> > when you use a HashMap or RocksDB or other "state stores", you don't > expect them to automatically know extra stuff about the record you're > storing.
So, I don't think there is any reason we *can't* retain the record context > in the StateStoreContext, and if any users came along with a clear use case > I'd find that convincing. > I agree with the principle of being conservative with the StateStoreContext API. Regarding user expectations or a clear use case, the only counterpoint I would offer is that we sort of have that use case already, which is the example I gave of the change logging store using the timestamp. I am curious if this functionality will be retained when using built in state stores, or will a low-level processor get a KeyValueStore that no longer writes to the changelog topic with the record's timestamp. While I personally don't care much about that functionality specifically, I have a general desire for custom state stores to easily do the things that built in state stores do. It genuinely did not occur to me that users might be looking up and/or > updating records of other keys from within a Processor. > I'm glad you said this Sophie, because it gives me an opportunity to say that this is actually a *huge* use case for my team. The state store usability improvements I was referring to in my previous message were about enabling the user to write custom stores while still easily hooking into the ability to write to a changelog topic. I think that is technically possible now, but I don't think it's trivial. Specifically, we have an interest in making a state store with a richer way of querying its data (like perhaps getting all values associated with a secondary key), while still ultimately writing to the changelog topic for later restoration. We recognize that this use case throws away some of what kafka streams (especially the DSL) is good at - easy parallelizability by partitioning all processing by key - and that our business logic would completely fall apart if we were consuming from multi-partition topics with multiple consumers. But we have found that using the low level processor API is good for the very simple stream processing primitives it provides: handling the plumbing of consuming from multiple kafka topics and potentially updating persistent local state in a reliable way. That in itself has proven to be a worthwhile programming model. Since I got off track a bit, let me summarize: I don't particularly care about the record context being available to state store implementations, and I think this KIP is headed in the right direction in that regard. But more generally, I wanted to express the importance of maintaining a powerful and flexible StateStore interface. Thanks! Paul On Thu, Sep 10, 2020 at 6:11 PM Sophie Blee-Goldman <sop...@confluent.io> wrote: > Aha, I did misinterpret the example in your previous response regarding the > range query after all. I thought you just meant a time-range query inside a > punctuator. It genuinely did not occur to me that users might be looking up > and/or updating records of other keys from within a Processor. Sorry for > being closed minded > > I won't drag out this discussion any further by asking whether that might > be > a valid use case or just a lurking bug in itself :) > > Thanks for humoring me. The current proposal for KIP-478 sounds good to me > > On Thu, Sep 10, 2020 at 3:43 PM John Roesler <vvcep...@apache.org> wrote: > > > Ah, thanks Sophie, > > > > I'm sorry for misinterpreting your resonse. Yes, we > > absolutely can and should clear the context before > > punctuating. > > > > My secondary concern is maybe more far-fetched. I was > > thinking that inside process(key,value), a Processor might > > do a get/put of a _different_ key. Consider, for example, > > the way that Suppress processors work. When they get a > > record, they add it to the store and then do a range scan > > and possibly forward a _different_ record. Of course, this > > is an operation that is deeply coupled to the internals, and > > the Suppress processor accordingly actually does get access > > to the internal context so that it can set the context > > before forwarding. > > > > Still, it seems like I've had a handful of conversations > > with people over the years in which they tell me they are > > using state stores in a way that transcends the "get and put > > the currently processing record" access pattern. I doubt > > that those folks would even have considered the possiblity > > that the currently processing record's _context_ could > > pollute their state store operations, as I myself never gave > > it a second thought until the current conversation began. In > > cases like that, we have actually set a trap for these > > people, and it seems better to dismantle the trap. > > > > As you noted, really the only people who would be negatively > > impacted are people who implement their own state stores. > > These folks will get the deprecation warning and try to > > adapt their stores to the new interface. If they needed > > access to the record context, they would find it's now > > missing. They'd ask us about it, and we'd have the ability > > to explain the lurking bug that they have had in their > > stores all along, as well as the new recommended pattern > > (just pass everything you need in the value). If that's > > unsatisfying, _then_ we should consider amending the API. > > > > Thanks, > > -John > > > > On Thu, 2020-09-10 at 15:21 -0700, Sophie Blee-Goldman > > wrote: > > > > Regarding your first sentence, "...the processor would null > > > > out the record context...", this is not possible, since the > > > > processor doesn't have write access to the context. We could > > > > add it, > > > > > > > > > > Sorry, this was poorly phrased, I definitely did not mean to imply that > > we > > > should make the context modifiable by the Processors themselves. I > meant > > > this should be handled by the internal processing framework that deals > > with > > > passing records from one Processor to the next, setting the record > > context > > > when a new record is picked up, invoking the punctuators, etc. I > believe > > > this > > > all currently happens in the StreamTask? It already can and does > > overwrite > > > the record context as new records are processed, and is also > responsible > > > for calling the punctuators, so it doesn't seem like a huge leap to > just > > say > > > "null out the current record before punctuating" > > > > > > To clarify, I was never advocating or even considering to give the > > > Processors > > > write access to the record context. Sorry if my last message (or all of > > > them) > > > was misleading. I just wanted to point out that the punctuator concern > is > > > orthogonal to the question of whether we should include the record > > context > > > in the StateStoreContext. It's definitely a real problem, but it's a > > > problem > > > that exists at the Processor level and not just the StateStore. > > > > > > So, I don't think there is any reason we *can't* retain the record > > context > > > in the > > > StateStoreContext, and if any users came along with a clear use case > I'd > > > find > > > that convincing. In the absence of any examples, the conservative > > approach > > > sounds good to me. > > > > > > If it turns out that someone did need the record context in their > custom > > > state > > > store, I'm sure they'll submit a politely worded bug report alerting us > > > that we > > > broke their application. > > > > > > On Thu, Sep 10, 2020 at 3:05 PM John Roesler <vvcep...@apache.org> > > wrote: > > > > > > > Thanks, Sophie, > > > > > > > > Yes, now that you point it out, I can see that the record > > > > context itself should be nulled out by Streams before > > > > invoking punctuators. From that perspective, we don't need > > > > to think about the second-order problem of what's in the > > > > context for the state store when called from a punctuator. > > > > > > > > Regarding your first sentence, "...the processor would null > > > > out the record context...", this is not possible, since the > > > > processor doesn't have write access to the context. We could > > > > add it, but then all kinds of strange effects would ensue > > > > when downstream processors execute but the context is empty, > > > > etc. Better to just let the framework manage the record > > > > context and keep it read-only for Processors. > > > > > > > > Reading between the lines of your last reply, it sounds that > > > > the disconnect may just have been a mutual misunderstanding > > > > about whether or not Processors currently have access to set > > > > the record context. Since they do not, if we wanted to add > > > > the record context to StateStoreContext in a well-defined > > > > way, we'd also have to add the ability for Processors to > > > > manipulate it. But then, we're just creating a side-channel > > > > for Processors to pass some information in arguments to > > > > "put()" and other information implicitly through the > > > > context. It seems better just to go for a single channel for > > > > now. > > > > > > > > It sounds like you're basically in favor of the conservative > > > > approach, and you just wanted to understand the blockers > > > > that I implied. Does my clarification make sense? > > > > > > > > Thanks, > > > > -John > > > > > > > > On Thu, 2020-09-10 at 10:54 -0700, Sophie Blee-Goldman > > > > wrote: > > > > > I was just thinking that the processor would null out the record > > context > > > > > after it > > > > > finished processing the record, so I'm not sure I follow why this > > would > > > > not > > > > > be > > > > > possible? AFAIK we never call a punctuator in the middle of > > processing a > > > > > record through the topology, and even if we did, we still know when > > it is > > > > > about > > > > > to be called and could set it to null beforehand. > > > > > > > > > > I'm not trying to advocate for it here, I'm in agreement that > > anything > > > > you > > > > > want > > > > > to access within the store can and should be accessed within the > > calling > > > > > Processor/Punctuator before reaching the store. The "we can always > > add it > > > > > later if necessary" argument is also pretty convincing. Just trying > > to > > > > > understand > > > > > why this wouldn't be possible. > > > > > > > > > > FWIW, the question of "what is the current record in the context > of a > > > > > Punctuator" > > > > > exists independently of whether we want to add this to the > > > > StateStoreContext > > > > > or not. The full ProcessorContext, including the current record > > context, > > > > is > > > > > already available within a Punctuator, so removing the current > record > > > > > context > > > > > from the StateStoreContext does not solve the problem. Users can -- > > and > > > > have > > > > > (see KAFKA-9584 <https://issues.apache.org/jira/browse/KAFKA-9584 > >;;) > > -- > > > > hit > > > > > such subtle bugs without ever invoking a StateStore > > > > > from their punctuator. > > > > > > > > > > Again, I think I do agree that we should leave the current record > > context > > > > > off of > > > > > the StateStoreContext, but I don't think the Punctuator argument > > against > > > > it > > > > > is > > > > > very convincing. It sounds to me like we need to disallow access to > > the > > > > > current > > > > > record context from within the Punctuator, independent of anything > > to do > > > > > with > > > > > state stores > > > > > > > > > > On Thu, Sep 10, 2020 at 7:12 AM John Roesler <vvcep...@apache.org> > > > > wrote: > > > > > > Thanks for the thoughts, Sophie. > > > > > > > > > > > > I agree that the extra information could be useful. My only > > concern is > > > > > > that it doesn’t seem like we can actually supply that extra > > information > > > > > > correctly. So, then we have a situation where the system offers > > useful > > > > API > > > > > > calls that are only correct in a narrow range of use cases. > > Outside of > > > > > > those use cases, you get incorrect behavior. > > > > > > > > > > > > If it were possible to null out the context before you put a > > document > > > > to > > > > > > which the context doesn’t apply, then the concern would be > > mitigated. > > > > But > > > > > > it would still be pretty weird from the perspective of the store > > that > > > > > > sometimes the context is populated and other times, it’s null. > > > > > > > > > > > > But that seems moot, since it doesn’t seem possible to null out > the > > > > > > context. Only the Processor could know whether it’s about to put > a > > > > document > > > > > > different from the context or not. And it would be inappropriate > to > > > > offer a > > > > > > public ProcessorContext api to manage the record context. > > > > > > > > > > > > Ultimately, it still seems like if you want to store headers, you > > can > > > > > > store them explicitly, right? That doesn’t seem onerous to me, > and > > it > > > > kind > > > > > > of seems better than relying on undefined or asymmetrical > behavior > > in > > > > the > > > > > > store itself. > > > > > > > > > > > > Anyway, I’m not saying that we couldn’t solve these problems. > Just > > > > that it > > > > > > seems a little that we can be conservative and avoid them for > now. > > If > > > > it > > > > > > turns out we really need to solve them, we can always do it > later. > > > > > > > > > > > > Thanks, > > > > > > John > > > > > > > > > > > > On Wed, Sep 9, 2020, at 22:46, Sophie Blee-Goldman wrote: > > > > > > > > If you were to call "put" from a punctuator, or do a > > > > > > > > `range()` query and then update one of those records with > > > > > > > > `put()`, you'd have a very subtle bug on your hands. > > > > > > > > > > > > > > Can you elaborate on this a bit? I agree that the punctuator > > case is > > > > an > > > > > > > obvious exemption to the assumption that store invocations > always > > > > > > > have a corresponding "current record", but I don't understand > the > > > > > > > second example. Are you envisioning a scenario where the > #process > > > > > > > method performs a range query and then updates records? Or were > > > > > > > you just giving another example of the punctuator case? > > > > > > > > > > > > > > I only bring it up because I agree that the current record > > > > information > > > > > > could > > > > > > > still be useful within the context of the store. As a non-user > my > > > > input > > > > > > on > > > > > > > this > > > > > > > definitely has limited value, but it just isn't striking me as > > > > obvious > > > > > > that > > > > > > > we > > > > > > > should remove access to the current record context from the > state > > > > stores. > > > > > > > If there is no current record, as in the punctuator case, we > > should > > > > just > > > > > > > set > > > > > > > the record context to null (or Optional.empty, etc). > > > > > > > > > > > > > > That said, the put() always has to come from somewhere, and > that > > > > > > > somewhere is always going to be either a Processor or a > > Punctuator, > > > > both > > > > > > > of which will still have access to the full context. So > > additional > > > > info > > > > > > > such as > > > > > > > the timestamp can and should probably be supplied to the store > > before > > > > > > > calling put(), rather than looked up by the store. But I can > see > > some > > > > > > other > > > > > > > things being useful, for example the current record's headers. > > Maybe > > > > > > if/when > > > > > > > we add better (or any) support for headers in state stores this > > will > > > > be > > > > > > > less true. > > > > > > > > > > > > > > Of course as John has made clear, it's pretty hard to judge > > without > > > > > > > examples > > > > > > > and more insight as to what actually goes on within a custom > > state > > > > store > > > > > > > On Wed, Sep 9, 2020 at 8:07 PM John Roesler < > vvcep...@apache.org > > > > > > > wrote: > > > > > > > > Hi Paul, > > > > > > > > > > > > > > > > It's good to hear from you! > > > > > > > > > > > > > > > > I'm glad you're in favor of the direction. Especially when > > > > > > > > it comes to public API and usability concens, I tend to > > > > > > > > think that "the folks who matter" are actually the folks who > > > > > > > > have to use the APIs to accomplish real tasks. It can be > > > > > > > > hard for me to be sure I'm thinking clearly from that > > > > > > > > perspective. > > > > > > > > > > > > > > > > Funny story, I also started down this road a couple of times > > > > > > > > already and backed them out before the KIP because I was > > > > > > > > afraid of the scope of the proposal. Unfortunately, needing > > > > > > > > to make a new ProcessorContext kind of forced my hand. > > > > > > > > > > > > > > > > I see you've called me out about the ChangeLogging stores :) > > > > > > > > In fact, I think these are the main/only reason that stores > > > > > > > > might really need to invoke "forward()". My secret plan was > > > > > > > > to cheat and either accomplish change-logging by a different > > > > > > > > mechanism than implementing the store interface, or by just > > > > > > > > breaking encapsulation to sneak the "real" ProcessorContext > > > > > > > > into the ChangeLogging stores. But those are all > > > > > > > > implementation details. I think the key question is whether > > > > > > > > anyone else has a store implementation that needs to call > > > > > > > > "forward()". It's not what you mentioned, but since you > > > > > > > > spoke up, I'll just ask: if you have a use case for calling > > > > > > > > "forward()" in a store, please share it. > > > > > > > > > > > > > > > > Regarding the other record-specific context methods, I think > > > > > > > > you have a good point, but I also can't quite wrap my head > > > > > > > > around how we can actually guarantee it to work in general. > > > > > > > > For example, the case you cited, where the implementation of > > > > > > > > `KeyValueStore#put(key, value)` uses the context to augment > > > > > > > > the record with timestamp information. This relies on the > > > > > > > > assumption that you would only call "put()" from inside a > > > > > > > > `Processor#process(key, value)` call in which the record > > > > > > > > being processed is the same record that you're trying to put > > > > > > > > into the store. > > > > > > > > > > > > > > > > If you were to call "put" from a punctuator, or do a > > > > > > > > `range()` query and then update one of those records with > > > > > > > > `put()`, you'd have a very subtle bug on your hands. Right > > > > > > > > now, the Streams component that actually calls the Processor > > > > > > > > takes care to set the right record context before invoking > > > > > > > > the method, and in the case of caching, etc., it also takes > > > > > > > > care to swap out the old context and keep it somewhere safe. > > > > > > > > But when it comes to public API Processors calling methods > > > > > > > > on StateStores, there's no opportunity for any component to > > > > > > > > make sure the context is always correct. > > > > > > > > > > > > > > > > In the face of that situation, it seemed better to just move > > > > > > > > in the direction of a "normal" data store. I.e., when you > > > > > > > > use a HashMap or RocksDB or other "state stores", you don't > > > > > > > > expect them to automatically know extra stuff about the > > > > > > > > record you're storing. If you need them to know something, > > > > > > > > you just put it in the value. > > > > > > > > > > > > > > > > All of that said, I'm just reasoning from first principles > > > > > > > > here. To really know if this is a mistake or not, I need to > > > > > > > > be in your place. So please push back if you think what I > > > > > > > > said is nonsense. My personal plan was to keep an eye out > > > > > > > > during the period where the old API was still present, but > > > > > > > > deprecated, to see if people were struggling to use the new > > > > > > > > API. If so, then we'd have a chance to address it before > > > > > > > > dropping the old API. But it's even better if you can help > > > > > > > > think it through now. > > > > > > > > > > > > > > > > It did also cross my mind to _not_ add the > > > > > > > > StateStoreContext, but just to continue to punt on the > > > > > > > > question by just dropping in the new ProcessorContext to the > > > > > > > > new init method. If StateStoreContext seems too bold, we can > > > > > > > > go that direction. But if we actually add some methods to > > > > > > > > StateStoreContext, I'd like to be able to ensure they would > > > > > > > > be well defined. I think the current situation was more of > > > > > > > > an oversight than a choice. > > > > > > > > > > > > > > > > Thanks again for your reply, > > > > > > > > -John > > > > > > > > > > > > > > > > > > > > > > > > On Wed, 2020-09-09 at 21:23 -0500, Paul Whalen wrote: > > > > > > > > > John, > > > > > > > > > > > > > > > > > > It's exciting to see this KIP head in this direction! In > the > > > > last > > > > > > year > > > > > > > > or > > > > > > > > > so I've tried to sketch out some usability improvements for > > > > custom > > > > > > state > > > > > > > > > stores, and I also ended up splitting out the > > StateStoreContext > > > > from > > > > > > the > > > > > > > > > ProcessorContext in an attempt to facilitate what I was > > doing. I > > > > > > sort of > > > > > > > > > abandoned it when I realized how large the ideal change > might > > > > have > > > > > > to be, > > > > > > > > > but it's great to see that there is other interest in > moving > > in > > > > this > > > > > > > > > direction (from the folks that matter :) ). > > > > > > > > > > > > > > > > > > Having taken a stab at it myself, I have a comment/question > > on > > > > this > > > > > > > > bullet > > > > > > > > > about StateStoreContext: > > > > > > > > > > > > > > > > > > It does *not* include anything processor- or record- > > specific, > > > > like > > > > > > > > > > `forward()` or any information about the "current" > record, > > > > which is > > > > > > > > only a > > > > > > > > > > well-defined in the context of the Processor. Processors > > > > process > > > > > > one > > > > > > > > record > > > > > > > > > > at a time, but state stores may be used to store and > fetch > > many > > > > > > > > records, so > > > > > > > > > > there is no "current record". > > > > > > > > > > > > > > > > > > > > > > > > > > > > I totally agree that record-specific or processor-specific > > > > context > > > > > > in a > > > > > > > > > state store is often not well-defined and it would be good > to > > > > > > separate > > > > > > > > that > > > > > > > > > out, but sometimes it (at least record-specific context) is > > > > actually > > > > > > > > > useful, for example, passing the record's timestamp through > > to > > > > the > > > > > > > > > underlying storage (or changelog topic): > > > > > > > > > > > > > > > > https://github.com/apache/kafka/blob/trunk/streams/src/main/java/org/apache/kafka/streams/state/internals/ChangeLoggingKeyValueBytesStore.java#L121 > > > > > > > > > You could have the writer client of the state store pass > this > > > > > > through, > > > > > > > > but > > > > > > > > > it would be nice to be able to write state stores where the > > > > client > > > > > > did > > > > > > > > not > > > > > > > > > have this responsibility. I'm not sure if the solution is > > to add > > > > > > some > > > > > > > > > things back to StateStoreContext, or make yet another > context > > > > that > > > > > > > > > represents record-specific context while inside a state > > store. > > > > > > > > > > > > > > > > > > Best, > > > > > > > > > Paul > > > > > > > > > > > > > > > > > > On Wed, Sep 9, 2020 at 5:43 PM John Roesler < > > j...@vvcephei.org> > > > > > > wrote: > > > > > > > > > > Hello all, > > > > > > > > > > > > > > > > > > > > I've been slowly pushing KIP-478 forward over the last > > year, > > > > > > > > > > and I'm happy to say that we're making good progress now. > > > > > > > > > > However, several issues with the original design have > come > > > > > > > > > > to light. > > > > > > > > > > > > > > > > > > > > The major changes: > > > > > > > > > > > > > > > > > > > > We discovered that the original plan of just adding > generic > > > > > > > > > > parameters to ProcessorContext was too disruptive, so we > > are > > > > > > > > > > now adding a new api.ProcessorContext. > > > > > > > > > > > > > > > > > > > > That choice forces us to add a new StateStore.init method > > > > > > > > > > for the new context, but ProcessorContext really isn't > > ideal > > > > > > > > > > for state stores to begin with, so I'm proposing a new > > > > > > > > > > StateStoreContext for this purpose. In a nutshell, there > > are > > > > > > > > > > quite a few methods in ProcessorContext that actually > > should > > > > > > > > > > never be called from inside a StateStore. > > > > > > > > > > > > > > > > > > > > Also, since there is a new ProcessorContext interface, we > > > > > > > > > > need a new MockProcessorContext implementation in the > test- > > > > > > > > > > utils module. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > The changeset for the KIP document is here: > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/pages/diffpagesbyversion.action?pageId=118172121&selectedPageVersions=14&selectedPageVersions=10 > > > > > > > > > > And the KIP itself is here: > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-478+-+Strongly+typed+Processor+API > > > > > > > > > > If you have any concerns, please let me know! > > > > > > > > > > > > > > > > > > > > Thanks, > > > > > > > > > > -John > > > > > > > > > > > > > > > > > > > > > > > > >
