Inline
rather sparse for the lack of time.
Sadly I can't agree to any of your arguments and I _hate_ how its gonna
look,
but we can't have this discussion for ever.
I think I explained everything in enough detail so my points can make sense.
if someone is interested and has specific questions, can always approach me.
Otherwise I am just going to drink the kool-aid now. :(
Best Jan
On 08.08.2017 20:37, Guozhang Wang wrote:
Hello Jan,
Thanks for your feedback. Trying to explain them a bit more here since I
think there are still a bit mis-communication here:
Here are a few things I need to clarify for KIP-182 first:
1. KIP-182 is mainly about refactoring the public APIs, NOT for making any
optimizations on the internal implementations. So we only care that these
public APIs changes do not forbid us to make the internal implementations
in the near future.
To give you a concrete example, as you mentioned that KTableValueGetterSupplier
is NOT used in IQ, and that a materialized physical store is always used
today. Yes that is true, and we do have plans to optimize this case soon;
for example, it is still doable with the proposed KIP-182 that we can
remove the physical materialized store but use KTableValueGetterSupplier to
read form a up-stream's physical store and apply the optimizations. Another
example you mentioned is stream-stream join, where each stream is
physically materialized into a store, we can definitely optimize this in
the future to remove the physical materialized store but use something
else, e.g. a in-memory buffer. Such optimizations are NOT blocked by the
updated public APIs of KIP-182.
One of the big goals of the refactoring at least was to get rid of the
overloads
to make implementation of new features easier as one has not to take
care about
all the overloads. Folding 2 Overloads into 1 with a Builder that has 2
way of beeing build
wont help much here.
Having the DSL express very closely what happens will only help people
not getting confused.
Having the store overload on every operation is just plain confusing
right now.
2. One concern you raise that KIP-182 may actually block such
optimizations, is that if users do specify a StateStoreSupplier then we
cannot optimize that away. That is true, but that is by design: if user do
specify a state store supplier in Materialized API, that is equal to say
"forget about doing any smart things library, just use what I give to you".
In other words, the above mentioned optimizations can be applied if users
do not enforce any specific StateStoreSupplier, for example in
public static <K, V, S extends StateStore> Materialized<K, V, S>
as(final String
storeName)
i.e. user only provide a store name, which is similar like handler token
for IQ; then the library still have the freedom to do smart things
internally which is totally hidden from the users. It is similar to, like
in RDBMS or some NoSQL stores like in HIVE / Cassandra: the store engine do
not have the freedom to do those query plan optimizations if users already
enforce the specs like join ordering, query plan generation rules, etc.
You call the same method with the builder build differently and its
going todo
different things. That is my definition of unituitive + The code
internally has to become
dead ugly as it needs to apply these optimisations basically in the same
method call or
at the place the Builder is evaluated. This just cries for ugly internal
code. There is no
way this can become pretty
3. About whether it is worthwhile to "not break the fluent interface", a
key point that we cannot just try to optimize one or two use cases, but
consider the whole user space, and ask what are the percentage of users
that may get affected. Note that in the DSL we have overloaded functions
where Materialized / Joined / other options are NOT needed so for most
normal users they do not need to worry about the specs at all.
So suppose there are only X% "advanced" users who would want to enforce
some state store suppliers, and Y% who like to use IQ, 100-X-Y percent of
normal users see no difference in terms of programming for either of these
two approaches: whether to separate the specs into a different set of APIs.
And for the Y percent of users they are most likely to just use the
simplest APIs which is `operator(..., Materialized.as(storeName))` which
does not sound too bad as to `table = operator(...);
table.materialize(storeName)`. In other words we use the first approach
then only X percent of users may have an awkward programming with complex
option specs along with the operator; if we use the second approach the X+Y
users need to break its programing fluency to call `table.materialize`
separately. And my personal guess is that
0 < X << Y < 1, and that X is very minor compared to Y. That is why I feel
this is a good trade-off.
The keypoint here is that It doesn't matter. Any sufficiently usefull
topology
will get broken up by the user anyways to keep his code sane. And it
will most
likely be broken up in the interesting parts where IQ would be usefull.
The math of
how many people are affected by this is therefore not important.
Additionally
the comparission doesnt make sense as the X-group can still go with a
fluent
interface only the Y-people need to break their fluent interface
Guozhang
On Fri, Aug 4, 2017 at 6:18 AM, Jan Filipiak <jan.filip...@trivago.com>
wrote:
Hi Guozhang,
thank you very much for the reply. It explained a lot more of your
reasoning to me
once again!
I have to disagree with you on the first point. As you mentioned the Join
Case.
A Join is usually a "logically" materialized table and its
KTableValueGetterSupplier
is to be used when one wants todo a lookup. But this is not at all what is
currently
Happening. The join merge processor currently maintains its own new
statestore
when join is invoked with Storename or supplier.
This describes the Issue I want to address perfectly. A Joined-Table
doesn't become
querieable because it is a JOINEDtable but because it is a joinedTABLE.
the emphasis here
is that we put the store logic with the join and not the table. It is part
of the join() method invocation and not the KTable Interface. This
abstraction is wrong.
This will always show its ugly face. Lets check your example:
stream.groupByKey(..).aggregate(.., Materializedas("store1")) //
this resulted KTable is materialized in order to complete the aggregation
operation
.filter(Materialized.as("store2"))
// this restuled KTable is not materialized but its
GetterSupplier is implemented to get values from "store1"
Currently this is only half true. For IQ a store is used that is maintained
by the KTableFilterProcessor, for downstream gets like joins the
ValueGetterSupplier is used
and indeed uses store1.
With the String overload (that you picked here on purpose I guess) it
works easier
as you can logically map those. But with the StateStoreSupplier it
wouldn't.
you could not optimize this away as the user is expecting puts and gets to
be called
on what he supplied.
table1.filter(() -> true, InMemoryStore).filter(()->true,SQlLiteStore)
There is no way to optimize these away.
The same argument with the join holds for filter. Its not querrieable
because it got filtered
it is querrieable because its a KTable. That's where the emphasis needs to
go.
The second point was new to me. So I had to think about this in more
detail.
For me the breaking of the flow comes in very natural.
One Stream app I put most of my heart in has the these key metrics:
It has:
8 input topics.
3 1:n Joins
6 Group bys
2 built in Joins
2 built in left joins
some filters and mappers.
this is spanning 390 lines, counting java imports and some more stuff.
The whole topology forms a tree in wich the input topics usually get
joined and then collected to maps
and then joined again and collected to maps again. until they get send to
1 final output topic for consumption in our application servers.
I would argue it is impossible to express this topology as a chain of
calls. What happened is that
usually each join + groupBy tuple became its method taking in the builder
and return the Table
expressing the result of the sub topology. All Ktables that meet each
other with the same key in the
process get joined (most of that happening on the top level). This leads
to breaking in the fluent interface
quite naturally. especially if you have 2 KTables expressing
sub-topologies joined together. One subtopology had to go into the method
call which is unreasonable IMHO.
Even inside these methods we broke the chains. The variable names we used
give intermediate KTables really helped in making the semantics clear. They
are much like CTE's in hive or the required name in Mysql Subquerries. They
help to mark milestones inside the topology.
I would argue that for big topologies. (I haven't seen others but I think
its big) these milestones would
be the most important ones for IQ aswell. So i would argue breaking the
chains is not really a problem in
reality and it can help in many cases. As I laid out, we broke our chained
calls intuitively and it helped
other developers debugging the logic a lot. Even without detailed streams
understanding.
If one really do not want to stop the flow. I could argue that one could
either do something like this
KTable joinresult;
KTable<Integer,Integer> t1 = b.table("laa");
KTable<Integer,Integer> t2 = b.table("luu");
(joinresult = t1.join(t2, (value1, value2) -> value1 + value2))
.filter((key, value) -> false);
or write a little snitch like that
KTable<Integer,Integer> rememberTableandContinue(KTable<Integer,Integer>
t){
joinresult = t;
return t;
}
for usuage as such
rememberTableandContinue(t1.join(t2, (value1, value2) -> value1 + value2))
.filter((key, value) -> false);
These suggestions might not looks so pretty. But in the context of
breaking bigger topology at milestones.
I think everything becomes acceptable really. Probably user would store
that intermediate KTable anyways just for clarity.
To summarize to give a KTable a name. I would always opt to the host
language variable names.
Tables used for IQ are probably tables that are of some sort more
important to the topology than
others and saving them separatly will increase the readability of
topologies by a lot IMO.
For these quick example Topologies that we have floating around in all
places:
I am pretty sure one can go unbroken on them and usually the last table
will be the one that
is needed for IQ then.
Thanks again. The second point really got me thinking, as your perspective
on the importance
of "not break the fluent interface" was not clear to me. I hope I managed
to line out why I
think it shouldn't have such a big weight in the discussion.
PS.: check out Hive CTE, everyone loves them and our Analytic team is
crazy for them
because you can name them and that brings clarity. and you get rid of the
nesting and can
split everything into logical chunks of SQL. KTable variables are the CTE
of kafka streams.
One can probably sell this to people :)
Best Jan
Enjoyed your feedback! hope mine makes sense
On 03.08.2017 00:10, Guozhang Wang wrote:
Hello Jan,
Thanks for your proposal. As Bill mentioned the main difference is that we
extract the user-customizable materialization logic out of the topology
building DSL workflow. And the main motivations are in two folds:
1) efficiency wise, it allows some KTables to not be materialized if
unnecessary, saving one state store instance and changelog topic.
2) programming wise, it looks nicer to separate the topology construction
code from the KTable materialization for IQ uses code.
Here are my thoughts regarding these two points:
Regarding 1), I think with whichever the public APIs (either Damian's
proposal or yours), we can always apply the internal optimization to not
physically materialize the KTable. You can take a look at the internal
interface of "KTableValueGetterSupplier", which is used exactly for this
purposes such that a get call on a "logically" materialized KTable can be
traced back to its parent KTables that are physically materialized in a
state store. So following proposed APIs, for example:
stream.groupByKey(..).aggregate(.., Materializedas("store1")) //
this resulted KTable is materialized in order to complete the aggregation
operation
.filter(Materialized.as("store2"))
// this restuled KTable is not materialized but its
GetterSupplier is implemented to get values from "store1"
Or
table1 = stream.groupByKey(..).aggregate(..);
table2 = table1.filter();
tabel1.queryHandle("store1"); // this resulted KTable is
materialized
in order to complete the aggregation operation
tabel1.queryHandle("store2") // this restuled KTable is not
materialized but its GetterSupplier is implemented to get values from
"store1"
When user query a value for "store2" which is not actually materialized
into a state store, the GetterSupplier will be triggered to in turn query
the store for "store1", and then apply the filter operator on-the-fly to
return the value. So the bottom line is, we can achieve the same
efficiency
optimization with either of the public APIs.
Regarding 2), I actually have proposed a similar API to yours earlier in
this discussion thread:
---------------------------------------
// specifying the topology, should be concise and conveniently
concatenated, no specs of materialization at all
KStream stream1 = builder.stream();
KTable table1 = stream1.groupby(...).aggregate(initializer, aggregator,
sessionMerger, sessionWindows); // do not allow to pass-in a state store
supplier here any more
// additional code to the topology above, could be more prescriptive
than descriptive
// only advanced users would want to code in both parts above; while other
users would only code the topology as above.
table1.materialize("queryableStoreName"); // or..
table1.materialize("queryableStoreName").enableCaching().enableLogging();
// or..
table1.materialize(stateStoreSupplier); // we check type (key-value
types,
windowed or not etc) at starting time and add the metrics / logging /
caching / windowing wrapper on top of the store, or..
table1.materialize(stateStoreSupplier).enableCaching().enableLogging();
//
etc..
---------------------------------------
But one caveat of that, as illustrated above, is that you need to have
separate object of the KTable in order to call either "queryHandle" or
"materialize" (whatever the function name is) for the specifications of
materialization options. This can break the concatenation of the topology
construction part of the code, that you cannot simply add one operator
directly after another. So I think this is a trade-off we have to make and
the current approach looks better in this regard.
Guozhang
On Wed, Aug 2, 2017 at 2:07 PM, Jan Filipiak<jan.filip...@trivago.com>
wrote:
Hi Bill,
totally! So in the original discussion it was mentioned that the
overloads
are nasty when implementing new features. So we wanted to get rid of
them.
But what I felt was that the
copy & pasted code in the KTableProcessors for maintaining IQ stores was
as big as a hurdle as the overloads.
With this proposal I try to shift things into the direction of getting IQ
for free if
KTableValueGetterSupplier is properly implemented (like getting join for
free then). Instead of having the code for maintaining IQ stores all the
places. I realized I can do that while getting rid of the overloads, that
makes me feel my proposal is superior.
Further I try to optimize by using as few stores as possible to give the
user what he needs. That should save all sorts of resources while
allowing
faster rebalances.
The target ultimately is to only have KTableSource and the Aggregators
maintain a Store and provide a ValueGetterSupplier.
Does this makes sense to you?
Best Jan
On 02.08.2017 18:09, Bill Bejeck wrote:
Hi Jan,
Thanks for the effort in putting your thoughts down on paper.
Comparing what I see from your proposal and what is presented in
KIP-182,
one of the main differences is the exclusion of an`Materialized`
instance
in the `KTable` methods.
Can you go into more detail why this is so and the specific problems is
avoids and or solves with this approach?
Thanks!
Bill
On Wed, Aug 2, 2017 at 4:19 AM, Damian Guy <damian....@gmail.com
<mailto:
damian....@gmail.com>> wrote:
Hi Jan,
Thanks for taking the time to put this together, appreciated. For
the
benefit of others would you mind explaining a bit about your
motivation?
Cheers,
Damian
On Wed, 2 Aug 2017 at 01:40 Jan Filipiak <jan.filip...@trivago.com
<mailto:jan.filip...@trivago.com>> wrote:
> Hi all,
>
> after some further discussions, the best thing to show my Idea
of how it
> should evolve would be a bigger mock/interface description.
> The goal is to reduce the store maintaining processors to only
the
> Aggregators + and KTableSource. While having KTableSource
optionally
> materialized.
>
> Introducing KTable:copy() will allow users to maintain state
twice if
> they really want to. KStream::join*() wasn't touched. I never
personally
> used that so I didn't feel
> comfortable enough touching it. Currently still making up my
mind. None
> of the suggestions made it querieable so far. Gouzhangs
'Buffered' idea
> seems ideal here.
>
> please have a look. Looking forward for your opinions.
>
> Best Jan
>
>
>
> On 21.06.2017 17 <tel:21.06.2017%2017>:24, Eno Thereska wrote:
> > (cc’ing user-list too)
> >
> > Given that we already have StateStoreSuppliers that are
configurable
> using the fluent-like API, probably it’s worth discussing the
other
> examples with joins and serdes first since those have many
overloads and
> are in need of some TLC.
> >
> > So following your example, I guess you’d have something like:
> > .join()
> > .withKeySerdes(…)
> > .withValueSerdes(…)
> > .withJoinType(“outer”)
> >
> > etc?
> >
> > I like the approach since it still remains declarative and
it’d reduce
> the number of overloads by quite a bit.
> >
> > Eno
> >
> >> On Jun 21, 2017, at 3:37 PM, Damian Guy <damian....@gmail.com
<mailto:damian....@gmail.com>> wrote:
> >>
> >> Hi,
> >>
> >> I'd like to get a discussion going around some of the API
choices we've
> >> made in the DLS. In particular those that relate to stateful
operations
> >> (though this could expand).
> >> As it stands we lean heavily on overloaded methods in the
API, i.e,
> there
> >> are 9 overloads for KGroupedStream.count(..)! It is becoming
noisy and i
> >> feel it is only going to get worse as we add more optional
params. In
> >> particular we've had some requests to be able to turn caching
off, or
> >> change log configs, on a per operator basis (note this can
be done now
> if
> >> you pass in a StateStoreSupplier, but this can be a bit
cumbersome).
> >>
> >> So this is a bit of an open question. How can we change the
DSL
> overloads
> >> so that it flows, is simple to use and understand, and is
easily
> extended
> >> in the future?
> >>
> >> One option would be to use a fluent API approach for
providing the
> optional
> >> params, so something like this:
> >>
> >> groupedStream.count()
> >> .withStoreName("name")
> >> .withCachingEnabled(false)
> >> .withLoggingEnabled(config)
> >> .table()
> >>
> >>
> >>
> >> Another option would be to provide a Builder to the count
method, so it
> >> would look something like this:
> >> groupedStream.count(new
> >> CountBuilder("storeName").withCachingEnabled(false).build())
> >>
> >> Another option is to say: Hey we don't need this, what are
you on about!
> >>
> >> The above has focussed on state store related overloads, but
the same
> ideas
> >> could be applied to joins etc, where we presently have many
join
> methods
> >> and many overloads.
> >>
> >> Anyway, i look forward to hearing your opinions.
> >>
> >> Thanks,
> >> Damian
>
>
