 150 processors is a bit misleading. In NiFi, processors cover data sources, data sinks, serialization/deserialization, *and* transformations. I haven't filtered the list to see how many fall into the first 3 categories, but it's a *lot* of the processors they have. Second, since transformations only apply to a single message and I'd think they generally shouldn't be interacting with external services (i.e. I think trying to do enrichment in SMT is probably a bad idea), the scope of possible transformations is reasonably limited and the transformations themselves tend to be small and easily maintainable. I think this is a dramatic difference from connectors, which are each substantial projects in their own right. While I get the slippery slope argument re: including specific transformations, I think we can come up with a reasonable policy (and via KIPs we can, as a community, come to an agreement based purely on taste if it comes down to that). In particular, I'd say keep the core general (i.e. no domain-specific transformations/parsing like HL7), pure data manipulation (i.e. no enrichment), and nothing that could just as well be done as a converter/serializer/deserializer/source connector/sink connector. I was very staunchly against including connectors (aside from a simple example) directly in Kafka, so this may seem like a reversal of position. But I think the % of use cases covered will look very different between connectors and transformations. Sure, some connectors are very popular, and moreso right now because they are the most thoroughly developed, tested, etc. But the top 3 most common transformations will probably be used across all the top 20 most popular connectors. I have no doubt people will end up writing custom ones (which is why it's nice to make them pluggable rather than choosing a fixed set), but they'll either be very niche (like people write custom connectors for their internal systems) or be more broadly applicable but very domain specific such that they are easy to reject for inclusion. @Gwen if we filtered the list of NiFi processors to ones that fit that criteria, would that still be too long a list for your taste? Similarly, let's say we were going to include some baked in; in that case, does anything look out of place to you in the list Shikhar has included in the KIP? -Ewen On Thu, Dec 15, 2016 at 2:01 PM, Gwen Shapira <g...@confluent.io> wrote: > I agree about the ease of use in adding a small-subset of built-in > transformations. > > But the same thing is true for connectors - there are maybe 5 super popular > OSS connectors and the rest is a very long tail. We drew the line at not > adding any, because thats the easiest and because we did not want to turn > Kafka into a collection of transformations. > > I really don't want to end up with 135 (or even 20) transformations in > Kafka. So either we have a super-clear definition of what belongs and what > doesn't - or we put in one minimal example and the rest goes into the > ecosystem. > > We can also start by putting transformations on github and just see if > there is huge demand for them in Apache. It is easier to add stuff to the > project later than to remove functionality. > > > > On Thu, Dec 15, 2016 at 11:59 AM, Shikhar Bhushan <shik...@confluent.io> > wrote: > > > I have updated KIP-66 > > https://cwiki.apache.org/confluence/display/KAFKA/KIP- > > 66%3A+Single+Message+Transforms+for+Kafka+Connect > > with > > the changes I proposed in the design. > > > > Gwen, I think the main downside to not including some transformations > with > > Kafka Connect is that it seems less user friendly if folks have to make > > sure to have the right transformation(s) on the classpath as well, > besides > > their connector(s). Additionally by going in with a small set included, > we > > can encourage a consistent configuration and implementation style and > > provide utilities for e.g. data transformations, which I expect we will > > definitely need (discussed under 'Patterns for data transformations'). > > > > It does get hard to draw the line once you go from 'none' to 'some'. To > get > > discussion going, if we get agreement on 'none' vs 'some', I added a > table > > under 'Bundled transformations' for transformations which I think are > worth > > including. > > > > For many of these, I have noticed their absence in the wild as a pain > point > > -- > > TimestampRouter: > > https://github.com/confluentinc/kafka-connect-elasticsearch/issues/33 > > Mask: > > https://groups.google.com/d/msg/confluent-platform/3yHb8_ > > mCReQ/sTQc3dNgBwAJ > > Insert: > > http://stackoverflow.com/questions/40664745/elasticsearch-connector-for- > > kafka-connect-offset-and-timestamp > > RegexRouter: > > https://groups.google.com/d/msg/confluent-platform/ > > yEBwu1rGcs0/gIAhRp6kBwAJ > > NumericCast: > > https://github.com/confluentinc/kafka-connect- > > jdbc/issues/101#issuecomment-249096119 > > TimestampConverter: > > https://groups.google.com/d/msg/confluent-platform/ > > gGAOsw3Qeu4/8JCqdDhGBwAJ > > ValueToKey: https://github.com/confluentinc/kafka-connect-jdbc/pull/166 > > > > In other cases, their functionality is already being implemented by > > connectors in divergent ways: RegexRouter, Insert, Replace, > HoistToStruct, > > ExtractFromStruct > > > > On Wed, Dec 14, 2016 at 6:00 PM Gwen Shapira <g...@confluent.io> wrote: > > > > I'm a bit concerned about adding transformations in Kafka. NiFi has 150 > > processors, presumably they are all useful for someone. I don't know if > I'd > > want all of that in Apache Kafka. What's the downside of keeping it out? > Or > > at least keeping the built-in set super minimal (Flume has like 3 > built-in > > interceptors)? > > > > Gwen > > > > On Wed, Dec 14, 2016 at 1:36 PM, Shikhar Bhushan <shik...@confluent.io> > > wrote: > > > > > With regard to a), just using `ConnectRecord` with `newRecord` as a new > > > abstract method would be a fine choice. In prototyping, both options > end > > up > > > looking pretty similar (in terms of how transformations are implemented > > and > > > the runtime initializes and uses them) and I'm starting to lean towards > > not > > > adding a new interface into the mix. > > > > > > On b) I think we should include a small set of useful transformations > > with > > > Connect, since they can be applicable across different connectors and > we > > > should encourage some standardization for common operations. I'll > update > > > KIP-66 soon including a spec of transformations that I believe are > worth > > > including. > > > > > > On Sat, Dec 10, 2016 at 11:52 PM Ewen Cheslack-Postava < > > e...@confluent.io> > > > wrote: > > > > > > If anyone has time to review here, it'd be great to get feedback. I'd > > > imagine that the proposal itself won't be too controversial -- keeps > > > transformations simple (by only allowing map/filter), doesn't affect > the > > > rest of the framework much, and fits in with general config structure > > we've > > > used elsewhere (although ConfigDef could use some updates to make this > > > easier...). > > > > > > I think the main open questions for me are: > > > > > > a) Is TransformableRecord worth it to avoid reimplementing small bits > of > > > code (it allows for a single implementation of the interface to > trivially > > > apply to both Source and SinkRecords). I think I prefer this, but it > does > > > come with some commitment to another interface on top of ConnectRecord. > > We > > > could alternatively modify ConnectRecord which would require fewer > > changes. > > > b) How do folks feel about built-in transformations and the set that > are > > > mentioned here? This brings us way back to the discussion of built-in > > > connectors. Transformations, especially when intended to be lightweight > > and > > > touch nothing besides the data already in the record, seem different > from > > > connectors -- there might be quite a few, but hopefully limited. Since > we > > > (hopefully) already factor out most serialization-specific stuff via > > > Converters, I think we can keep this pretty limited. That said, I have > no > > > doubt some folks will (in my opinion) abuse this feature to do data > > > enrichment by querying external systems, so building a bunch of > > > transformations in could potentially open the floodgates, or at least > > make > > > decisions about what is included vs what should be 3rd party muddy. > > > > > > -Ewen > > > > > > > > > On Wed, Dec 7, 2016 at 11:46 AM, Shikhar Bhushan <shik...@confluent.io > > > > > wrote: > > > > > > > Hi all, > > > > > > > > I have another iteration at a proposal for this feature here: > > > > https://cwiki.apache.org/confluence/display/KAFKA/ > > > > Connect+Transforms+-+Proposed+Design > > > > > > > > I'd welcome your feedback and comments. > > > > > > > > Thanks, > > > > > > > > Shikhar > > > > > > > > On Tue, Aug 2, 2016 at 7:21 PM Ewen Cheslack-Postava < > > e...@confluent.io> > > > > wrote: > > > > > > > > On Thu, Jul 28, 2016 at 11:58 PM, Shikhar Bhushan < > > shik...@confluent.io> > > > > wrote: > > > > > > > > > > > > > > > > > > > > > > Hmm, operating on ConnectRecords probably doesn't work since you > > need > > > > to > > > > > > emit the right type of record, which might mean instantiating a > new > > > > one. > > > > > I > > > > > > think that means we either need 2 methods, one for SourceRecord, > > one > > > > for > > > > > > SinkRecord, or we'd need to limit what parts of the message you > can > > > > > modify > > > > > > (e.g. you can change the key/value via something like > > > > > > transformKey(ConnectRecord) and transformValue(ConnectRecord), > but > > > > other > > > > > > fields would remain the same and the fmwk would handle allocating > > new > > > > > > Source/SinkRecords if needed) > > > > > > > > > > > > > > > > Good point, perhaps we could add an abstract method on > ConnectRecord > > > that > > > > > takes all the shared fields as parameters and the implementations > > > return > > > > a > > > > > copy of the narrower SourceRecord/SinkRecord type as appropriate. > > > > > Transformers would only operate on ConnectRecord rather than caring > > > about > > > > > SourceRecord or SinkRecord (in theory they could instanceof/cast, > but > > > the > > > > > API should discourage it) > > > > > > > > > > > > > > > > Is there a use case for hanging on to the original? I can't think > > of > > > a > > > > > > transformation where you'd need to do that (or couldn't just > order > > > > things > > > > > > differently so it isn't a problem). > > > > > > > > > > > > > > > Yeah maybe this isn't really necessary. No strong preference here. > > > > > > > > > > That said, I do worry a bit that farming too much stuff out to > > > > transformers > > > > > > can result in "programming via config", i.e. a lot of the > > simplicity > > > > you > > > > > > get from Connect disappears in long config files. Standardization > > > would > > > > > be > > > > > > nice and might just avoid this (and doesn't cost that much > > > implementing > > > > > it > > > > > > in each connector), and I'd personally prefer something a bit > less > > > > > flexible > > > > > > but consistent and easy to configure. > > > > > > > > > > > > > > > Not sure what the you're suggesting :-) Standardized config > > properties > > > > for > > > > > a small set of transformations, leaving it upto connectors to > > > integrate? > > > > > > > > > > > > > I just mean that you get to the point where you're practically > writing > > a > > > > Kafka Streams application, you're just doing it through either an > > > > incredibly convoluted set of transformers and configs, or a single > > > > transformer with incredibly convoluted set of configs. You basically > > get > > > to > > > > the point where you're config is a mini DSL and you're not really > > saving > > > > that much. > > > > > > > > The real question is how much we want to venture into the "T" part of > > > ETL. > > > > I tend to favor minimizing how much we take on since the rest of > > Connect > > > > isn't designed for it, it's designed around the E & L parts. > > > > > > > > -Ewen > > > > > > > > > > > > > Personally I'm skeptical of that level of flexibility in > transformers > > > -- > > > > > > its getting awfully complex and certainly takes us pretty far > from > > > > > "config > > > > > > only" realtime data integration. It's not clear to me what the > use > > > > cases > > > > > > are that aren't covered by a small set of common transformations > > that > > > > can > > > > > > be chained together (e.g. rename/remove fields, mask values, and > > > maybe > > > > a > > > > > > couple more). > > > > > > > > > > > > > > > > I agree that we should have some standard transformations that we > > ship > > > > with > > > > > connect that users would ideally lean towards for routine tasks. > The > > > ones > > > > > you mention are some good candidates where I'd imagine can expose > > > simple > > > > > config, e.g. > > > > > transform.filter.whitelist=x,y,z # filter to a whitelist of > > fields > > > > > transfom.rename.spec=oldName1=>newName1, oldName2=>newName2 > > > > > topic.rename.replace=-/_ > > > > > topic.rename.prefix=kafka_ > > > > > etc.. > > > > > > > > > > However the ecosystem will invariably have more complex > transformers > > if > > > > we > > > > > make this pluggable. And because ETL is messy, that's probably a > good > > > > thing > > > > > if folks are able to do their data munging orthogonally to > > connectors, > > > so > > > > > that connectors can focus on the logic of how data should be copied > > > > from/to > > > > > datastores and Kafka. > > > > > > > > > > > > > > > > In any case, we'd probably also have to change configs of > > connectors > > > if > > > > > we > > > > > > allowed configs like that since presumably transformer configs > will > > > > just > > > > > be > > > > > > part of the connector config. > > > > > > > > > > > > > > > > Yeah, haven't thought much about how all the configuration would > tie > > > > > together... > > > > > > > > > > I think we'd need the ability to: > > > > > - spec transformer chain (fully-qualified class names? perhaps > > special > > > > > aliases for built-in ones? perhaps third-party fqcns can be > assigned > > > > > aliases by users in the chain spec, for easier configuration and to > > > > > uniquely identify a transformation when it occurs more than one > time > > in > > > a > > > > > chain?) > > > > > - configure each transformer -- all properties prefixed with that > > > > > transformer's ID (fqcn / alias) get destined to it > > > > > > > > > > Additionally, I think we would probably want to allow for > > > topic-specific > > > > > overrides <https://issues.apache.org/jira/browse/KAFKA-3962> (e.g. > > you > > > > > want > > > > > certain transformations for one topic, but different ones for > > > another...) > > > > > > > > > > > > > > > > > > > > > -- > > > > Thanks, > > > > Ewen > > > > > > > > > > > > > > > -- > > *Gwen Shapira* > > Product Manager | Confluent > > 650.450.2760 <(650)%20450-2760> | @gwenshap > > Follow us: Twitter <https://twitter.com/ConfluentInc> | blog > > <http://www.confluent.io/blog> > > > > > > -- > *Gwen Shapira* > Product Manager | Confluent > 650.450.2760 | @gwenshap > Follow us: Twitter <https://twitter.com/ConfluentInc> | blog > <http://www.confluent.io/blog> >
