On 30 Jan 2017, at 03:56, Guozhang Wang <wangg...@gmail.com> wrote:
Thinking loud here about the API options (materialize v.s. overloaded
functions) and its impact on IQ:
1. The first issue of the current DSL is that, there is inconsistency upon
whether / how KTables should be materialized:
a) in many cases the library HAS TO materialize KTables no matter what,
e.g. KStream / KTable aggregation resulted KTables, and hence we enforce
users to provide store names and throw RTE if it is null;
b) in some other cases, the KTable can be materialized or not; for
example in KStreamBuilder.table(), store names can be nullable and in which
case the KTable would not be materialized;
c) in some other cases, the KTable will never be materialized, for
example KTable.filter() resulted KTables, and users have no options to
enforce them to be materialized;
d) this is related to a), where some KTables are required to be
materialized, but we do not enforce users to provide a state store name,
e.g. KTables involved in joins; a RTE will be thrown not immediately but
later in this case.
2. The second issue is related to IQ, where state stores are accessed by
their state stores; so only those KTable's that have user-specified state
stores will be queryable. But because of 1) above, many stores may not be
interested to users for IQ but they still need to provide a (dummy?) state
store name for them; while on the other hand users cannot query some state
stores, e.g. the ones generated by KTable.filter() as there is no APIs for
them to specify a state store name.
3. We are aware from user feedbacks that such backend details would be
better be abstracted away from the DSL layer, where app developers should
just focus on processing logic, while state stores along with their
changelogs etc would better be in a different mechanism; same arguments
have been discussed for serdes / windowing triggers as well. For serdes
specifically, we had a very long discussion about it and concluded that, at
least in Java7, we cannot completely abstract serde away in the DSL, so we
choose the other extreme to enforce users to be completely aware of the
serde requirements when some KTables may need to be materialized vis
overloaded API functions. While for the state store names, I feel it is a
different argument than serdes (details below).
So to me, for either materialize() v.s. overloaded functions directions,
the first thing I'd like to resolve is the inconsistency issue mentioned
above. So in either case: KTable materialization will not be affect by user
providing state store name or not, but will only be decided by the library
when it is necessary. More specifically, only join operator and
builder.table() resulted KTables are not always materialized, but are still
likely to be materialized lazily (e.g. when participated in a join
operator).
For overloaded functions that would mean:
a) we have an overloaded function for ALL operators that could result
in a KTable, and allow it to be null (i.e. for the function without this
param it is null by default);
b) null-state-store-name do not indicate that a KTable would not be
materialized, but that it will not be used for IQ at all (internal state
store names will be generated when necessary).
For materialize() that would mean:
a) we will remove state store names from ALL operators that could
result in a KTable.
b) KTables that not calling materialized do not indicate that a KTable
would not be materialized, but that it will not be used for IQ at all
(internal state store names will be generated when necessary).
Again, in either ways the API itself does not "hint" about anything for
materializing a KTable or not at all; it is still purely determined by the
library when parsing the DSL for now.
Following these thoughts, I feel that 1) we should probably change the name
"materialize" since it may be misleading to users as what actually happened
behind the scene, to e.g. Damian suggested "queryableStore(String storeName)",
which returns a QueryableStateStore, and can replace the
`KafkaStreams.store` function; 2) comparing those two options assuming we
get rid of the misleading function name, I personally favor not adding more
overloading functions as it keeps the API simpler.
Guozhang
On Sat, Jan 28, 2017 at 2:32 PM, Jan Filipiak <jan.filip...@trivago.com>
wrote:
Hi,
thanks for your mail, felt like this can clarify some things! The thread
unfortunately split but as all branches close in on what my suggestion was
about Ill pick this to continue
Of course only the table the user wants to query would be materialized.
(retrieving the queryhandle implies materialisation). So In the example of
KTable::filter if you call
getIQHandle on both tables only the one source that is there would
materialize and the QueryHandleabstraction would make sure it gets mapped
and filtered and what not uppon read as usual.
Of Course the Object you would retrieve would maybe only wrap the
storeName / table unique identifier and a way to access the streams
instance and then basically uses the same mechanism that is currently used.
From my point of view this is the least confusing way for DSL users. If
its to tricky to get a hand on the streams instance one could ask the user
to pass it in before executing queries, therefore making sure the streams
instance has been build.
The effort to implement this is indeed some orders of magnitude higher
than the overloaded materialized call. As long as I could help getting a
different view I am happy.
Best Jan
On 28.01.2017 09:36, Eno Thereska wrote:
Hi Jan,
I understand your concern. One implication of not passing any store name
and just getting an IQ handle is that all KTables would need to be
materialised. Currently the store name (or proposed .materialize() call)
act as hints on whether to materialise the KTable or not. Materialising
every KTable can be expensive, although there are some tricks one can play,
e.g., have a virtual store rather than one backed by a Kafka topic.
However, even with the above, after getting an IQ handle, the user would
still need to use IQ APIs to query the state. As such, we would still
continue to be outside the original DSL so this wouldn't address your
original concern.
So I read this suggestion as simplifying the APIs by removing the store
name, at the cost of having to materialise every KTable. It's definitely an
option we'll consider as part of this KIP.
Thanks
Eno
On 28 Jan 2017, at 06:49, Jan Filipiak <jan.filip...@trivago.com> wrote:
Hi Exactly
I know it works from the Processor API, but my suggestion would prevent
DSL users dealing with storenames what so ever.
In general I am pro switching between DSL and Processor API easily. (In
my Stream applications I do this a lot with reflection and instanciating
KTableImpl) Concerning this KIP all I say is that there should be a DSL
concept of "I want to expose this __KTable__. This can be a Method like
KTable::retrieveIQHandle():InteractiveQueryHandle, the table would know
to materialize, and the user had a reference to the "store and the
distributed query mechanism by the Interactive Query Handle" under the hood
it can use the same mechanism as the PIP people again.
I hope you see my point J
Best Jan
#DeathToIQMoreAndBetterConnectors :)
On 27.01.2017 21:59, Matthias J. Sax wrote:
Jan,
the IQ feature is not limited to Streams DSL but can also be used for
Stores used in PAPI. Thus, we need a mechanism that does work for PAPI
and DSL.
Nevertheless I see your point and I think we could provide a better API
for KTable stores including the discovery of remote shards of the same
KTable.
@Michael: Yes, right now we do have a lot of overloads and I am not a
big fan of those -- I would rather prefer a builder pattern. But that
might be a different discussion (nevertheless, if we would aim for a API
rework, we should get the changes with regard to stores right from the
beginning on, in order to avoid a redesign later on.)
something like:
stream.groupyByKey()
.window(TimeWindow.of(5000))
.aggregate(...)
.withAggValueSerde(new CustomTypeSerde())
.withStoreName("storeName);
(This would also reduce JavaDoc redundancy -- maybe a personal pain
point right now :))
-Matthias
On 1/27/17 11:10 AM, Jan Filipiak wrote:
Yeah,
Maybe my bad that I refuse to look into IQ as i don't find them
anywhere
close to being interesting. The Problem IMO is that people need to know
the Store name), so we are working on different levels to achieve a
single goal.
What is your peoples opinion on having a method on KTABLE that returns
them something like a Keyvalue store. There is of course problems like
"it cant be used before the streamthreads are going and groupmembership
is established..." but the benefit would be that for the user there is
a
consistent way of saying "Hey I need it materialized as querries gonna
be comming" + already get a Thing that he can execute the querries on
in
1 step.
What I think is unintuitive here is you need to say materialize on this
Ktable and then you go somewhere else and find its store name and then
you go to the kafkastreams instance and ask for the store with this
name.
So one could the user help to stay in DSL land and therefore maybe
confuse him less.
Best Jan
#DeathToIQMoreAndBetterConnectors :)
On 27.01.2017 16:51, Damian Guy wrote:
I think Jan is saying that they don't always need to be materialized,
i.e.,
filter just needs to apply the ValueGetter, it doesn't need yet
another
physical state store.
On Fri, 27 Jan 2017 at 15:49 Michael Noll <mich...@confluent.io>
wrote:
Like Damian, and for the same reasons, I am more in favor of
overloading
methods rather than introducing `materialize()`.
FWIW, we already have a similar API setup for e.g.
`KTable#through(topicName, stateStoreName)`.
A related but slightly different question is what e.g. Jan Filipiak
mentioned earlier in this thread:
I think we need to explain more clearly why KIP-114 doesn't propose
the
seemingly simpler solution of always materializing tables/state
stores.
On Fri, Jan 27, 2017 at 4:38 PM, Jan Filipiak <
jan.filip...@trivago.com>
wrote:
Hi,
Yeah its confusing, Why shoudn't it be querable by IQ? If you uses
the
ValueGetter of Filter it will apply the filter and should be
completely
transparent as to if another processor or IQ is accessing it? How
can
this
new method help?
I cannot see the reason for the additional materialize method being
required! Hence I suggest leave it alone.
regarding removing the others I dont have strong opinions and it
seems to
be unrelated.
Best Jan
On 26.01.2017 20:48, Eno Thereska wrote:
Forwarding this thread to the users list too in case people would
like
to
comment. It is also on the dev list.
Thanks
Eno
Begin forwarded message:
From: "Matthias J. Sax" <matth...@confluent.io>
Subject: Re: [DISCUSS] KIP-114: KTable materialization and
improved
semantics
Date: 24 January 2017 at 19:30:10 GMT
To: d...@kafka.apache.org
Reply-To: d...@kafka.apache.org
That not what I meant by "huge impact".
I refer to the actions related to materialize a KTable: creating a
RocksDB store and a changelog topic -- users should be aware about
runtime implication and this is better expressed by an explicit
method
call, rather than implicitly triggered by using a different
overload of
a method.
-Matthias
On 1/24/17 1:35 AM, Damian Guy wrote:
I think your definition of a huge impact and mine are rather
different
;-P
Overloading a few methods is not really a huge impact IMO. It is
also a
sacrifice worth making for readability, usability of the API.
On Mon, 23 Jan 2017 at 17:55 Matthias J. Sax <
matth...@confluent.io>
wrote:
I understand your argument, but do not agree with it.
Your first version (even if the "flow" is not as nice) is more
explicit
than the second version. Adding a stateStoreName parameter is quite
implicit but has a huge impact -- thus, I prefer the rather more
verbose
but explicit version.
-Matthias
On 1/23/17 1:39 AM, Damian Guy wrote:
I'm not a fan of materialize. I think it interrupts the flow,
i.e,
table.mapValue(..).materialize().join(..).materialize()
compared to:
table.mapValues(..).join(..)
I know which one i prefer.
My preference is stil to provide overloaded methods where
people can
specify the store names if they want, otherwise we just
generate
them.
On Mon, 23 Jan 2017 at 05:30 Matthias J. Sax
<matth...@confluent.io
wrote:
Hi,
thanks for the KIP Eno! Here are my 2 cents:
1) I like Guozhang's proposal about removing store name from
all
KTable
methods and generate internal names (however, I would do this
as
overloads). Furthermore, I would not force users to call
.materialize()
if they want to query a store, but add one more method
.stateStoreName()
that returns the store name if the KTable is materialized.
Thus,
also
.materialize() must not necessarily have a parameter storeName
(ie,
we
should have some overloads here).
I would also not allow to provide a null store name (to
indicate no
materialization if not necessary) but throw an exception.
This yields some simplification (see below).
2) I also like Guozhang's proposal about KStream#toTable()
3)
3. What will happen when you call materialize on KTable
that is
already
materialized? Will it create another StateStore
(providing
the
name
is
different), throw an Exception?
Currently an exception is thrown, but see below.
If we follow approach (1) from Guozhang, there is no need to
worry
about
a second materialization and also no exception must be
throws. A
call to
.materialize() basically sets a "materialized flag" (ie,
idempotent
operation) and sets a new name.
4)
Rename toStream() to toKStream() for consistency.
Not sure whether that is really required. We also use
`KStreamBuilder#stream()` and `KStreamBuilder#table()`, for
example,
and
don't care about the "K" prefix.
Eno's reply:
I think changing it to `toKStream` would make it absolutely
clear
what
we are converting it to.
I'd say we should probably change the KStreamBuilder methods
(but
not
in
this KIP).
I would keep #toStream(). (see below)
5) We should not remove any methods but only deprecate them.
A general note:
I do not understand your comments "Rejected Alternatives". You
say
"Have
the KTable be the materialized view" was rejected. But your
KIP
actually
does exactly this -- the changelog abstraction of KTable is
secondary
after those changes and the "view" abstraction is what a
KTable is.
And
just to be clear, I like this a lot:
- it aligns with the name KTable
- is aligns with stream-table-duality
- it aligns with IQ
I would say that a KTable is a "view abstraction" (as
materialization is
optional).
-Matthias
On 1/22/17 5:05 PM, Guozhang Wang wrote:
Thanks for the KIP Eno, I have a few meta comments and a few
detailed
comments:
1. I like the materialize() function in general, but I would
like
to
see
how other KTable functions should be updated accordingly. For
example,
1)
KStreamBuilder.table(..) has a state store name parameter, and
we
will
always materialize the KTable unless its state store name is set
to
null;
2) KTable.agg requires the result KTable to be materialized,
and
hence
it
also have a state store name; 3) KTable.join requires the
joining
table
to
be materialized. And today we do not actually have a
mechanism to
enforce
that, but will only throw an exception at runtime if it is not
(e.g.
if
you
have "builder.table("topic", null).join()" a RTE will be
thrown).
I'd make an extended proposal just to kick off the discussion
here:
let's
remove all the state store params in other KTable functions,
and if
in
some
cases KTable have to be materialized (e.g. KTable resulted
from
KXX.agg)
and users do not call materialize(), then we treat it as "users
are
not
interested in querying it at all" and hence use an internal
name
generated
for the materialized KTable; i.e. although it is materialized
the
state
store is not exposed to users. And if users call
materialize()
afterwards
but we have already decided to materialize it, we can replace
the
internal
name with the user's provided names. Then from a user's
point-view,
if
they
ever want to query a KTable, they have to call materialize()
with
a
given
state store name. This approach has one awkwardness though,
that
serdes
and
state store names param are not separated and could be
overlapped
(see
detailed comment #2 below).
2. This step does not need to be included in this KIP, but
just
as a
reference / future work: as we have discussed before, we may
enforce
materialize KTable.join resulted KTables as well in the
future. If
we
do
that, then:
a) KXX.agg resulted KTables are always materialized;
b) KTable.agg requires the aggregating KTable to always be
materialized
(otherwise we would not know the old value);
c) KTable.join resulted KTables are always materialized, and
so
are
the
joining KTables to always be materialized.
d) KTable.filter/mapValues resulted KTables materialization
depend
on
its
parent's materialization;
By recursive induction all KTables are actually always
materialized,
and
then the effect of the "materialize()" is just for specifying
the
state
store names. In this scenario, we do not need to send
Change<V> in
repartition topics within joins any more, but only for
repartitions
topics
within aggregations. Instead, we can just send a "tombstone"
without
the
old value and we do not need to calculate joins twice (one more
time
when
old value is received).
3. I'm wondering if it is worth-while to add a
"KStream#toTable()"
function
which is interpreted as a dummy-aggregation where the new
value
always
replaces the old value. I have seen a couple of use cases of
this,
for
example, users want to read a changelog topic, apply some
filters,
and
then
materialize it into a KTable with state stores without
creating
duplicated
changelog topics. With materialize() and toTable I'd imagine
users
can
specify sth. like:
"
KStream stream = builder.stream("topic1").filter(..);
KTable table = stream.toTable(..);
table.materialize("state1");
"
And the library in this case could set store "state1" 's
changelog
topic
to
be "topic1", and applying the filter on the fly while
(re-)storing
its
state by reading from this topic, instead of creating a
second
changelog
topic like "appID-state1-changelog" which is a semi-duplicate
of
"topic1".
Detailed:
1. I'm +1 with Michael regarding "#toStream"; actually I was
thinking
about
renaming to "#toChangeLog" but after thinking a bit more I
think
#toStream
is still better, and we can just mention in the javaDoc that
it is
transforming its underlying changelog stream to a normal
stream.
2. As Damian mentioned, there are a few scenarios where the
serdes
are
already specified in a previous operation whereas it is not
known
before
calling materialize, for example:
stream.groupByKey.agg(serde).materialize(serde) v.s.
table.mapValues(/*no
serde specified*/).materialize(serde). We need to specify
what are
the
handling logic here.
3. We can remove "KTable#to" call as well, and enforce users
to
call "
KTable.toStream.to" to be more clear.
Guozhang
On Wed, Jan 18, 2017 at 3:22 AM, Eno Thereska <
eno.there...@gmail.com>
wrote:
I think changing it to `toKStream` would make it absolutely
clear
what
we
are converting it to.
I'd say we should probably change the KStreamBuilder methods
(but
not
in
this KIP).
Thanks
Eno
On 17 Jan 2017, at 13:59, Michael Noll <
mich...@confluent.io>
wrote:
Rename toStream() to toKStream() for consistency.
Not sure whether that is really required. We also use
`KStreamBuilder#stream()` and `KStreamBuilder#table()`, for
example,
and
don't care about the "K" prefix.
On Tue, Jan 17, 2017 at 10:55 AM, Eno Thereska <
eno.there...@gmail.com
wrote:
Thanks Damian, answers inline:
On 16 Jan 2017, at 17:17, Damian Guy <
damian....@gmail.com>
wrote:
Hi Eno,
Thanks for the KIP. Some comments:
1. I'd probably rename materialized to materialize.
Ok.
2. I don't think the addition of the new Log compaction
mechanism
is
necessary for this KIP, i.e, the KIP is useful without it.
Maybe
that
should be a different KIP?
Agreed, already removed. Will do a separate KIP for that.
3. What will happen when you call materialize on KTable
that is
already
materialized? Will it create another StateStore (providing
the
name
is
different), throw an Exception?
Currently an exception is thrown, but see below.
4. Have you considered overloading the existing KTable
operations
to
add
a state store name? So if a state store name is provided,
then
materialize
a state store? This would be my preferred approach as i
don't
think
materialize is always a valid operation.
Ok I can see your point. This will increase the KIP size
since
I'll
need
to enumerate all overloaded methods, but it's not a problem.
5. The materialize method will need ta value Serde as some
operations,
i.e., mapValues, join etc can change the value types
6. https://issues.apache.org/jira/browse/KAFKA-4609 - might
mean
that
we
always need to materialize the StateStore for KTable-KTable
joins.
If
that
is the case, then the KTable Join operators will also need
Serde
information.
I'll update the KIP with the serdes.
Thanks
Eno
Cheers,
Damian
On Mon, 16 Jan 2017 at 16:44 Eno Thereska <
eno.there...@gmail.com>
wrote:
Hello,
We created "KIP-114: KTable materialization and improved
semantics"
to
solidify the KTable semantics in Kafka Streams:
https://cwiki.apache.org/confluence/display/KAFKA/KIP-
114%3A+KTable+materialization+and+improved+semantics
<
https://cwiki.apache.org/confluence/display/KAFKA/KIP-
114:+KTable+materialization+and+improved+semantics
Your feedback is appreciated.
Thanks
Eno
--
-- Guozhang
