Hi Walker,
Sorry for the delay in responding. Thanks for your response earlier.
I think there might be a subtlety getting overlooked in considering
whether we're talking about streams versus tables in cogroup. As I'm
sure you know, Kafka Streams treats "stream" records as independent,
immutable, and opaque "facts", whereas we treat "table" records as a
sequence of updates to an entity identified by the record key (where
"update" means that each record's value represents the new state after
applying the update). For the most part, this is a clean separation,
but there is one special case where records with a "null" value are
interpreted as a tombstone in the table context (i.e., the record
indicates not that the new value of the entity is "null", but rather
that the entity has been deleted). In contrast, a record with a null
value in the stream context is _just_ a record with a null value; no
special semantics.
The difficulty is that these two semantics clash at the stream/table
boundary. So, operations that convert streams to tables (like
KGroupedStream#aggregate) have to cope with ambiguity about whether to
treat null values opaquely as null values, or as tombstones. I think
I'll make a long story short and just say that this is a very, very
complex issue. As a result (and as a bit of a punt), our
KGroupedStream operations actually just discard null-valued records.
This means that the following are _not_ equivalent programs:
table1 =
stream<Id,Record>("records")
.filter(Record::isOk)
.groupByKey()
.aggregate(() -> new Record(), (key, value, agg) -> value)
table2 =
table<Id,Record>("record")
.filter(Record::isOk)
They look about the same, in that they'll both produce a
KTable<Id,Record> with the value being the latest state. But if a
record is deleted in the upstream data (represented as a "null"
value), that record would also be deleted in table2, but not in
table1. Table1 would just perpetually contain the value immediately
prior to the delete.
This is why it makes me nervous to propose a new kind of _stream_
operation ostensibly in order to solve a problem that presents itself
in the _table_ context.
If the goal is to provide a more efficient and convenient multi-way
KTable join, I think it would be a good idea to consider an extension
to the KTable API, not the KStream API. On the other hand, if this is
not the goal, then the motivation of the KIP shouldn't say that it is.
Instead, the KIP could provide some other motivation specifically for
augmenting the KStream API.
There is a third alternative that comes to mind, if you wish to
resolve the long-standing dilemma around this semantic problem and
specify in the KIP how exactly nulls are handled in this operator. But
(although this seems on the face to be a good option), I think it
might be a briarpatch. Even if we are able to reach a suitable design,
we'd have to contend with the fact that it looks like the
KGroupedStream API, but behaves differently.
What do you think about all this?
Thanks again for the KIP and the discussion!
-John
On Mon, Oct 28, 2019 at 3:32 PM Walker Carlson <wcarl...@confluent.io> wrote:
>
> Hi Gouzhang,
>
> Matthias and I did talk about overloading different a type of aggregate
> methods in the cogroup that would take in the windows and returns a
> windowed KTable. We decided that it would break too much with the current
> pattern that was established in the normal KStream. We can revisit this if
> you have a different opinion on the tradeoff.
>
> Walker
>
> On Mon, Oct 28, 2019 at 12:14 PM Guozhang Wang <wangg...@gmail.com> wrote:
>
> > Hi Walker,
> >
> > On Fri, Oct 25, 2019 at 1:34 PM Walker Carlson <wcarl...@confluent.io>
> > wrote:
> >
> > > Hi Guozhang,
> > >
> > > 1. I am familiar with the cogroup of spark, it is very similar to
> > > their join operator but instead it makes the values iterable. I think
> > that
> > > the use cases are different enough that it makes sense to specify the
> > > aggregator when we do.
> > >
> > > I like the idea of "absorb" and I think it could be useful. Although I do
> > > not think it is as intuitive.
> > >
> > > If we were to go that route we would either use more processors or do
> > > essentially the same thing but would have to store the information
> > > required to cogroup inside that KTable. I think this would violate some
> > > design principles. I would argue that we should consider adding absorb as
> > > well and auto re-write it to use cogroup.
> > >
> >
> > Yeah I think I agree with you about the internal design complexity with
> > "absorb"; I was primarily thinking if we can save ourselves from
> > introducing 3 more public classes with co-group. But it seems that without
> > introducing new classes there's no easy way for us to bound the scope of
> > co-grouping (like how many streams will be co-grouped together).
> >
> > LMK if you have some better ideas: generally speaking the less new public
> > interfaces we are introducing to fulfill a new feature the better, so I'd
> > push us to think twice and carefully before we go down the route.
> >
> >
> > >
> > > 2. We have not considered this thought that would be a convenient
> > > operation.
> > >
> > > 3. There is only one processor made. We are actually having the naming
> > > conversation right now in the above thread
> > >
> > > 4, 5. fair points
> > >
> > > Walker
> > >
> > > On Fri, Oct 25, 2019 at 11:58 AM Guozhang Wang <wangg...@gmail.com>
> > wrote:
> > >
> > > > Hi Walker, thanks for the KIP! I made a pass on the writeup and have
> > some
> > > > comments below:
> > > >
> > > > Meta:
> > > >
> > > > 1. Syntax-wise, I'm wondering if we have compared our current proposal
> > > with
> > > > Spark's co-group syntax (I know they are targeting for different use
> > > cases,
> > > > but wondering if their syntax is closer to the join operator), what are
> > > the
> > > > syntax / semantics trade-off here?
> > > >
> > > > Just playing a devil's advocate here, if the main motivation is to
> > > provide
> > > > a more convienent multi-way join syntax, and in order to only have one
> > > > materialized store we need to specify the final joined format at the
> > > > beginning, then what about the following alternative (with the given
> > > > example in your wiki page):
> > > >
> > > >
> > > > KGroupedStream<K, V1> grouped1 = builder.stream("topic1").groupByKey();
> > > > KGroupedStream<K, V2> grouped2 = builder.stream("topic2").groupByKey();
> > > > KGroupedStream<K, V3> grouped3 = builder.stream("topic3").groupByKey();
> > > >
> > > > KTable<K, CG> aggregated = grouped1.aggregate(initializer,
> > materialized,
> > > > aggregator1);
> > > >
> > > > aggregated.absorb(grouped2, aggregator2); // I'm just using a random
> > > name
> > > > on top of my head here
> > > > .absorb(grouped3, aggregator3);
> > > >
> > > > In this way, we just add a new API to the KTable to "absorb" new
> > streams
> > > as
> > > > aggregated results without needing to introduce new first citizen
> > > classes.
> > > >
> > > > 2. From the DSL optimization, have we considered if we can auto
> > re-write
> > > > the user written old fashioned multi-join into this new DSL operator?
> > > >
> > > > 3. Although it is not needed for the wiki page itself, for internal
> > > > implementation how many processor nodes would we create for the new
> > > > operator, and how we can allow users to name them?
> > > >
> > > > Minor:
> > > >
> > > > 4. In "Public Interfaces", better add the templated generics to
> > > > "KGroupedStream" as "KGroupedStream<K, V>".
> > > >
> > > > 5. Naming wise, I'd suggest we keep the "K" together with Stream/Table,
> > > > e.g. "TimeWindowed*CogroupedKStream*<K, V>".
> > > >
> > > >
> > > > Guozhang
> > > >
> > > >
> > > >
> > > >
> > > > On Thu, Oct 24, 2019 at 11:43 PM Matthias J. Sax <
> > matth...@confluent.io>
> > > > wrote:
> > > >
> > > > > Walker,
> > > > >
> > > > > I am not sure if I can follow your argument. What do you exactly mean
> > > by
> > > > >
> > > > > > I also
> > > > > >> think that in this case it would be better to separate the 2
> > option
> > > > out
> > > > > >> into separate overloads.
> > > > >
> > > > > Maybe you can give an example what method signature you have in mind?
> > > > >
> > > > > >> We could take a named parameter from upstream or add an extra
> > naming
> > > > > option
> > > > > >> however I don't really see the advantage that would give.
> > > > >
> > > > > Are you familiar with KIP-307? Before KIP-307, KS generated all names
> > > > > for all Processors. This makes it hard to reason about a Topology if
> > > > > it's getting complex. Adding `Named` to the new co-group operator
> > would
> > > > > actually align with KIP-307.
> > > > >
> > > > > > It seems to go in
> > > > > >> the opposite direction from the cogroup configuration idea you
> > > > proposed.
> > > > >
> > > > > Can you elaborate? Not sure if I can follow.
> > > > >
> > > > >
> > > > >
> > > > > -Matthias
> > > > >
> > > > >
> > > > > On 10/24/19 10:20 AM, Walker Carlson wrote:
> > > > > > While I like the idea Sophie I don't think that it is necessary. I
> > > also
> > > > > > think that in this case it would be better to separate the 2 option
> > > out
> > > > > > into separate overloads.
> > > > > > We could take a named parameter from upstream or add an extra
> > naming
> > > > > option
> > > > > > however I don't really see the advantage that would give. It seems
> > to
> > > > go
> > > > > in
> > > > > > the opposite direction from the cogroup configuration idea you
> > > > proposed.
> > > > > >
> > > > > > John, I think it could be both. It depends on when you aggregate
> > and
> > > > what
> > > > > > kind of data you have. In the example it is aggregating before
> > > joining,
> > > > > > there are probably some cases where you could join before
> > > aggregating.
> > > > > IMHO
> > > > > > it would be easier to group all the streams together then perform
> > the
> > > > one
> > > > > > operation that results in a single KTable.
> > > > > >
> > > > > >
> > > > > >
> > > > > > On Wed, Oct 23, 2019 at 9:58 PM Sophie Blee-Goldman <
> > > > sop...@confluent.io
> > > > > >
> > > > > > wrote:
> > > > > >
> > > > > >>> I can personally not see any need to add other configuration
> > > > > >> Famous last words?
> > > > > >>
> > > > > >> Just kidding, 95% confidence is more than enough to me (and
> > better
> > > to
> > > > > >> optimize for current
> > > > > >> design than for hypothetical future changes).
> > > > > >>
> > > > > >> One last question I have then is about the
> > > operator/store/repartition
> > > > > >> naming -- seems like
> > > > > >> we can name the underlying store/changelog through the
> > Materialized
> > > > > >> parameter, but do we
> > > > > >> also want to include an overload taking a Named parameter for the
> > > > > operator
> > > > > >> name (as in the
> > > > > >> KTable#join variations)?
> > > > > >>
> > > > > >> On Wed, Oct 23, 2019 at 5:14 PM Matthias J. Sax <
> > > > matth...@confluent.io>
> > > > > >> wrote:
> > > > > >>
> > > > > >>> Interesting idea, Sophie.
> > > > > >>>
> > > > > >>> So far, we tried to reuse existing config objects and only add
> > new
> > > > ones
> > > > > >>> when needed to avoid creating "redundant" classes. This is of
> > > course
> > > > a
> > > > > >>> reactive approach (with the drawback to deprecate stuff if we
> > > change
> > > > > it,
> > > > > >>> as you described).
> > > > > >>>
> > > > > >>> I can personally not see any need to add other configuration
> > > > parameters
> > > > > >>> atm, so it's a 95% obvious "no" IMHO. The final `aggregate()` has
> > > > only
> > > > > a
> > > > > >>> single state store that we need to configure, and reusing
> > > > > `Materialized`
> > > > > >>> seems to be appropriate.
> > > > > >>>
> > > > > >>> Also note, that the `Initializer` is a mandatory parameter and
> > not
> > > a
> > > > > >>> configuration and should be passed directly, and not via a
> > > > > configuration
> > > > > >>> object.
> > > > > >>>
> > > > > >>>
> > > > > >>> -Matthias
> > > > > >>>
> > > > > >>> On 10/23/19 11:37 AM, Sophie Blee-Goldman wrote:
> > > > > >>>> Thanks for the explanation, makes sense to me! As for the API,
> > one
> > > > > >> other
> > > > > >>>> thought I had is might we ever want or need to introduce any
> > other
> > > > > >>> configs
> > > > > >>>> or parameters in the future? Obviously that's difficult to say
> > now
> > > > (or
> > > > > >>>> maybe the
> > > > > >>>> answer seems obviously "no") but we seem to often end up needing
> > > to
> > > > > add
> > > > > >>> new
> > > > > >>>> overloads and/or deprecate old ones as new features or
> > > requirements
> > > > > >> come
> > > > > >>>> into
> > > > > >>>> play.
> > > > > >>>>
> > > > > >>>> What do you (and others?) think about wrapping the config
> > > parameters
> > > > > >> (ie
> > > > > >>>> everything
> > > > > >>>> except the actual grouped streams) in a new config object? For
> > > > > example,
> > > > > >>> the
> > > > > >>>> CogroupedStream#aggregate field could take a single Cogrouped
> > > > object,
> > > > > >>>> which itself would have an initializer and a materialized. If we
> > > > ever
> > > > > >>> need
> > > > > >>>> to add
> > > > > >>>> a new parameter, we can just add it to the Cogrouped class.
> > > > > >>>>
> > > > > >>>> Also, will the backing store be available for IQ if a
> > Materialized
> > > > is
> > > > > >>>> passed in?
> > > > > >>>>
> > > > > >>>> On Wed, Oct 23, 2019 at 10:49 AM Walker Carlson <
> > > > > wcarl...@confluent.io
> > > > > >>>
> > > > > >>>> wrote:
> > > > > >>>>
> > > > > >>>>> Hi Sophie,
> > > > > >>>>>
> > > > > >>>>> Thank you for your comments. As for the different methods
> > > > signatures
> > > > > I
> > > > > >>> have
> > > > > >>>>> not really considered any other options but while I do agree
> > it
> > > is
> > > > > >>>>> confusing, I don't see any obvious solutions. The problem is
> > that
> > > > the
> > > > > >>>>> cogroup essentially pairs a group stream with an aggregator and
> > > > when
> > > > > >> it
> > > > > >>> is
> > > > > >>>>> first made the method is called on a groupedStream already.
> > > However
> > > > > >> each
> > > > > >>>>> subsequent stream-aggregator pair is added on to a cogroup
> > stream
> > > > so
> > > > > >> it
> > > > > >>>>> needs both arguments.
> > > > > >>>>>
> > > > > >>>>> For the second question you should not need a joiner. The idea
> > is
> > > > > that
> > > > > >>> you
> > > > > >>>>> can collect many grouped streams with overlapping key spaces
> > and
> > > > any
> > > > > >>> kind
> > > > > >>>>> of value types. Once aggregated its value will be reduced into
> > > one
> > > > > >> type.
> > > > > >>>>> This is why you need only one initializer. Each aggregator will
> > > > need
> > > > > >> to
> > > > > >>>>> integrate the new value with the new object made in the
> > > > initializer.
> > > > > >>>>> Does that make sense?
> > > > > >>>>>
> > > > > >>>>> This is a good question and I will include this explanation in
> > > the
> > > > > kip
> > > > > >>> as
> > > > > >>>>> well.
> > > > > >>>>>
> > > > > >>>>> Thanks,
> > > > > >>>>> Walker
> > > > > >>>>>
> > > > > >>>>> On Tue, Oct 22, 2019 at 8:59 PM Sophie Blee-Goldman <
> > > > > >>> sop...@confluent.io>
> > > > > >>>>> wrote:
> > > > > >>>>>
> > > > > >>>>>> Hey Walker,
> > > > > >>>>>>
> > > > > >>>>>> Thanks for the KIP! I have just a couple of questions:
> > > > > >>>>>>
> > > > > >>>>>> 1) It seems a little awkward to me that with the current API,
> > we
> > > > > >> have a
> > > > > >>>>>> nearly identical
> > > > > >>>>>> "add stream to cogroup" method, except for the first which
> > has a
> > > > > >>>>> different
> > > > > >>>>>> signature
> > > > > >>>>>> (ie the first stream is joined as stream.cogroup(Aggregator)
> > > while
> > > > > >> the
> > > > > >>>>>> subsequent ones
> > > > > >>>>>> are joined as .cogroup(Stream, Aggregator) ). I'm not sure
> > what
> > > it
> > > > > >>> would
> > > > > >>>>>> look like exactly,
> > > > > >>>>>> but I was just wondering if you'd considered and/or rejected
> > any
> > > > > >>>>>> alternative APIs?
> > > > > >>>>>>
> > > > > >>>>>> 2) This might just be my lack of familiarity with "cogroup"
> > as a
> > > > > >>> concept,
> > > > > >>>>>> but with the
> > > > > >>>>>> current (non-optimal) API the user seems to have some control
> > > over
> > > > > >> how
> > > > > >>>>>> exactly
> > > > > >>>>>> the different streams are joined through the ValueJoiners.
> > Would
> > > > > this
> > > > > >>> new
> > > > > >>>>>> cogroup
> > > > > >>>>>> simply concatenate the values from the different cogroup
> > > streams,
> > > > or
> > > > > >>>>> could
> > > > > >>>>>> users
> > > > > >>>>>> potentially pass some kind of Joiner to the cogroup/aggregate
> > > > > >> methods?
> > > > > >>>>> Or,
> > > > > >>>>>> is the
> > > > > >>>>>> whole point of cogroups that you no longer ever need to
> > specify
> > > a
> > > > > >>> Joiner?
> > > > > >>>>>> If so, you
> > > > > >>>>>> should add a short line to the KIP explaining that for those
> > of
> > > us
> > > > > >> who
> > > > > >>>>>> aren't fluent
> > > > > >>>>>> in cogroup semantics :)
> > > > > >>>>>>
> > > > > >>>>>> Cheers,
> > > > > >>>>>> Sophie
> > > > > >>>>>>
> > > > > >>>>>> On Thu, Oct 17, 2019 at 3:06 PM Walker Carlson <
> > > > > >> wcarl...@confluent.io>
> > > > > >>>>>> wrote:
> > > > > >>>>>>
> > > > > >>>>>>> Good catch I updated that.
> > > > > >>>>>>>
> > > > > >>>>>>> I have made a PR for this KIP
> > > > > >>>>>>>
> > > > > >>>>>>> I then am splitting it into 3 parts, first cogroup for a
> > > > key-value
> > > > > >>>>> store
> > > > > >>>>>> (
> > > > > >>>>>>> here <https://github.com/apache/kafka/pull/7538>), then for
> > a
> > > > > >>>>>>> timeWindowedStore, and then a sessionWindowedStore + ensuring
> > > > > >>>>>> partitioning.
> > > > > >>>>>>>
> > > > > >>>>>>> Walker
> > > > > >>>>>>>
> > > > > >>>>>>> On Tue, Oct 15, 2019 at 12:47 PM Matthias J. Sax <
> > > > > >>>>> matth...@confluent.io>
> > > > > >>>>>>> wrote:
> > > > > >>>>>>>
> > > > > >>>>>>>> Walker,
> > > > > >>>>>>>>
> > > > > >>>>>>>> thanks for picking up the KIP and reworking it for the
> > changed
> > > > > API.
> > > > > >>>>>>>>
> > > > > >>>>>>>> Overall, the updated API suggestions make sense to me. The
> > > seem
> > > > to
> > > > > >>>>>> align
> > > > > >>>>>>>> quite nicely with our current API design.
> > > > > >>>>>>>>
> > > > > >>>>>>>> One nit: In `CogroupedKStream#aggregate(...)` the type
> > > parameter
> > > > > of
> > > > > >>>>>>>> `Materialized` should be `V`, not `VR`?
> > > > > >>>>>>>>
> > > > > >>>>>>>>
> > > > > >>>>>>>> -Matthias
> > > > > >>>>>>>>
> > > > > >>>>>>>>
> > > > > >>>>>>>>
> > > > > >>>>>>>> On 10/14/19 2:57 PM, Walker Carlson wrote:
> > > > > >>>>>>>>>
> > > > > >>>>>>>>
> > > > > >>>>>>>
> > > > > >>>>>>
> > > > > >>>>>
> > > > > >>>
> > > > > >>
> > > > >
> > > >
> > >
> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-150+-+Kafka-Streams+Cogroup
> > > > > >>>>>>>>> here
> > > > > >>>>>>>>> is a link
> > > > > >>>>>>>>>
> > > > > >>>>>>>>> On Mon, Oct 14, 2019 at 2:52 PM Walker Carlson <
> > > > > >>>>>> wcarl...@confluent.io>
> > > > > >>>>>>>>> wrote:
> > > > > >>>>>>>>>
> > > > > >>>>>>>>>> Hello all,
> > > > > >>>>>>>>>>
> > > > > >>>>>>>>>> I have picked up and updated KIP-150. Due to changes to
> > the
> > > > > >>>>> project
> > > > > >>>>>>>> since
> > > > > >>>>>>>>>> KIP #150 was written there are a few items that need to be
> > > > > >>>>> updated.
> > > > > >>>>>>>>>>
> > > > > >>>>>>>>>> First item that changed is the adoption of the
> > Materialized
> > > > > >>>>>> parameter.
> > > > > >>>>>>>>>>
> > > > > >>>>>>>>>> The second item is the WindowedBy. How the old KIP handles
> > > > > >>>>> windowing
> > > > > >>>>>>> is
> > > > > >>>>>>>>>> that it overloads the aggregate function to take in a
> > Window
> > > > > >>>>> object
> > > > > >>>>>> as
> > > > > >>>>>>>> well
> > > > > >>>>>>>>>> as the other parameters. The current practice to window
> > > > > >>>>>>> grouped-streams
> > > > > >>>>>>>> is
> > > > > >>>>>>>>>> to call windowedBy and receive a windowed stream object.
> > The
> > > > > >>>>>> existing
> > > > > >>>>>>>>>> interface for a windowed stream made from a grouped stream
> > > > will
> > > > > >>>>> not
> > > > > >>>>>>> work
> > > > > >>>>>>>>>> for cogrouped streams. Hence, we have to make new
> > interfaces
> > > > for
> > > > > >>>>>>>> cogrouped
> > > > > >>>>>>>>>> windowed streams.
> > > > > >>>>>>>>>>
> > > > > >>>>>>>>>> Please take a look, I would like to hear your feedback,
> > > > > >>>>>>>>>>
> > > > > >>>>>>>>>> Walker
> > > > > >>>>>>>>>>
> > > > > >>>>>>>>>
> > > > > >>>>>>>>
> > > > > >>>>>>>>
> > > > > >>>>>>>
> > > > > >>>>>>
> > > > > >>>>>
> > > > > >>>>
> > > > > >>>
> > > > > >>>
> > > > > >>
> > > > > >
> > > > >
> > > > >
> > > >
> > > > --
> > > > -- Guozhang
> > > >
> > >
> >
> >
> > --
> > -- Guozhang
> >
