Hi Tom, With regards to the potential Transformation::validate method, I don't quite follow your objections. The AbstractHerder class, ConnectorConfig class, and embedding of ConfigDefs that happens with the two is all internal logic and we're allowed to modify it however we want, as long as it doesn't alter any public-facing APIs (and if it does, we just have to document those changes in a KIP). We don't have to embed the ConfigDef for a transformation chain inside another ConfigDef if the API we want to present to our users doesn't play well with that part of the code base. Additionally, beyond aligning with the existing API provided for connectors, another advantage is that it becomes possible to validate properties for a configuration in the context of all other properties, so to be clear, it's not just about preserving what may be perceived as a superficial similarity and comes with real functional benefits that can't be provided (at least, not as easily) by dynamic construction of a ConfigDef object.
As far as the new proposal goes, I hate to say it, but I think we're stuck with the worst of both worlds here. Adding the new RecordPredicate interface seems like it defeats the whole purpose of SMTs, which is to allow manipulation of a connector's event stream by users who don't necessarily know how or have the time to write Java code of their own. This is also why I'm in favor of adding a lightweight DSL for the condition; emphasizing readability for people who aren't very familiar with Connect and just want to get something going quickly should be a priority for SMTs. But if that's not going to happen with this KIP, I'd still choose the simpler, less-flexible approach initially outlined, in order to keep things simple for people creating connectors and try to let them accomplish what they want via configuration, not code. With regards to the question about where the line should be drawn and how much is too much and comparisons to other stream processing frameworks, I think the nature of SMTs draws the line quite nicely: you can only process one message at a time. There's plenty of use cases out there for heavier-duty processing frameworks like Kafka Streams, with aggregate operations, joining of streams, expanding a single message into multiple messages, etc. With SMTs, none of this is possible; the general use case is to filter and clean a stream of data. If any of the heavier-weight logic provided by, e.g., Streams, isn't required for a project, it should be possible to get along with just a collection of sink connectors, source connectors, a converter or two, and SMTs that smooth over any differences in data format between what source connectors produce and what sink connectors expect. This is why I'm comfortable suggesting heavy expansion of the out-of-the-box SMTs that we provide with Connect; as long as they're reasonable to configure, they can greatly reduce the operational burden on anyone running and/or using a Connect cluster since they might entirely replace additional services that would otherwise be required. All that said--this is a huge ask of someone who just wants to support the SMT equivalent of an "if" statement, and it's totally understandable if that's too much to ask. The one concern I have left is that if we expand the SMT in the future, there become compatibility concerns since SMTs are pretty tightly-coupled with the worker on which they run (although technically, with some classpath/plugin path shuffling, you can run different versions of the out-of-the-box SMTs from the Connect worker on which they're run). Someone might write a connector config with the If/Conditional/Whatever SMT with a condition type that works on one worker, but that doesn't work on another worker that's running an earlier version of Connect. This is why I'm in favor of adding extra predicates now instead of later; if we're going to implement what has the potential to be a bit of a Swiss army knife SMT with room for future expansion of configuration, and we can think of a reasonable way to add that functionality now, it seems better for users and administrators of Connect to try to do that now instead of later. Cheers, Chris On Wed, Apr 8, 2020 at 9:45 AM Tom Bentley <tbent...@redhat.com> wrote: > Since no one objected I've updated the KIP with the revised way to > configure this transformation. > > On Mon, Apr 6, 2020 at 2:52 PM Tom Bentley <tbent...@redhat.com> wrote: > > > To come back about a point Chris made: > > > > 1. Instead of the new "ConfigDef config(Map<String, String> props)" > method, > >> what would you think about adopting a similar approach as the framework > >> uses with connectors, by adding a "Config validate(Map<String, String> > >> props)" method that can perform custom validation outside of what can be > >> performed by the ConfigDef's single-property-at-a-time validation? It > may > >> be a little heavyweight for use with this particular SMT, but it'd > provide > >> more flexibility for other SMT implementations and would mirror an API > >> that > >> developers targeting the framework are likely already familiar with. > >> > > > > The validate() + config() approach taken for Connectors doesn't quite > work > > for Transformations. > > > > The default Connector.validate() basically just calls > > `config().validate(connectorConfigs)` and returns a Config. Separately > the > > AbstractHeader also calls `config()` and uses the Config and ConfigDef to > > build the ConfigInfos. The contract is not really defined well enough in > > the javadoc, but this means a connector can use validate() to build the > > ConfigDef which it will return from config(). > > > > For Transformations we don't really want validate() to perform validation > > at all since ultimately the transformations config will be embedded in > the > > connector's and will be validated by the connector itself. It wouldn't be > > harmful if it _did_ perform validation, just unnecessary. But having a > > validate() method that ought not to validate() seems like a recipe for > > confusion. > > > > Also problematic is that there's no use for the Config returned from > > Transformations.validate(). > > > > So I'm not convinced that the superficial similarity really gains > anything. > > > > Kind regards, > > > > Tom > > > > On Mon, Apr 6, 2020 at 2:29 PM Tom Bentley <tbent...@redhat.com> wrote: > > > >> Hi, > >> > >> Hi all, > >> > >> Thanks for the discussion so far. > >> > >> It seems a bit weird to me that when configuring the Conditional SMT > with > >> a DSL you would use a concise, intuitive DSL for expressing the > condition, > >> but not for the transforms that it's guarding. It also seems natural, if > >> you support this for conditionally applying SMTs, that you'd soon want > to > >> support the same thing for a generic filter transformer. Then, if you > can > >> configure a filter transformation using this DSL, it becomes odd that > you > >> can't do this for mapping transformations. I think it would be a > mistake to > >> go specifying an expression language for conditions when really that > might > >> just be part of a language for transformations. > >> > >> I think it would be possible today to write an SMT which allowed you to > >> express the transformation in a DSL. Concretely, it's possible to > imagine a > >> single transformation taking a DSL something like this: > >> > >> compose( > >> Flatten.Key(delimiter: ':'), > >> If(condition: TopicMatches(pattern: "blah-.*"), > >> then: Flatten.Value(delimiter: '/'))) > >> > >> All I've really done here, beyond what was already proposed, is rename > >> Conditional to If, imagine a couple more bits of syntax > >> (SMTs being constructed by class name and named argument invocation > >> syntax to support SMT constructors) and add a higher level compose() > >> function for chaining SMTs (which could easily be replaced with brace > >> delimited blocks). > >> > >> That may be a discussion we should have, but I think in _this_ KIP we > >> should focus on the conditional part, since the above example hopefully > >> shows that it would be possible to reuse it in a DSL if there was > appetite > >> for that. > >> > >> With that in mind, and since my original suggestion for an abbreviated > >> config syntax didn't appeal to people, I propose that we stick with the > >> existing norms for configuring this. > >> The previous example would look like this: > >> > >> transformation: flattenKey,if > >> transformation.flattenKey.type: Flatten.Key > >> transformation.flattenKey.delimiter: : > >> transformation.if.type: If > >> transformation.if.condition.type: TopicMatches > >> transformation.if.condition.pattern: blah-.* > >> transformation.if.then: flattenValue > >> transformation.if.then.flattenValue.type: Flatten.Value > >> transformation.if.then.flattenValue.delimiter: / > >> > >> Also, I'm inclined to think we should stick to just supporting > >> TopicMatches and Not, since we've not identified an actual need for > >> 'has-header', 'or' and 'and'. > >> > >> An example of the usage of Not would look like this: > >> > >> transformation: flattenKey,if > >> transformation.flattenKey.type: Flatten.Key > >> transformation.flattenKey.delimiter: : > >> transformation.if.type: If > >> transformation.if.condition.type: Not > >> transformation.if.condition.operand: startsWithBlah > >> transformation.if.condition.operand.startsWithBlah.type: TopicMatches > >> transformation.if.condition.operand.startsWithBlah.pattern: blah-.* > >> transformation.if.then: flattenValue > >> transformation.if.then.flattenValue.type: Flatten.Value > >> transformation.if.then.flattenValue.delimiter: / > >> > >> Hopefully we can agree that this allows the conditional SMT part to make > >> progress without getting bogged down. > >> > >> Thoughts? If this seems broadly reasonable, I'll update the KIP. > >> > >> Kind regards, > >> > >> Tom > >> > >> On Fri, Apr 3, 2020 at 5:55 PM Gunnar Morling <gun...@hibernate.org> > >> wrote: > >> > >>> Hi all, > >>> > >>> Thanks a lot for this initiative, Tom! > >>> > >>> To shed some light, the use case where this first came up, were issues > >>> we saw with SMTs being applied to the different topics produced by the > >>> Debezium change data capture connectors. There are different kinds of > >>> topics (for change data, schema history, heartbeats etc.) and the > >>> record structure to expect may vary between those. Hence we saw issues > >>> with SMTs like ExtractField, which for instance only should be applied > >>> to all change data topics but not the other ones. > >>> > >>> I like the overall approach; for Debezium's purposes, the simple topic > >>> matching and negation operators would be sufficient already. I agree > >>> with Chris and would prefer one single condition attribute, which > >>> contains a single condition or potentially a logical expression with > >>> not, and, etc. I think it's less ambiguous, in particular when it > >>> comes to ordering of the different conditions and determining their > >>> precedence. > >>> > >>> Would love to see this feature in one or another way in Connect. > >>> > >>> Best, > >>> > >>> --Gunnar > >>> > >>> > >>> > >>> Am Do., 2. Apr. 2020 um 18:48 Uhr schrieb Tom Bentley < > >>> tbent...@redhat.com>: > >>> > > >>> > Hi Chris and Sönke, > >>> > > >>> > Using the numbering from Chris's email... > >>> > > >>> > 1. That's a good point, I'll see what is needed to make that work. > >>> > > >>> > 2. I'm happy enough to add support for "and" and "or" as part of this > >>> KIP > >>> > if people can see a need for it. > >>> > > >>> > In a similar vein, I was wondering about whether it would be > worthwhile > >>> > having the equivalent of an "else" clause (what's in the KIP is > >>> basically > >>> > an "if" statement). Without support for "else" I think people would > >>> often > >>> > need two conditionals, with the condition of one being the negation > of > >>> the > >>> > condition of another. > >>> > > >>> > 3. I can see the attraction of an expression language. The pros > include > >>> > being terse and familiar to programmers and potentially very flexible > >>> if > >>> > that's needed in the future. I had a play and implemented it using > >>> ANTLR > >>> > and it's not difficult to write a grammar and implement the functions > >>> we've > >>> > already discussed and get decent error messages when the expression > is > >>> > malformed. So on the one hand I quite like the idea. On the other > hand > >>> it > >>> > feels like overkill for the use cases that have actually been > >>> identified so > >>> > far. > >>> > > >>> > @Sönke what do you make of the expression language idea? > >>> > > >>> > Kind regards, > >>> > > >>> > Tom > >>> > > >>> > On Wed, Apr 1, 2020 at 9:49 PM Christopher Egerton < > >>> chr...@confluent.io> > >>> > wrote: > >>> > > >>> > > Hi Tom, > >>> > > > >>> > > This looks great and I'd love to see the out-of-the-box SMTs become > >>> even > >>> > > more powerful with the improvements you've proposed! Got a few > >>> remarks and > >>> > > would be interested in your thoughts: > >>> > > > >>> > > 1. Instead of the new "ConfigDef config(Map<String, String> props)" > >>> method, > >>> > > what would you think about adopting a similar approach as the > >>> framework > >>> > > uses with connectors, by adding a "Config validate(Map<String, > >>> String> > >>> > > props)" method that can perform custom validation outside of what > >>> can be > >>> > > performed by the ConfigDef's single-property-at-a-time validation? > >>> It may > >>> > > be a little heavyweight for use with this particular SMT, but it'd > >>> provide > >>> > > more flexibility for other SMT implementations and would mirror an > >>> API that > >>> > > developers targeting the framework are likely already familiar > with. > >>> > > 2. The possibility for adding the logical operators "and" and "or" > is > >>> > > mentioned, but only as a potential future change and not one > >>> proposed by > >>> > > this KIP. Why not include those operators sooner rather than later? > >>> > > 3. The syntax for named conditions that are then referenced in > >>> logical > >>> > > operators is... tricky. It took me a few attempts to grok the > example > >>> > > provided in the KIP after reading Sönke's question about the > example > >>> for > >>> > > negation. What about a more sophisticated but less verbose syntax > >>> that > >>> > > supports a single configuration for the condition, even with > logical > >>> > > operators? I'm thinking something like > >>> > > "transforms.conditionalExtract.condition: > not(has-header:my-header)" > >>> > > instead of the "transforms.conditionalExtract.condition: > >>> not:hasMyHeader" > >>> > > and "transforms.conditionalExtract.condition.hasMyHeader: > >>> > > has-header:my-header" properties. If support for a logical "and" is > >>> added, > >>> > > this could then be expanded to something like > >>> > > "transforms.conditionalExtract.condition: > and(has-header(my-header), > >>> > > not(topic-matches(my-prefix-.*)))". There would be additional > >>> complexity > >>> > > here with the need to escape parentheses and commas that are > >>> intended to be > >>> > > treated literally (as part of a header name, for example) instead > of > >>> as > >>> > > part of the syntax for the condition itself, but a little > additional > >>> > > complexity for edge cases like that may be warranted if it heavily > >>> reduces > >>> > > complexity for the common cases. The rationale for the proposed > >>> > > parentheses-based syntax here instead of what's mentioned in the > KIP > >>> > > (something like "and: <condition1>, <condition2>") is to help with > >>> > > readability; we probably wouldn't need that with the approach of > >>> naming > >>> > > conditions via separate properties, but things may get a little > >>> nasty with > >>> > > literal conditions included there, especially with the possibility > of > >>> > > nesting. Finally, the shift in syntax here should make it easier to > >>> handle > >>> > > cases like the header value matching brought up by Sönke; it might > >>> look > >>> > > something like "header-matches(header-name, header-value-pattern)". > >>> > > > >>> > > Cheers, > >>> > > > >>> > > Chris > >>> > > > >>> > > On Wed, Apr 1, 2020 at 9:00 AM Tom Bentley <tbent...@redhat.com> > >>> wrote: > >>> > > > >>> > > > Hi Sönke, > >>> > > > > >>> > > > Thanks for taking a look. > >>> > > > > >>> > > > Let me answer in reverse order, since I think it might make more > >>> sense > >>> > > that > >>> > > > way... > >>> > > > > >>> > > > Also, while writing that, I noticed that you have a version with > >>> and > >>> > > > > without "name" for your transformation in the KIP: > >>> > > > > > >>> > > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> > > has-header:my-header > >>> > > > > and > >>> > > > > transforms.conditionalExtract.condition: has-header:my-header > >>> > > > > > >>> > > > > >>> > > > The example > >>> > > > transforms.conditionalExtract.condition: has-header:my-header > >>> > > > is a "has-header" condition (the prefix of the config value), > >>> which will > >>> > > > match records with a "my-header" header (given in the suffix). > >>> > > > > >>> > > > The other example given is: > >>> > > > transforms.conditionalExtract.condition: not:hasMyHeader > >>> > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> > > > has-header:my-header > >>> > > > The root of the condition is a "not" condition (the prefix of the > >>> value > >>> > > for > >>> > > > the transforms.conditionalExtract.condition key) of another named > >>> > > condition > >>> > > > called "hasMyHeader" (the suffix). Any name could be used for the > >>> other > >>> > > > condition. That other condition is configured at > >>> > > > "transforms.conditionalExtract.condition.<conditionName>". That > >>> condition > >>> > > > is a "has-header" condition (the prefix), which will match > records > >>> with a > >>> > > > "my-header" header (given in the suffix). So the "has-header:" > >>> condition > >>> > > > type would always require a suffix, as would the "not:" condition > >>> type. > >>> > > > Hypothetically you could have a "true" condition type (which > would > >>> not > >>> > > > require a suffix), and the hypothetical binary conditions "and:" > >>> and > >>> > > "or:" > >>> > > > would require a pair of other condition names. > >>> > > > > >>> > > > So what's proposed is a scheme for encoding conditions where the > >>> > > condition > >>> > > > type is the prefix of the value of some "....condition" config > >>> key, and > >>> > > the > >>> > > > optional suffix provides parameters for the condition. This makes > >>> those > >>> > > > parameters a bit inflexible, but is relatively succinct. > >>> > > > > >>> > > > This leads on to your first point. You're right that use cases > >>> might > >>> > > appear > >>> > > > which need other conditions, and we should make it flexible > enough > >>> to be > >>> > > > able to cope with future use cases. On the other hand, I was > >>> concerned > >>> > > that > >>> > > > we end up with something which is quite complicated to configure. > >>> (There > >>> > > > comes a point where it might makes more sense for the user to > >>> write their > >>> > > > own SMT). > >>> > > > > >>> > > > Just of the top of my head it might look like: > >>> > > > > > >>> > > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> type:has-header > >>> > > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> > > > header-name:my-header > >>> > > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> > > field-value:my-value > >>> > > > > > >>> > > > > >>> > > > That won't work because the format is basically a > >>> Properties/Map<String, > >>> > > > String> and what you've suggested has duplicate keys. > >>> > > > > >>> > > > One thing I did briefly consider what the ability to treat > >>> conditions as > >>> > > > Configurable objects in their own right (opening up the > >>> possibility of > >>> > > > people supplying their own Conditions, just like they can supply > >>> their > >>> > > own > >>> > > > SMTs). That might be configured something like this: > >>> > > > > >>> > > > transforms.conditionalExtract.condition: not > >>> > > > transforms.conditionalExtract.condition.not.type: Not > >>> > > > transforms.conditionalExtract.condition.not.negated: foo > >>> > > > transforms.conditionalExtract.condition.foo.type: > >>> HasHeaderWithValue > >>> > > > transforms.conditionalExtract.condition.foo.header: my-header > >>> > > > transforms.conditionalExtract.condition.foo.value: my-value > >>> > > > > >>> > > > I didn't propose that I couldn't see the use cases to justify > this > >>> kind > >>> > > of > >>> > > > complexity, especially as the common case would surely be > matching > >>> > > against > >>> > > > topic name (to be honest I wasn't completely convinced by the > need > >>> for > >>> > > > "has-header"). In the current version of the KIP that's just > >>> > > > > >>> > > > transforms.conditionalExtract.condition: topic-matches: > >>> my-prefix-.* > >>> > > > > >>> > > > but using the more flexible scheme that would require something > >>> more like > >>> > > > this: > >>> > > > > >>> > > > transforms.conditionalExtract.condition: bar > >>> > > > transforms.conditionalExtract.condition.bar.type: TopicMatch > >>> > > > transforms.conditionalExtract.condition.bar.pattern: > >>> my-prefix-.* > >>> > > > > >>> > > > If people know of use cases which would justify more > >>> sophistication, I'm > >>> > > > happy to reconsider. > >>> > > > > >>> > > > Thanks again for taking a look! > >>> > > > > >>> > > > Tom > >>> > > > > >>> > > > On Wed, Apr 1, 2020 at 2:05 PM Sönke Liebau > >>> > > > <soenke.lie...@opencore.com.invalid> wrote: > >>> > > > > >>> > > > > Hi Tom, > >>> > > > > > >>> > > > > sounds useful to me, thanks for the KIP! > >>> > > > > The only thought that I had while reading was that this will > >>> probably > >>> > > > raise > >>> > > > > questions about more involved conditions fairly quickly. For > >>> example > >>> > > the > >>> > > > > "has-header" will cause an appetite for conditions like > >>> > > > > "this-header-has-that-value". > >>> > > > > This would necessitate two parameters to be passed into the > >>> condition, > >>> > > > > which I think is not currently included in the KIP. I am not > >>> saying add > >>> > > > > this now, but might it make sense to discuss a concept of how > >>> that > >>> > > might > >>> > > > > look now, to avoid potential changes later on. > >>> > > > > > >>> > > > > Just of the top of my head it might look like: > >>> > > > > > >>> > > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> type:has-header > >>> > > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> > > > header-name:my-header > >>> > > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> > > field-value:my-value > >>> > > > > > >>> > > > > > >>> > > > > Also, while writing that, I noticed that you have a version > with > >>> and > >>> > > > > without "name" for your transformation in the KIP: > >>> > > > > > >>> > > > > transforms.conditionalExtract.condition.hasMyHeader: > >>> > > has-header:my-header > >>> > > > > and > >>> > > > > transforms.conditionalExtract.condition: has-header:my-header > >>> > > > > > >>> > > > > > >>> > > > > Is this intentional and the name is optional? > >>> > > > > > >>> > > > > > >>> > > > > Best regards, > >>> > > > > > >>> > > > > Sönke > >>> > > > > > >>> > > > > > >>> > > > > > >>> > > > > On Wed, 1 Apr 2020 at 11:12, Tom Bentley <tbent...@redhat.com> > >>> wrote: > >>> > > > > > > >>> > > > > > Does anyone have any comments, feedback or thoughts about > this? > >>> > > > > > > >>> > > > > > Thanks, > >>> > > > > > > >>> > > > > > Tom > >>> > > > > > > >>> > > > > > On Tue, Mar 24, 2020 at 11:56 AM Tom Bentley < > >>> tbent...@redhat.com> > >>> > > > > wrote: > >>> > > > > > > >>> > > > > > > Hi, > >>> > > > > > > > >>> > > > > > > I've opened KIP-585 which is intended to provide a > mechanism > >>> to > >>> > > > > > > conditionally apply SMTs in Kafka Connect. I'd be grateful > >>> for any > >>> > > > > > > feedback: > >>> > > > > > > > >>> > > > > > >>> > > > > > >>> > > > > >>> > > > >>> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-585%3A+Conditional+SMT > >>> > > > > > > > >>> > > > > > > Many thanks, > >>> > > > > > > > >>> > > > > > > Tom > >>> > > > > > > > >>> > > > > > >>> > > > > >>> > > > >>> > >>> >
