Just want to make a quick comment on the question that John raised about whether we should introduce a separate config for "key" and "value" window sizes:
My short answer is No, I don't think that's necessary. First of all, as you said, there is no first-class concept of a "Windowed value" in the DSL. Second, to engage in your rhetorical question, if there's no default window size for a Streams program then how can there be a sensible default for the key AND a separate sensible default for a value? I don't think we need to follow the existing pattern if it doesn't make sense, and to be honest I'm a bit skeptical that anyone was even using these default windowed inner classes since the config wasn't even defined/documented until quite recently. I'd actually be in favor of deprecating StreamsConfig.DEFAULT_WINDOWED_VALUE_SERDE_INNER_CLASS but I don't want to drag that into this discussion as well. My understanding is that these were meant to mirror the default key/value serde configs, but the real use of the DEFAULT_WINDOWED_SERDE_INNER_CLASS config is actually that you can at least use it to configure the inner class for a Consumer, thus making the TimeWindowed serdes functional at a basic level. With the window size configs, the point is not really to set a default but to make it actually work with a Consumer which instantiates the deserializer by reflection. So I don't think we should position this new config as a "default" (although it may technically behave as one) -- within Streams users can and should always supply the window size through the constructor. I don't think that's such an inconvenience, vs the amount of confusion that will (and has) been caused by default serde-related configs in streams. Regarding the fixed vs variable sized config, one idea I had was to just keep the fixed-size config and constructor and let users of enumerable windows override the TimeWindowedSerde class(es) to do whatever it is they need. IIUC you already have to override some other windows-related classes to get variable-sized windows so doing the same for the serdes sounds reasonable to me. Just my take on the "simple things should be easy, difficult things should be possible" mantra One last quick side note: the reason we don't really need to discuss SessionWindows here is that they already encode both the start and end time for the window. This is probably the best way to go for TimeWindows as well, but making this change in a backwards compatible way is a much larger scope of work. And even then, we might want to consider making it possible to still just encode the start time to save space, thus requiring this config either way On Fri, Aug 21, 2020 at 9:26 AM Leah Thomas <ltho...@confluent.io> wrote: > Thanks John and Walker for your thoughts. > > I agree with your two scenarios John, that you configure fully in the > constructor, or you don't need to call `init()`. IIUC, if we pass the > deserializer to the consumer, we want to make sure it has the window size > is set using the newly required constructor. If we don't pass in the > deserializer, the window size will be set through the configs. To answer > Walker's question directly, because the configs aren't passed to the > constructor, we can't set the window size unless we pass it to the > constructor or configure the constructor after initializing it. > > For users who would rather not set a strict window size (outside of the > variable size scenario), they can pass in Long.MAX_VALUE. The way I see > this is instead of having the default be for scenarios that don't require a > window size, we have the default be the scenarios that *do*, flipping the > current implementation to fit with typical use cases. > > On your points John: > 1. I agree that it makes sense to store it in StreamsConfig, this shouldn't > cause any issues. I've updated the KIP accordingly. > > 2. The non-fixed time windows issue is a good point. It seems like calendar > windows in particular are quite useful, so I think we want to make sure > that this wouldn't inhibit flexible sized windows. I think having two > different configs and functions makes sense, although it is slightly > messier. While requiring all time windows to use the WindowFunction > constructor would work, I think that allowing users to access the > WindowSize constructor is preferable because it seems easier to use for > people who are not at all interested in delving into variably sized > windows. This assumption could be wrong though, and perhaps users would > adapt quickly to the new WindowFunction style, but my immediate reaction is > to support both configs and constructors. > > One note on this is that Session Windows are handled separately from time > windows and also have variable window sizes. I assume that the TimeWindowed > option is preferable for variably sized windows because you still want to > access the window end times? But I think one alternative could be > separating the variably sized windows from the current implementation of > time windows, although I think KIP-645 > < > https://cwiki.apache.org/confluence/display/KAFKA/KIP-645%3A+Replace+Windows+with+a+proper+interface > > > would make this not strictly necessary. > > Cheers, > Leah > > On Fri, Aug 21, 2020 at 10:04 AM John Roesler <vvcep...@apache.org> wrote: > > > Hi Leah, > > > > Thanks for the KIP! This has been a real pain for some use > > cases, so it's really good to see a proposal to fix it. > > > > We do need a default constructor so that it can be > > dynamically instantiated by the consumer (or any other > > component). But I'm +1 on deprecating the constructor you're > > proposing to deprecate, which only partially configures the > > class. It seems like there are exactly two patterns: either > > you fully configure the class in the constructor and don't > > call `init()`, or you call the default constructor and then > > configure the class by calling `init()`. > > > > I can appreciate Walker's point, but stepping back, it > > doesn't actually seem that useful to partially configure the > > class in the constructor and then finish up the > > configuration by calling `init()`. I could see the argument > > if there were a sensible default, but for this particular > > class, there isn't one. Rhetorical question: what is the > > default window size for Streams programs? > > > > I have a couple of concerns to discuss: > > > > 1. Config Location > > > > I don't think I would add the new configs to ConsumerConfig, > > but would add it to StreamsConfig instead. The deserailzier > > itself is in Streams (it is > > o.a.k.streams.kstream.TimeWindowedDeserializer), so it seems > > odd to have one of its configurations in a completely > > different module. > > > > Also, this class already has two configs, which are in > > StreamsConfig: > > StreamsConfig.DEFAULT_WINDOWED_KEY_SERDE_INNER_CLASS > > StreamsConfig.DEFAULT_WINDOWED_VALUE_SERDE_INNER_CLASS > > > > It seems like the new config belongs right next to the > > existing ones. > > > > For me, it raises a secondary question: > > 1b: Should there be a KEY_WINDOW_SIZE and a > > VALUE_WINDOW_SIZE? I'm honestly not sure what a "windowed > > value" even is, but the fact that we can configure serdes > > for it implies that perhaps we should symmetrically > > configure its size as well. > > > > 2. Fixed Size Assumption > > > > In KIP-645, I'm proposing to lift the assumption that > > TimeWindows have a fixed size at all, but KIP-659 is > > currently built on that assumption. > > > > For details on why this is not a good assumtion, see: > > https://issues.apache.org/jira/browse/KAFKA-10408 > > > > In fact, in my POC PR for KIP-659, I'm dropping the > > constructor that takes a "window size" parameter in favor of > > one that takes a window function, mapping a window start > > time to a full Window(start, end). > > > > In that context, it seems incongruous to introduce a > > configuration that specifies a window size. Of course, my > > KIP is also under discussion, so my proposal may not > > eventually be accepted. But it is necessary to consider both > > of these concerns together. > > > > One option seems to be to accept both. Namely, we keep the > > "fixed size" constructor AND add my new constructor (for > > variably sized windows). Likewise, we accept your proposal, > > and KIP-659 would propose to add a new config specifying a > > windowing function, such as: > > > > > StreamsConfig.WINDOW_FUNCTION_CONFIG > > > > which would be an instance of: > > > > > public interface WindowFunction implements Function<Long, > > Window>; > > > > I'm not bringing these up for discussion in your KIP right > > now, just demonstrating the feasibility of merging both > > proposals. > > > > My question for you: do you think the general strategy of > > having two constructors and two configs, one for fixed and > > one for variable windows, makes sense? Is it too > > complicated? Do you have a better idea? > > > > Thanks! > > -John > > > > On Thu, 2020-08-20 at 14:49 -0700, Walker Carlson wrote: > > > Hi Leah, > > > > > > Could you explain a bit more why we do not wish to > > > let TimeWindowedDeserializer and WindowedSerdes be created without a > > > specified time as a parameter? > > > > > > I understand the long.MAX_VALUE could cause problems but would it not > be > > a > > > good idea to have a usable default or fetch from the config if > available? > > > After all you are proposing to add "window.size.ms" > > > > > > We definitely need a fix to this problem and adding "window.size.ms" > > makes > > > sense to me. > > > > > > Thanks for the KIP, > > > Walker > > > > > > On Thu, Aug 20, 2020 at 2:22 PM Leah Thomas <ltho...@confluent.io> > > wrote: > > > > > > > Hi all, > > > > > > > > I'd like to start a discussion for KIP-659: > > > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-659%3A+Improve+TimeWindowedDeserializer+and+TimeWindowedSerde+to+handle+window+size > > > > > > > > > > > > The goal of the KIP is to ensure that window size is passed to the > > consumer > > > > when needed, which will generally be for testing purposes, and to > avoid > > > > runtime errors when the *TimeWindowedSerde* is created without a > window > > > > size. > > > > > > > > Looking forward to hearing your feedback. > > > > > > > > Cheers, > > > > Leah > > > > > > > > >
