[DISCUSS] Enhance Support for Multicast Communication Pattern

[SHI Xiaogang]

Hi, Yun Gao

The discussion seems to move in a different direction, changing from
supporting multicasting to implementing new iteration libraries on data
streams.

Regarding the broadcast events in iterations, many details of new iteration
libraries are unclear,
1. How the iteration progress is determined and notified? The iterations
are synchronous or asynchronous? As far as i know, progress tracking for
asynchronous iterations is very difficult.
2. Do async I/O operators allowed in the iterations? If so, how the
broadcast events are checkpointed and restored? How broadcast events are
distributed when the degree of parallelism changes?
3. Do the emitted broadcast events carry the sender's index? Will they be
aligned in a similar way to checkpoint barriers in downstream operators?
4. In the case of synchronous iterations, do we need something similar to
barrier buffers to guarantee the correctness of iterations?
5. Will checkpointing be enabled in iterations? If checkpointing is
enabled, how will checkpoint barriers interact with broadcast events?

I think a detailed design document for iterations will help understand
these problems, hencing improving the discussion.

I also suggest a new thread for the discussion on iterations.
This thread should focus on multicasting and discuss those problems related
to multicasting, including how data is delivered and states are partitioned.

Regards,
Xiaogang

Yun Gao <[hidden email]> 于2019年8月26日周一 下午11:35写道：

>
> Hi,
>
> Very thanks for all the points raised !
>
> @Piotr For using another edge to broadcast the event, I think it may not
> be able to address the iteration case. The primary problem is that with
> two edges we cannot ensure the order of records. However, In the iteration
> case, the broadcasted event is used to mark the progress of the iteration
> and it works like watermark, thus its position relative to the normal
> records can not change.
> And @Piotr, @Xiaogang, for the requirements on the state, I think
> different options seems vary. The first option is to allow Operator<T> to
> broadcast a separate event and have a separate process method for this
> event. To be detail, we may add a new type of StreamElement called Event
> and allow Operator<T> to broadcastEmit Event. Then in the received side, we
> could add a new `processEvent` method to the (Keyed)ProcessFunction.
> Similar to the broadcast side of KeyedBroadcastProcessFunction, in this new
> method users cannot access keyed state with specific key, but can register
> a state function to touch all the elements in the keyed state. This option
> needs to modify the runtime to support the new type of StreamElement, but
> it does not affect the semantics of states and thus it has no requirements
> on state.
> The second option is to allow Operator<T> to broadcastEmit T and in the
> receiver side, user can process the broadcast element with the existing
> process method. This option is consistent with the OperatorState, but for
> keyedState we may send a record to tasks that do not containing the
> corresponding keyed state, thus it should require some changes on the State.
> The third option is to support the generic Multicast. For keyedState it
> also meets the problem of inconsistency between network partitioner and
> keyed state partitioner, and if we want to rely on it to implement the
> non-key join, it should be also meet the problem of cannot control the
> partitioning of operator state. Therefore, it should also require some
> changes on the State.
> Then for the different scenarios proposed, the iteration case in fact
> requires exactly the ability to broadcast a different event type. In the
> iteration the fields of the progress event are in fact different from that
> of normal records. It does not contain actual value but contains some
> fields for the downstream operators to align the events and track the
> progress. Therefore, broadcasting a different event type is able to solve
> the iteration case without the requirements on the state. Besides, allowing
> the operator to broadcast a separate event may also facilitate some other
> user cases, for example, users may notify the downstream operators to
> change logic if some patterns are matched. The notification might be
> different from the normal records and users do not need to uniform them
> with a wrapper type manually if the operators are able to broadcast a
> separate event. However, it truly cannot address the non-key join
> scenarios.
> Since allowing broadcasting a separate event seems to be able to serve as
> a standalone functionality, and it does not require change on the state, I
> am thinking that is it possible for us to partition to multiple steps and
> supports broadcasting events first ? At the same time we could also
> continue working on other options to support more scenarios like non-key
> join and they seems to requires more thoughts.
>
> Best,
> Yun
>
>
>
> ------------------------------------------------------------------
> From:Piotr Nowojski <[hidden email]>
> Send Time:2019 Aug. 26 (Mon.) 18:59
> To:dev <[hidden email]>
> Cc:Yun Gao <[hidden email]>
> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern
>
> Hi,
>
> Xiaogang, those things worry me the most.
> 1. Regarding the broadcasting, doesn’t the BroadcastState [1] cover our
> issues? Can not we construct a job graph, where one operator has two
> outputs, one keyed another broadcasted, which are wired together back to
> the KeyedBroadcastProcessFunction or BroadcastProcessFunction?
>
> 2. Multicast on keyed streams, might be done by iterating over all of the
> keys. However I have a feeling that might not be the feature which
> distributed cross/theta joins would want, since they would probably need a
> guarantee to have only a single key per operator instance.
>
> Kurt, by broadcast optimisation do you mean [2]?
>
> I’m not sure if we should split the discussion yet. Most of the changes
> required by either multicast or broadcast will be in the API/state layers.
> Runtime changes for broadcast would be almost none (just exposing existing
> features) and for multicast they shouldn't be huge as well. However maybe
> we should consider those two things together at the API level, so that we
> do not make wrong decisions when just looking at the simpler/more narrow
> broadcast support?
>
> Piotrek
>
> [1]
> https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/state/broadcast_state.html
> [2] https://github.com/apache/flink/pull/7713
>
>
> On 26 Aug 2019, at 09:35, Kurt Young <[hidden email]> wrote:
> From SQL's perspective, distributed cross join is a valid feature but not
> very
> urgent. Actually this discuss reminds me about another useful feature
> (sorry
> for the distraction):
>
> when doing broadcast in batch shuffle mode, we can make each producer only
> write one copy of the output data, but not for every consumer. Broadcast
> join
> is much more useful, and this is a very important optimization. Not sure if
> we
> have already consider this.
>
> Best,
> Kurt
>
>
> On Mon, Aug 26, 2019 at 12:16 PM Guowei Ma <[hidden email]> wrote:
>
> Thanks Yun for bringing up this discussion and very thanks for all the deep
> thoughts!
>
> For now, I think this discussion contains two scenarios: one if for
> iteration library support and the other is for SQL join support. I think
> both of the two scenarios are useful but they seem to have different best
> suitable solutions. For making the discussion more clear, I would suggest
> to split the discussion into two threads.
>
> And I agree with Piotr that it is very tricky that a keyed stream received
> a "broadcast element". So we may add some new interfaces, which could
> broadcast or process some special "broadcast event". In that way "broadcast
> event" will not be sent with the normal process.
>
> Best,
> Guowei
>
>
> SHI Xiaogang <[hidden email]> 于2019年8月26日周一 上午9:27写道：
>
> Hi all,
>
> I also think that multicasting is a necessity in Flink, but more details
> are needed to be considered.
>
> Currently network is tightly coupled with states in Flink to achieve
> automatic scaling. We can only access keyed states in keyed streams and
> operator states in all streams.
> In the concrete example of theta-joins implemented with mutlticasting,
> the
> following questions exist:
>
>    - In which type of states will the data be stored? Do we need another
>    type of states which is coupled with multicasting streams?
>    - How to ensure the consistency between network and states when jobs
>    scale out or scale in?
>
> Regards,
> Xiaogang
>
> Xingcan Cui <[hidden email]> 于2019年8月25日周日 上午10:03写道：
>
> Hi all,
>
> Sorry for joining this thread late. Basically, I think enabling
> multicast
> pattern could be the right direction, but more detailed implementation
> policies need to be discussed.
>
> Two years ago, I filed an issue [1] about the multicast API. However,
> due
> to some reasons, it was laid aside. After that, when I tried to
> cherry-pick
> the change for experimental use, I found the return type of
> `selectChannels()` method had changed from `int[]` to `int`, which
> makes
> the old implementation not work anymore.
>
> From my side, the multicast has always been used for theta-join. As far
> as
> I know, it’s an essential requirement for some sophisticated joining
> algorithms. Until now, the Flink non-equi joins can still only be
> executed
> single-threaded. If we'd like to make some improvements on this, we
> should
> first take some measures to support multicast pattern.
>
> Best,
> Xingcan
>
> [1] https://issues.apache.org/jira/browse/FLINK-6936
>
> On Aug 24, 2019, at 5:54 AM, Zhu Zhu <[hidden email]> wrote:
>
> Hi Piotr,
>
> Thanks for the explanation.
> Agreed that the broadcastEmit(record) is a better choice for
> broadcasting
> for the iterations.
> As broadcasting for the iterations is the first motivation, let's
> support
> it first.
>
> Thanks,
> Zhu Zhu
>
> Yun Gao <[hidden email]> 于2019年8月23日周五 下午11:56写道：
>
>     Hi Piotr,
>
>      Very thanks for the suggestions!
>
>     Totally agree with that we could first focus on the broadcast
> scenarios and exposing the broadcastEmit method first considering
> the
> semantics and performance.
>
>     For the keyed stream, I also agree with that broadcasting keyed
> records to all the tasks may be confused considering the semantics
> of
> keyed
> partitioner. However, in the iteration case supporting broadcast
> over
> keyed
> partitioner should be required since users may create any subgraph
> for
> the
> iteration body, including the operators with key. I think a possible
> solution to this issue is to introduce another data type for
> 'broadcastEmit'. For example, for an operator Operator<T>, it may
> broadcast
> emit another type E instead of T, and the transmitting E will bypass
> the
> partitioner and setting keyed context. This should result in the
> design
> to
> introduce customized operator event (option 1 in the document). The
> cost of
> this method is that we need to introduce a new type of StreamElement
> and
> new interface for this type, but it should be suitable for both
> keyed
> or
> non-keyed partitioner.
>
> Best,
> Yun
>
>
>
> ------------------------------------------------------------------
> From:Piotr Nowojski <[hidden email]>
> Send Time:2019 Aug. 23 (Fri.) 22:29
> To:Zhu Zhu <[hidden email]>
> Cc:dev <[hidden email]>; Yun Gao <[hidden email]>
> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
> Pattern
>
> Hi,
>
> If the primary motivation is broadcasting (for the iterations) and
> we
> have
> no immediate need for multicast (cross join), I would prefer to
> first
> expose broadcast via the DataStream API and only later, once we
> finally
> need it, support multicast. As I wrote, multicast would be more
> challenging
> to implement, with more complicated runtime and API. And re-using
> multicast
> just to support broadcast doesn’t have much sense:
>
> 1. It’s a bit obfuscated. It’s easier to understand
> collectBroadcast(record) or broadcastEmit(record) compared to some
> multicast channel selector that just happens to return all of the
> channels.
> 2. There are performance benefits of explicitly calling
> `RecordWriter#broadcastEmit`.
>
>
> On a different note, what would be the semantic of such broadcast
> emit
> on
> KeyedStream? Would it be supported? Or would we limit support only
> to
> the
> non-keyed streams?
>
> Piotrek
>
> On 23 Aug 2019, at 12:48, Zhu Zhu <[hidden email]> wrote:
>
> Thanks Piotr,
>
> Users asked for this feature sometimes ago when they migrating
> batch
> jobs to Flink(Blink).
> It's not very urgent as they have taken some workarounds to solve
> it.(like partitioning data set to different job vertices)
> So it's fine to not make it top priority.
>
> Anyway, as a commonly known scenario, I think users can benefit
> from
> cross join sooner or later.
>
> Thanks,
> Zhu Zhu
>
> Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
> 于2019年8月23日周五 下午6:19写道：
> Hi,
>
> Thanks for the answers :) Ok I understand the full picture now. +1
> from
> my side on solving this issue somehow. But before we start
> discussing
> how
> to solve it one last control question:
>
> I guess this multicast is intended to be used in blink planner,
> right?
> Assuming that we implement the multicast support now, when would it
> be
> used
> by the blink? I would like to avoid a scenario, where we implement
> an
> unused feature and we keep maintaining it for a long period of time.
>
> Piotrek
>
> PS, try to include motivating examples, including concrete ones in
> the
> proposals/design docs, for example in the very first paragraph.
> Especially
> if it’s a commonly known feature like cross join :)
>
> On 23 Aug 2019, at 11:38, Yun Gao <[hidden email]>
> wrote:
>
>    Hi Piotr,
>
>       Thanks a lot for sharing the thoughts!
>
>       For the iteration, agree with that multicasting is not
> necessary. Exploring the broadcast interface to Output of the
> operators
> in
> some way should also solve this issue, and I think it should be even
> more
> convenient to have the broadcast method for the iteration.
>
>       Also thanks Zhu Zhu for the cross join case!
> Best,
>  Yun
>
>
>
> ------------------------------------------------------------------
> From:Zhu Zhu <[hidden email] <mailto:[hidden email]>>
> Send Time:2019 Aug. 23 (Fri.) 17:25
> To:dev <[hidden email] <mailto:[hidden email]>>
> Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
> Pattern
>
> Hi Piotr,
>
> Yes you are right it's a distributed cross join requirement.
> Broadcast join can help with cross join cases. But users cannot
> use
> it
> if the data set to join is too large to fit into one subtask.
>
> Sorry for left some details behind.
>
> Thanks,
> Zhu Zhu
> Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
> 于2019年8月23日周五 下午4:57写道：
> Hi Yun and Zhu Zhu,
>
> Thanks for the more detailed example Zhu Zhu.
>
> As far as I understand for the iterations example we do not need
> multicasting. Regarding the Join example, I don’t fully understand
> it.
> The
> example that Zhu Zhu presented has a drawback of sending both tables
> to
> multiple nodes. What’s the benefit of using broadcast join over a
> hash
> join
> in such case? As far as I know, the biggest benefit of using
> broadcast
> join
> instead of hash join is that we can avoid sending the larger table
> over
> the
> network, because we can perform the join locally. In this example we
> are
> sending both of the tables to multiple nodes, which should defeat
> the
> purpose.
>
> Is it about implementing cross join or near cross joins in a
> distributed fashion?
>
> if we introduce a new MulticastRecordWriter
>
> That’s one of the solutions. It might have a drawback of 3 class
> virtualisation problem (We have RecordWriter and
> BroadcastRecordWriter
> already). With up to two implementations, JVM is able to
> devirtualise
> the
> calls.
>
> Previously I was also thinking about just providing two different
> ChannelSelector interfaces. One with `int[]` and
> `SingleChannelSelector`
> with plain `int` and based on that, RecordWriter could perform some
> magic
> (worst case scenario `instaceof` checks).
>
> Another solution might be to change `ChannelSelector` interface
> into
> an iterator.
>
> But let's discuss the details after we agree on implementing this.
>
> Piotrek
>
> On 23 Aug 2019, at 10:20, Yun Gao <[hidden email] <mailto:
> [hidden email]>> wrote:
>
>  Hi Piotr,
>
>       Thanks a lot for the suggestions!
>
>       The core motivation of this discussion is to implement a
> new
> iteration library on the DataStream, and it requires to insert
> special
> records in the stream to notify the progress of the iteration. The
> mechanism of such records is very similar to the current Watermark,
> and
> we
> meet the problem of sending normal records according to the
> partition
> (Rebalance, etc..) and also be able to broadcast the inserted
> progress
> records to all the connected records. I have read the notes in the
> google
> doc and I totally agree with that exploring the broadcast interface
> in
> RecordWriter in some way is able to solve this issue.
>
>      Regarding to `int[] ChannelSelector#selectChannels()`, I'm
> wondering if we introduce a new MulticastRecordWriter and left the
> current
> RecordWriter untouched, could we avoid the performance degradation ?
> Since
> with such a modification the normal RecordWriter does not need to
> iterate
> the return array by ChannelSelector, and the only difference will be
> returning an array instead of an integer, and accessing the first
> element
> of the returned array instead of reading the integer directly.
>
> Best,
> Yun
>
>
> ------------------------------------------------------------------
> From:Piotr Nowojski <[hidden email] <mailto:
> [hidden email]
>
> Send Time:2019 Aug. 23 (Fri.) 15:20
> To:dev <[hidden email] <mailto:[hidden email]>>
> Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
> Pattern
>
> Hi,
>
> Yun:
>
> Thanks for proposing the idea. I have checked the document and
> left
> couple of questions there, but it might be better to answer them
> here.
>
> What is the exact motivation and what problems do you want to
> solve?
> We have dropped multicast support from the network stack [1] for two
> reasons:
> 1. Performance
> 2. Code simplicity
>
> The proposal to re introduce `int[]
> ChannelSelector#selectChannels()`
> would revert those changes. At that time we were thinking about a
> way
> how
> to keep the multicast support on the network level, while keeping
> the
> performance and simplicity for non multicast cases and there are
> ways
> to
> achieve that. However they would add extra complexity to Flink,
> which
> it
> would be better to avoid.
>
> On the other hand, supporting dual pattern: standard partitioning
> or
> broadcasting is easy to do, as LatencyMarkers are doing exactly
> that.
> It
> would be just a matter of exposing this to the user in some way. So
> before
> we go any further, can you describe your use cases/motivation? Isn’t
> mix of
> standard partitioning and broadcasting enough? Do we need
> multicasting?
>
> Zhu:
>
> Could you rephrase your example? I didn’t quite understand it.
>
> Piotrek
>
> [1] https://issues.apache.org/jira/browse/FLINK-10662 <
> https://issues.apache.org/jira/browse/FLINK-10662> <
> https://issues.apache.org/jira/browse/FLINK-10662 <
> https://issues.apache.org/jira/browse/FLINK-10662>>
>
> On 23 Aug 2019, at 09:17, Zhu Zhu <[hidden email] <mailto:
> [hidden email]> <mailto:[hidden email] <mailto:
> [hidden email]
>
> wrote:
>
> Thanks Yun for starting this discussion.
> I think the multicasting can be very helpful in certain cases.
>
> I have received requirements from users that they want to do
> broadcast
> join, while the data set to broadcast is too large to fit in one
> task.
> Thus the requirement turned out to be to support cartesian
> product
> of
> 2
> data set(one of which can be infinite stream).
> For example, A(parallelism=2) broadcast join B(parallelism=2) in
> JobVertex
> C.
> The idea to is have 4 C subtasks to deal with different
> combinations
> of A/B
> partitions, like C1(A1,B1), C2(A1,B2), C3(A2,B1), C4(A2,B2).
> This requires one record to be sent to multiple downstream
> subtasks,
> but
> not to all subtasks.
>
> With current interface this is not supported, as one record can
> only
> be
> sent to one subtask, or to all subtasks of a JobVertex.
> And the user had to split the broadcast data set manually to
> several
> different JobVertices, which is hard to maintain and extend.
>
> Thanks,
> Zhu Zhu
>
> Yun Gao <[hidden email] <mailto:
> [hidden email] <mailto:[hidden email]
>
> 于2019年8月22日周四 下午8:42写道：
>
> Hi everyone,
>    In some scenarios we met a requirement that some operators
> want
> to
> send records to theirs downstream operators with an multicast
> communication
> pattern. In detail, for some records, the operators want to send
> them
> according to the partitioner (for example, Rebalance), and for
> some
> other
> records, the operators want to send them to all the connected
> operators and
> tasks. Such a communication pattern could be viewed as a kind of
> multicast:
> it does not broadcast every record, but some record will indeed
> be
> sent to
> multiple downstream operators.
>
> However, we found that this kind of communication pattern seems
> could
> not
> be implemented rightly if the operators have multiple consumers
> with
> different parallelism, using the customized partitioner. To solve
> the
> above
> problem, we propose to enhance the support for such kind of
> irregular
> communication pattern. We think there may be two options:
>
>   1. Support a kind of customized operator events, which share
> much
> similarity with Watermark, and these events can be broadcasted to
> the
> downstream operators separately.
>   2. Let the channel selector supports multicast, and also add
> the
> separate RecordWriter implementation to avoid impacting the
> performance of
> the channel selector that does not need multicast.
>
> The problem and options are detailed in
>
>
>
>
>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
> <
>
>
>
>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>
> <
>
>
>
>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
> <
>
>
>
>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>
>
> We are also wondering if there are other methods to implement
> this
> requirement with or without changing Runtime. Very thanks for any
> feedbacks
> !
>
>
> Best,
> Yun
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>

[Yun Gao]

     Hi Xiaogang,

      Very thanks for also considering the iteration case! :) These points are really important for iteration. As a whole, we are implementing a new iteration library on top of Stream API. As a library, most of its implementation does not need to touch Runtime layer, but it really has some new requirements on the API, like the one for being able to broadcast the progressive events. To be more detail, these events indeed carry the sender's index and the downstream operators need to do alignment the events from all the upstream operators. It works very similar to watermark, thus these events do not need to be contained in checkpoints.

Some other points are also under implementation. However, since some part of the design is still under discussion internally, we may not be able to start a new discussion on iteration immediately. Besides, we should also need to fix the problems that may have new requirements on the Runtime, like broadcasting events, to have a complete design. Therefore, I think we may still first have the broadcasting problem settled in this thread? Based on the points learned in the discussion, now I think that we might be able to decouple the broadcasting events requirements and more generalized multicasting mechanism. :)

Best,
Yun



------------------------------------------------------------------
From:SHI Xiaogang <[hidden email]>
Send Time:2019 Aug. 27 (Tue.) 09:16
To:dev <[hidden email]>; Yun Gao <[hidden email]>
Cc:Piotr Nowojski <[hidden email]>
Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern

Hi, Yun Gao

The discussion seems to move in a different direction, changing from supporting multicasting to implementing new iteration libraries on data streams.

Regarding the broadcast events in iterations, many details of new iteration libraries are unclear,
1. How the iteration progress is determined and notified? The iterations are synchronous or asynchronous? As far as i know, progress tracking for asynchronous iterations is very difficult.
2. Do async I/O operators allowed in the iterations? If so, how the broadcast events are checkpointed and restored? How broadcast events are distributed when the degree of parallelism changes?
3. Do the emitted broadcast events carry the sender's index? Will they be aligned in a similar way to checkpoint barriers in downstream operators?
4. In the case of synchronous iterations, do we need something similar to barrier buffers to guarantee the correctness of iterations?
5. Will checkpointing be enabled in iterations? If checkpointing is enabled, how will checkpoint barriers interact with broadcast events?

I think a detailed design document for iterations will help understand these problems, hencing improving the discussion.

I also suggest a new thread for the discussion on iterations.
This thread should focus on multicasting and discuss those problems related to multicasting, including how data is delivered and states are partitioned.

Regards,
Xiaogang
Yun Gao <[hidden email]> 于2019年8月26日周一 下午11:35写道：

 Hi,

 Very thanks for all the points raised !

 @Piotr For using another edge to broadcast the event, I think it may not be able to address the iteration case. The primary problem is that with  two edges we cannot ensure the order of records. However, In the iteration case, the broadcasted event is used to mark the progress of the iteration and it works like watermark, thus its position relative to the normal records can not change.
 And @Piotr, @Xiaogang, for the requirements on the state, I think different options seems vary. The first option is to allow Operator<T> to broadcast a separate event and have a separate process method for this event. To be detail, we may add a new type of StreamElement called Event and allow Operator<T> to broadcastEmit Event. Then in the received side, we could add a new `processEvent` method to the (Keyed)ProcessFunction. Similar to the broadcast side of KeyedBroadcastProcessFunction, in this new method users cannot access keyed state with specific key, but can register a state function to touch all the elements in the keyed state. This option needs to modify the runtime to support the new type of StreamElement, but it does not affect the semantics of states and thus it has no requirements on state.
 The second option is to allow Operator<T> to broadcastEmit T and in the receiver side, user can process the broadcast element with the existing process method. This option is consistent with the OperatorState, but for keyedState we may send a record to tasks that do not containing the corresponding keyed state, thus it should require some changes on the State.
 The third option is to support the generic Multicast. For keyedState it also meets the problem of inconsistency between network partitioner and keyed state partitioner, and if we want to rely on it to implement the non-key join, it should be also meet the problem of cannot control the partitioning of operator state. Therefore, it should also require some changes on the State.
 Then for the different scenarios proposed, the iteration case in fact requires exactly the ability to broadcast a different event type. In the iteration the fields of the progress event are in fact different from that of normal records. It does not contain actual value but contains some fields for the downstream operators to align the events and track the progress. Therefore, broadcasting a different event type is able to solve the iteration case without the requirements on the state. Besides, allowing the operator to broadcast a separate event may also facilitate some other user cases, for example, users may notify the downstream operators to change logic if some patterns are matched. The notification might be different from the normal records and users do not need to uniform them with a wrapper type manually if the operators are able to broadcast a separate event. However, it truly cannot address the non-key join scenarios.
 Since allowing broadcasting a separate event seems to be able to serve as a standalone functionality, and it does not require change on the state, I am thinking that is it possible for us to partition to multiple steps and supports broadcasting events first ? At the same time we could also continue working on other options to support more scenarios like non-key join and they seems to requires more thoughts.

 Best,
 Yun



 ------------------------------------------------------------------
 From:Piotr Nowojski <[hidden email]>
 Send Time:2019 Aug. 26 (Mon.) 18:59
 To:dev <[hidden email]>
 Cc:Yun Gao <[hidden email]>
 Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern

 Hi,

 Xiaogang, those things worry me the most.
 1. Regarding the broadcasting, doesn’t the BroadcastState [1] cover our issues? Can not we construct a job graph, where one operator has two outputs, one keyed another broadcasted, which are wired together back to the KeyedBroadcastProcessFunction or BroadcastProcessFunction?

 2. Multicast on keyed streams, might be done by iterating over all of the keys. However I have a feeling that might not be the feature which distributed cross/theta joins would want, since they would probably need a guarantee to have only a single key per operator instance.

 Kurt, by broadcast optimisation do you mean [2]?

 I’m not sure if we should split the discussion yet. Most of the changes required by either multicast or broadcast will be in the API/state layers. Runtime changes for broadcast would be almost none (just exposing existing features) and for multicast they shouldn't be huge as well. However maybe we should consider those two things together at the API level, so that we do not make wrong decisions when just looking at the simpler/more narrow broadcast support?

 Piotrek

 [1] https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/state/broadcast_state.html
 [2] https://github.com/apache/flink/pull/7713


 On 26 Aug 2019, at 09:35, Kurt Young <[hidden email]> wrote:
 From SQL's perspective, distributed cross join is a valid feature but not
 very
 urgent. Actually this discuss reminds me about another useful feature
 (sorry
 for the distraction):

 when doing broadcast in batch shuffle mode, we can make each producer only
 write one copy of the output data, but not for every consumer. Broadcast
 join
 is much more useful, and this is a very important optimization. Not sure if
 we
 have already consider this.

 Best,
 Kurt


 On Mon, Aug 26, 2019 at 12:16 PM Guowei Ma <[hidden email]> wrote:

 Thanks Yun for bringing up this discussion and very thanks for all the deep
 thoughts!

 For now, I think this discussion contains two scenarios: one if for
 iteration library support and the other is for SQL join support. I think
 both of the two scenarios are useful but they seem to have different best
 suitable solutions. For making the discussion more clear, I would suggest
 to split the discussion into two threads.

 And I agree with Piotr that it is very tricky that a keyed stream received
 a "broadcast element". So we may add some new interfaces, which could
 broadcast or process some special "broadcast event". In that way "broadcast
 event" will not be sent with the normal process.

 Best,
 Guowei


 SHI Xiaogang <[hidden email]> 于2019年8月26日周一 上午9:27写道：

 Hi all,

 I also think that multicasting is a necessity in Flink, but more details
 are needed to be considered.

 Currently network is tightly coupled with states in Flink to achieve
 automatic scaling. We can only access keyed states in keyed streams and
 operator states in all streams.
 In the concrete example of theta-joins implemented with mutlticasting,
 the
 following questions exist:

    - In which type of states will the data be stored? Do we need another
    type of states which is coupled with multicasting streams?
    - How to ensure the consistency between network and states when jobs
    scale out or scale in?

 Regards,
 Xiaogang

 Xingcan Cui <[hidden email]> 于2019年8月25日周日 上午10:03写道：

 Hi all,

 Sorry for joining this thread late. Basically, I think enabling
 multicast
 pattern could be the right direction, but more detailed implementation
 policies need to be discussed.

 Two years ago, I filed an issue [1] about the multicast API. However,
 due
 to some reasons, it was laid aside. After that, when I tried to
 cherry-pick
 the change for experimental use, I found the return type of
 `selectChannels()` method had changed from `int[]` to `int`, which
 makes
 the old implementation not work anymore.

 From my side, the multicast has always been used for theta-join. As far
 as
 I know, it’s an essential requirement for some sophisticated joining
 algorithms. Until now, the Flink non-equi joins can still only be
 executed
 single-threaded. If we'd like to make some improvements on this, we
 should
 first take some measures to support multicast pattern.

 Best,
 Xingcan

 [1] https://issues.apache.org/jira/browse/FLINK-6936

 On Aug 24, 2019, at 5:54 AM, Zhu Zhu <[hidden email]> wrote:

 Hi Piotr,

 Thanks for the explanation.
 Agreed that the broadcastEmit(record) is a better choice for
 broadcasting
 for the iterations.
 As broadcasting for the iterations is the first motivation, let's
 support
 it first.

 Thanks,
 Zhu Zhu

 Yun Gao <[hidden email]> 于2019年8月23日周五 下午11:56写道：

     Hi Piotr,

      Very thanks for the suggestions!

     Totally agree with that we could first focus on the broadcast
 scenarios and exposing the broadcastEmit method first considering
 the
 semantics and performance.

     For the keyed stream, I also agree with that broadcasting keyed
 records to all the tasks may be confused considering the semantics
 of
 keyed
 partitioner. However, in the iteration case supporting broadcast
 over
 keyed
 partitioner should be required since users may create any subgraph
 for
 the
 iteration body, including the operators with key. I think a possible
 solution to this issue is to introduce another data type for
 'broadcastEmit'. For example, for an operator Operator<T>, it may
 broadcast
 emit another type E instead of T, and the transmitting E will bypass
 the
 partitioner and setting keyed context. This should result in the
 design
 to
 introduce customized operator event (option 1 in the document). The
 cost of
 this method is that we need to introduce a new type of StreamElement
 and
 new interface for this type, but it should be suitable for both
 keyed
 or
 non-keyed partitioner.

 Best,
 Yun



 ------------------------------------------------------------------
 From:Piotr Nowojski <[hidden email]>
 Send Time:2019 Aug. 23 (Fri.) 22:29
 To:Zhu Zhu <[hidden email]>
 Cc:dev <[hidden email]>; Yun Gao <[hidden email]>
 Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
 Pattern

 Hi,

 If the primary motivation is broadcasting (for the iterations) and
 we
 have
 no immediate need for multicast (cross join), I would prefer to
 first
 expose broadcast via the DataStream API and only later, once we
 finally
 need it, support multicast. As I wrote, multicast would be more
 challenging
 to implement, with more complicated runtime and API. And re-using
 multicast
 just to support broadcast doesn’t have much sense:

 1. It’s a bit obfuscated. It’s easier to understand
 collectBroadcast(record) or broadcastEmit(record) compared to some
 multicast channel selector that just happens to return all of the
 channels.
 2. There are performance benefits of explicitly calling
 `RecordWriter#broadcastEmit`.


 On a different note, what would be the semantic of such broadcast
 emit
 on
 KeyedStream? Would it be supported? Or would we limit support only
 to
 the
 non-keyed streams?

 Piotrek

 On 23 Aug 2019, at 12:48, Zhu Zhu <[hidden email]> wrote:

 Thanks Piotr,

 Users asked for this feature sometimes ago when they migrating
 batch
 jobs to Flink(Blink).
 It's not very urgent as they have taken some workarounds to solve
 it.(like partitioning data set to different job vertices)
 So it's fine to not make it top priority.

 Anyway, as a commonly known scenario, I think users can benefit
 from
 cross join sooner or later.

 Thanks,
 Zhu Zhu

 Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
 于2019年8月23日周五 下午6:19写道：
 Hi,

 Thanks for the answers :) Ok I understand the full picture now. +1
 from
 my side on solving this issue somehow. But before we start
 discussing
 how
 to solve it one last control question:

 I guess this multicast is intended to be used in blink planner,
 right?
 Assuming that we implement the multicast support now, when would it
 be
 used
 by the blink? I would like to avoid a scenario, where we implement
 an
 unused feature and we keep maintaining it for a long period of time.

 Piotrek

 PS, try to include motivating examples, including concrete ones in
 the
 proposals/design docs, for example in the very first paragraph.
 Especially
 if it’s a commonly known feature like cross join :)

 On 23 Aug 2019, at 11:38, Yun Gao <[hidden email]>
 wrote:

    Hi Piotr,

       Thanks a lot for sharing the thoughts!

       For the iteration, agree with that multicasting is not
 necessary. Exploring the broadcast interface to Output of the
 operators
 in
 some way should also solve this issue, and I think it should be even
 more
 convenient to have the broadcast method for the iteration.

       Also thanks Zhu Zhu for the cross join case!
 Best,
  Yun



 ------------------------------------------------------------------
 From:Zhu Zhu <[hidden email] <mailto:[hidden email]>>
 Send Time:2019 Aug. 23 (Fri.) 17:25
 To:dev <[hidden email] <mailto:[hidden email]>>
 Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
 Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
 Pattern

 Hi Piotr,

 Yes you are right it's a distributed cross join requirement.
 Broadcast join can help with cross join cases. But users cannot
 use
 it
 if the data set to join is too large to fit into one subtask.

 Sorry for left some details behind.

 Thanks,
 Zhu Zhu
 Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
 于2019年8月23日周五 下午4:57写道：
 Hi Yun and Zhu Zhu,

 Thanks for the more detailed example Zhu Zhu.

 As far as I understand for the iterations example we do not need
 multicasting. Regarding the Join example, I don’t fully understand
 it.
 The
 example that Zhu Zhu presented has a drawback of sending both tables
 to
 multiple nodes. What’s the benefit of using broadcast join over a
 hash
 join
 in such case? As far as I know, the biggest benefit of using
 broadcast
 join
 instead of hash join is that we can avoid sending the larger table
 over
 the
 network, because we can perform the join locally. In this example we
 are
 sending both of the tables to multiple nodes, which should defeat
 the
 purpose.

 Is it about implementing cross join or near cross joins in a
 distributed fashion?

 if we introduce a new MulticastRecordWriter

 That’s one of the solutions. It might have a drawback of 3 class
 virtualisation problem (We have RecordWriter and
 BroadcastRecordWriter
 already). With up to two implementations, JVM is able to
 devirtualise
 the
 calls.

 Previously I was also thinking about just providing two different
 ChannelSelector interfaces. One with `int[]` and
 `SingleChannelSelector`
 with plain `int` and based on that, RecordWriter could perform some
 magic
 (worst case scenario `instaceof` checks).

 Another solution might be to change `ChannelSelector` interface
 into
 an iterator.

 But let's discuss the details after we agree on implementing this.

 Piotrek

 On 23 Aug 2019, at 10:20, Yun Gao <[hidden email] <mailto:
[hidden email]>> wrote:

  Hi Piotr,

       Thanks a lot for the suggestions!

       The core motivation of this discussion is to implement a
 new
 iteration library on the DataStream, and it requires to insert
 special
 records in the stream to notify the progress of the iteration. The
 mechanism of such records is very similar to the current Watermark,
 and
 we
 meet the problem of sending normal records according to the
 partition
 (Rebalance, etc..) and also be able to broadcast the inserted
 progress
 records to all the connected records. I have read the notes in the
 google
 doc and I totally agree with that exploring the broadcast interface
 in
 RecordWriter in some way is able to solve this issue.

      Regarding to `int[] ChannelSelector#selectChannels()`, I'm
 wondering if we introduce a new MulticastRecordWriter and left the
 current
 RecordWriter untouched, could we avoid the performance degradation ?
 Since
 with such a modification the normal RecordWriter does not need to
 iterate
 the return array by ChannelSelector, and the only difference will be
 returning an array instead of an integer, and accessing the first
 element
 of the returned array instead of reading the integer directly.

 Best,
 Yun


 ------------------------------------------------------------------
 From:Piotr Nowojski <[hidden email] <mailto:
[hidden email]

 Send Time:2019 Aug. 23 (Fri.) 15:20
 To:dev <[hidden email] <mailto:[hidden email]>>
 Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
 Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
 Pattern

 Hi,

 Yun:

 Thanks for proposing the idea. I have checked the document and
 left
 couple of questions there, but it might be better to answer them
 here.

 What is the exact motivation and what problems do you want to
 solve?
 We have dropped multicast support from the network stack [1] for two
 reasons:
 1. Performance
 2. Code simplicity

 The proposal to re introduce `int[]
 ChannelSelector#selectChannels()`
 would revert those changes. At that time we were thinking about a
 way
 how
 to keep the multicast support on the network level, while keeping
 the
 performance and simplicity for non multicast cases and there are
 ways
 to
 achieve that. However they would add extra complexity to Flink,
 which
 it
 would be better to avoid.

 On the other hand, supporting dual pattern: standard partitioning
 or
 broadcasting is easy to do, as LatencyMarkers are doing exactly
 that.
 It
 would be just a matter of exposing this to the user in some way. So
 before
 we go any further, can you describe your use cases/motivation? Isn’t
 mix of
 standard partitioning and broadcasting enough? Do we need
 multicasting?

 Zhu:

 Could you rephrase your example? I didn’t quite understand it.

 Piotrek

 [1] https://issues.apache.org/jira/browse/FLINK-10662 <
https://issues.apache.org/jira/browse/FLINK-10662> <
https://issues.apache.org/jira/browse/FLINK-10662 <
https://issues.apache.org/jira/browse/FLINK-10662>>

 On 23 Aug 2019, at 09:17, Zhu Zhu <[hidden email] <mailto:
[hidden email]> <mailto:[hidden email] <mailto:
[hidden email]

 wrote:

 Thanks Yun for starting this discussion.
 I think the multicasting can be very helpful in certain cases.

 I have received requirements from users that they want to do
 broadcast
 join, while the data set to broadcast is too large to fit in one
 task.
 Thus the requirement turned out to be to support cartesian
 product
 of
 2
 data set(one of which can be infinite stream).
 For example, A(parallelism=2) broadcast join B(parallelism=2) in
 JobVertex
 C.
 The idea to is have 4 C subtasks to deal with different
 combinations
 of A/B
 partitions, like C1(A1,B1), C2(A1,B2), C3(A2,B1), C4(A2,B2).
 This requires one record to be sent to multiple downstream
 subtasks,
 but
 not to all subtasks.

 With current interface this is not supported, as one record can
 only
 be
 sent to one subtask, or to all subtasks of a JobVertex.
 And the user had to split the broadcast data set manually to
 several
 different JobVertices, which is hard to maintain and extend.

 Thanks,
 Zhu Zhu

 Yun Gao <[hidden email] <mailto:
 [hidden email] <mailto:[hidden email]

 于2019年8月22日周四 下午8:42写道：

 Hi everyone,
    In some scenarios we met a requirement that some operators
 want
 to
 send records to theirs downstream operators with an multicast
 communication
 pattern. In detail, for some records, the operators want to send
 them
 according to the partitioner (for example, Rebalance), and for
 some
 other
 records, the operators want to send them to all the connected
 operators and
 tasks. Such a communication pattern could be viewed as a kind of
 multicast:
 it does not broadcast every record, but some record will indeed
 be
 sent to
 multiple downstream operators.

 However, we found that this kind of communication pattern seems
 could
 not
 be implemented rightly if the operators have multiple consumers
 with
 different parallelism, using the customized partitioner. To solve
 the
 above
 problem, we propose to enhance the support for such kind of
 irregular
 communication pattern. We think there may be two options:

   1. Support a kind of customized operator events, which share
 much
 similarity with Watermark, and these events can be broadcasted to
 the
 downstream operators separately.
   2. Let the channel selector supports multicast, and also add
 the
 separate RecordWriter implementation to avoid impacting the
 performance of
 the channel selector that does not need multicast.

 The problem and options are detailed in




https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
 <



https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing

 <



https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
 <



https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing


 We are also wondering if there are other methods to implement
 this
 requirement with or without changing Runtime. Very thanks for any
 feedbacks
 !


 Best,
 Yun

[SHI Xiaogang]

Hi Yun Gao,

Thanks a lot for your clarification.

Now that the notification of broadcast events requires alignment whose
implementation, in my opinion, will affect the correctness of synchronous
iterations, I prefer to postpone the discussion until you have completed
the design of the new iteration library, or at least the progress tracking
part. Otherwise, the discussion for broadcasting events may become an empty
talk if it does not fit in with the final design.

What do you think?

Regards,
Xiaogang

Yun Gao <[hidden email]> 于2019年8月27日周二 上午11:33写道：

>      Hi Xiaogang,
>
>       Very thanks for also considering the iteration case! :) These points
> are really important for iteration. As a whole, we are implementing a new
> iteration library on top of Stream API. As a library, most of its
> implementation does not need to touch Runtime layer, but it really has some
> new requirements on the API, like the one for being able to broadcast the
> progressive events. To be more detail, these events indeed carry the
> sender's index and the downstream operators need to do alignment the events
> from all the upstream operators. It works very similar to watermark, thus
> these events do not need to be contained in checkpoints.
>
> Some other points are also under implementation. However, since some part
> of the design is still under discussion internally, we may not be able to
> start a new discussion on iteration immediately. Besides, we should also
> need to fix the problems that may have new requirements on the Runtime,
> like broadcasting events, to have a complete design. Therefore, I think we
> may still first have the broadcasting problem settled in this thread? Based
> on the points learned in the discussion, now I think that we might be able
> to decouple the broadcasting events requirements and more generalized
> multicasting mechanism. :)
>
> Best,
> Yun
>
>
>
> ------------------------------------------------------------------
> From:SHI Xiaogang <[hidden email]>
> Send Time:2019 Aug. 27 (Tue.) 09:16
> To:dev <[hidden email]>; Yun Gao <[hidden email]>
> Cc:Piotr Nowojski <[hidden email]>
> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern
>
> Hi, Yun Gao
>
> The discussion seems to move in a different direction, changing from
> supporting multicasting to implementing new iteration libraries on data
> streams.
>
> Regarding the broadcast events in iterations, many details of new
> iteration libraries are unclear,
> 1. How the iteration progress is determined and notified? The iterations
> are synchronous or asynchronous? As far as i know, progress tracking for
> asynchronous iterations is very difficult.
> 2. Do async I/O operators allowed in the iterations? If so, how the
> broadcast events are checkpointed and restored? How broadcast events are
> distributed when the degree of parallelism changes?
> 3. Do the emitted broadcast events carry the sender's index? Will they be
> aligned in a similar way to checkpoint barriers in downstream operators?
> 4. In the case of synchronous iterations, do we need something similar to
> barrier buffers to guarantee the correctness of iterations?
> 5. Will checkpointing be enabled in iterations? If checkpointing is
> enabled, how will checkpoint barriers interact with broadcast events?
>
> I think a detailed design document for iterations will help understand
> these problems, hencing improving the discussion.
>
> I also suggest a new thread for the discussion on iterations.
> This thread should focus on multicasting and discuss those problems
> related to multicasting, including how data is delivered and states are
> partitioned.
>
> Regards,
> Xiaogang
> Yun Gao <[hidden email]> 于2019年8月26日周一 下午11:35写道：
>
>  Hi,
>
>  Very thanks for all the points raised !
>
>  @Piotr For using another edge to broadcast the event, I think it may not
> be able to address the iteration case. The primary problem is that with
> two edges we cannot ensure the order of records. However, In the iteration
> case, the broadcasted event is used to mark the progress of the iteration
> and it works like watermark, thus its position relative to the normal
> records can not change.
>  And @Piotr, @Xiaogang, for the requirements on the state, I think
> different options seems vary. The first option is to allow Operator<T> to
> broadcast a separate event and have a separate process method for this
> event. To be detail, we may add a new type of StreamElement called Event
> and allow Operator<T> to broadcastEmit Event. Then in the received side, we
> could add a new `processEvent` method to the (Keyed)ProcessFunction.
> Similar to the broadcast side of KeyedBroadcastProcessFunction, in this new
> method users cannot access keyed state with specific key, but can register
> a state function to touch all the elements in the keyed state. This option
> needs to modify the runtime to support the new type of StreamElement, but
> it does not affect the semantics of states and thus it has no requirements
> on state.
>  The second option is to allow Operator<T> to broadcastEmit T and in the
> receiver side, user can process the broadcast element with the existing
> process method. This option is consistent with the OperatorState, but for
> keyedState we may send a record to tasks that do not containing the
> corresponding keyed state, thus it should require some changes on the State.
>  The third option is to support the generic Multicast. For keyedState it
> also meets the problem of inconsistency between network partitioner and
> keyed state partitioner, and if we want to rely on it to implement the
> non-key join, it should be also meet the problem of cannot control the
> partitioning of operator state. Therefore, it should also require some
> changes on the State.
>  Then for the different scenarios proposed, the iteration case in fact
> requires exactly the ability to broadcast a different event type. In the
> iteration the fields of the progress event are in fact different from that
> of normal records. It does not contain actual value but contains some
> fields for the downstream operators to align the events and track the
> progress. Therefore, broadcasting a different event type is able to solve
> the iteration case without the requirements on the state. Besides, allowing
> the operator to broadcast a separate event may also facilitate some other
> user cases, for example, users may notify the downstream operators to
> change logic if some patterns are matched. The notification might be
> different from the normal records and users do not need to uniform them
> with a wrapper type manually if the operators are able to broadcast a
> separate event. However, it truly cannot address the non-key join
> scenarios.
>  Since allowing broadcasting a separate event seems to be able to serve as
> a standalone functionality, and it does not require change on the state, I
> am thinking that is it possible for us to partition to multiple steps and
> supports broadcasting events first ? At the same time we could also
> continue working on other options to support more scenarios like non-key
> join and they seems to requires more thoughts.
>
>  Best,
>  Yun
>
>
>
>  ------------------------------------------------------------------
>  From:Piotr Nowojski <[hidden email]>
>  Send Time:2019 Aug. 26 (Mon.) 18:59
>  To:dev <[hidden email]>
>  Cc:Yun Gao <[hidden email]>
>  Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern
>
>  Hi,
>
>  Xiaogang, those things worry me the most.
>  1. Regarding the broadcasting, doesn’t the BroadcastState [1] cover our
> issues? Can not we construct a job graph, where one operator has two
> outputs, one keyed another broadcasted, which are wired together back to
> the KeyedBroadcastProcessFunction or BroadcastProcessFunction?
>
>  2. Multicast on keyed streams, might be done by iterating over all of the
> keys. However I have a feeling that might not be the feature which
> distributed cross/theta joins would want, since they would probably need a
> guarantee to have only a single key per operator instance.
>
>  Kurt, by broadcast optimisation do you mean [2]?
>
>  I’m not sure if we should split the discussion yet. Most of the changes
> required by either multicast or broadcast will be in the API/state layers.
> Runtime changes for broadcast would be almost none (just exposing existing
> features) and for multicast they shouldn't be huge as well. However maybe
> we should consider those two things together at the API level, so that we
> do not make wrong decisions when just looking at the simpler/more narrow
> broadcast support?
>
>  Piotrek
>
>  [1]
> https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/state/broadcast_state.html
>  [2] https://github.com/apache/flink/pull/7713
>
>
>  On 26 Aug 2019, at 09:35, Kurt Young <[hidden email]> wrote:
>  From SQL's perspective, distributed cross join is a valid feature but not
>  very
>  urgent. Actually this discuss reminds me about another useful feature
>  (sorry
>  for the distraction):
>
>  when doing broadcast in batch shuffle mode, we can make each producer only
>  write one copy of the output data, but not for every consumer. Broadcast
>  join
>  is much more useful, and this is a very important optimization. Not sure
> if
>  we
>  have already consider this.
>
>  Best,
>  Kurt
>
>
>  On Mon, Aug 26, 2019 at 12:16 PM Guowei Ma <[hidden email]> wrote:
>
>  Thanks Yun for bringing up this discussion and very thanks for all the
> deep
>  thoughts!
>
>  For now, I think this discussion contains two scenarios: one if for
>  iteration library support and the other is for SQL join support. I think
>  both of the two scenarios are useful but they seem to have different best
>  suitable solutions. For making the discussion more clear, I would suggest
>  to split the discussion into two threads.
>
>  And I agree with Piotr that it is very tricky that a keyed stream received
>  a "broadcast element". So we may add some new interfaces, which could
>  broadcast or process some special "broadcast event". In that way
> "broadcast
>  event" will not be sent with the normal process.
>
>  Best,
>  Guowei
>
>
>  SHI Xiaogang <[hidden email]> 于2019年8月26日周一 上午9:27写道：
>
>  Hi all,
>
>  I also think that multicasting is a necessity in Flink, but more details
>  are needed to be considered.
>
>  Currently network is tightly coupled with states in Flink to achieve
>  automatic scaling. We can only access keyed states in keyed streams and
>  operator states in all streams.
>  In the concrete example of theta-joins implemented with mutlticasting,
>  the
>  following questions exist:
>
>     - In which type of states will the data be stored? Do we need another
>     type of states which is coupled with multicasting streams?
>     - How to ensure the consistency between network and states when jobs
>     scale out or scale in?
>
>  Regards,
>  Xiaogang
>
>  Xingcan Cui <[hidden email]> 于2019年8月25日周日 上午10:03写道：
>
>  Hi all,
>
>  Sorry for joining this thread late. Basically, I think enabling
>  multicast
>  pattern could be the right direction, but more detailed implementation
>  policies need to be discussed.
>
>  Two years ago, I filed an issue [1] about the multicast API. However,
>  due
>  to some reasons, it was laid aside. After that, when I tried to
>  cherry-pick
>  the change for experimental use, I found the return type of
>  `selectChannels()` method had changed from `int[]` to `int`, which
>  makes
>  the old implementation not work anymore.
>
>  From my side, the multicast has always been used for theta-join. As far
>  as
>  I know, it’s an essential requirement for some sophisticated joining
>  algorithms. Until now, the Flink non-equi joins can still only be
>  executed
>  single-threaded. If we'd like to make some improvements on this, we
>  should
>  first take some measures to support multicast pattern.
>
>  Best,
>  Xingcan
>
>  [1] https://issues.apache.org/jira/browse/FLINK-6936
>
>  On Aug 24, 2019, at 5:54 AM, Zhu Zhu <[hidden email]> wrote:
>
>  Hi Piotr,
>
>  Thanks for the explanation.
>  Agreed that the broadcastEmit(record) is a better choice for
>  broadcasting
>  for the iterations.
>  As broadcasting for the iterations is the first motivation, let's
>  support
>  it first.
>
>  Thanks,
>  Zhu Zhu
>
>  Yun Gao <[hidden email]> 于2019年8月23日周五 下午11:56写道：
>
>      Hi Piotr,
>
>       Very thanks for the suggestions!
>
>      Totally agree with that we could first focus on the broadcast
>  scenarios and exposing the broadcastEmit method first considering
>  the
>  semantics and performance.
>
>      For the keyed stream, I also agree with that broadcasting keyed
>  records to all the tasks may be confused considering the semantics
>  of
>  keyed
>  partitioner. However, in the iteration case supporting broadcast
>  over
>  keyed
>  partitioner should be required since users may create any subgraph
>  for
>  the
>  iteration body, including the operators with key. I think a possible
>  solution to this issue is to introduce another data type for
>  'broadcastEmit'. For example, for an operator Operator<T>, it may
>  broadcast
>  emit another type E instead of T, and the transmitting E will bypass
>  the
>  partitioner and setting keyed context. This should result in the
>  design
>  to
>  introduce customized operator event (option 1 in the document). The
>  cost of
>  this method is that we need to introduce a new type of StreamElement
>  and
>  new interface for this type, but it should be suitable for both
>  keyed
>  or
>  non-keyed partitioner.
>
>  Best,
>  Yun
>
>
>
>  ------------------------------------------------------------------
>  From:Piotr Nowojski <[hidden email]>
>  Send Time:2019 Aug. 23 (Fri.) 22:29
>  To:Zhu Zhu <[hidden email]>
>  Cc:dev <[hidden email]>; Yun Gao <[hidden email]>
>  Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
>  Pattern
>
>  Hi,
>
>  If the primary motivation is broadcasting (for the iterations) and
>  we
>  have
>  no immediate need for multicast (cross join), I would prefer to
>  first
>  expose broadcast via the DataStream API and only later, once we
>  finally
>  need it, support multicast. As I wrote, multicast would be more
>  challenging
>  to implement, with more complicated runtime and API. And re-using
>  multicast
>  just to support broadcast doesn’t have much sense:
>
>  1. It’s a bit obfuscated. It’s easier to understand
>  collectBroadcast(record) or broadcastEmit(record) compared to some
>  multicast channel selector that just happens to return all of the
>  channels.
>  2. There are performance benefits of explicitly calling
>  `RecordWriter#broadcastEmit`.
>
>
>  On a different note, what would be the semantic of such broadcast
>  emit
>  on
>  KeyedStream? Would it be supported? Or would we limit support only
>  to
>  the
>  non-keyed streams?
>
>  Piotrek
>
>  On 23 Aug 2019, at 12:48, Zhu Zhu <[hidden email]> wrote:
>
>  Thanks Piotr,
>
>  Users asked for this feature sometimes ago when they migrating
>  batch
>  jobs to Flink(Blink).
>  It's not very urgent as they have taken some workarounds to solve
>  it.(like partitioning data set to different job vertices)
>  So it's fine to not make it top priority.
>
>  Anyway, as a commonly known scenario, I think users can benefit
>  from
>  cross join sooner or later.
>
>  Thanks,
>  Zhu Zhu
>
>  Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
>  于2019年8月23日周五 下午6:19写道：
>  Hi,
>
>  Thanks for the answers :) Ok I understand the full picture now. +1
>  from
>  my side on solving this issue somehow. But before we start
>  discussing
>  how
>  to solve it one last control question:
>
>  I guess this multicast is intended to be used in blink planner,
>  right?
>  Assuming that we implement the multicast support now, when would it
>  be
>  used
>  by the blink? I would like to avoid a scenario, where we implement
>  an
>  unused feature and we keep maintaining it for a long period of time.
>
>  Piotrek
>
>  PS, try to include motivating examples, including concrete ones in
>  the
>  proposals/design docs, for example in the very first paragraph.
>  Especially
>  if it’s a commonly known feature like cross join :)
>
>  On 23 Aug 2019, at 11:38, Yun Gao <[hidden email]>
>  wrote:
>
>     Hi Piotr,
>
>        Thanks a lot for sharing the thoughts!
>
>        For the iteration, agree with that multicasting is not
>  necessary. Exploring the broadcast interface to Output of the
>  operators
>  in
>  some way should also solve this issue, and I think it should be even
>  more
>  convenient to have the broadcast method for the iteration.
>
>        Also thanks Zhu Zhu for the cross join case!
>  Best,
>   Yun
>
>
>
>  ------------------------------------------------------------------
>  From:Zhu Zhu <[hidden email] <mailto:[hidden email]>>
>  Send Time:2019 Aug. 23 (Fri.) 17:25
>  To:dev <[hidden email] <mailto:[hidden email]>>
>  Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
>  Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
>  Pattern
>
>  Hi Piotr,
>
>  Yes you are right it's a distributed cross join requirement.
>  Broadcast join can help with cross join cases. But users cannot
>  use
>  it
>  if the data set to join is too large to fit into one subtask.
>
>  Sorry for left some details behind.
>
>  Thanks,
>  Zhu Zhu
>  Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
>  于2019年8月23日周五 下午4:57写道：
>  Hi Yun and Zhu Zhu,
>
>  Thanks for the more detailed example Zhu Zhu.
>
>  As far as I understand for the iterations example we do not need
>  multicasting. Regarding the Join example, I don’t fully understand
>  it.
>  The
>  example that Zhu Zhu presented has a drawback of sending both tables
>  to
>  multiple nodes. What’s the benefit of using broadcast join over a
>  hash
>  join
>  in such case? As far as I know, the biggest benefit of using
>  broadcast
>  join
>  instead of hash join is that we can avoid sending the larger table
>  over
>  the
>  network, because we can perform the join locally. In this example we
>  are
>  sending both of the tables to multiple nodes, which should defeat
>  the
>  purpose.
>
>  Is it about implementing cross join or near cross joins in a
>  distributed fashion?
>
>  if we introduce a new MulticastRecordWriter
>
>  That’s one of the solutions. It might have a drawback of 3 class
>  virtualisation problem (We have RecordWriter and
>  BroadcastRecordWriter
>  already). With up to two implementations, JVM is able to
>  devirtualise
>  the
>  calls.
>
>  Previously I was also thinking about just providing two different
>  ChannelSelector interfaces. One with `int[]` and
>  `SingleChannelSelector`
>  with plain `int` and based on that, RecordWriter could perform some
>  magic
>  (worst case scenario `instaceof` checks).
>
>  Another solution might be to change `ChannelSelector` interface
>  into
>  an iterator.
>
>  But let's discuss the details after we agree on implementing this.
>
>  Piotrek
>
>  On 23 Aug 2019, at 10:20, Yun Gao <[hidden email] <mailto:
> [hidden email]>> wrote:
>
>   Hi Piotr,
>
>        Thanks a lot for the suggestions!
>
>        The core motivation of this discussion is to implement a
>  new
>  iteration library on the DataStream, and it requires to insert
>  special
>  records in the stream to notify the progress of the iteration. The
>  mechanism of such records is very similar to the current Watermark,
>  and
>  we
>  meet the problem of sending normal records according to the
>  partition
>  (Rebalance, etc..) and also be able to broadcast the inserted
>  progress
>  records to all the connected records. I have read the notes in the
>  google
>  doc and I totally agree with that exploring the broadcast interface
>  in
>  RecordWriter in some way is able to solve this issue.
>
>       Regarding to `int[] ChannelSelector#selectChannels()`, I'm
>  wondering if we introduce a new MulticastRecordWriter and left the
>  current
>  RecordWriter untouched, could we avoid the performance degradation ?
>  Since
>  with such a modification the normal RecordWriter does not need to
>  iterate
>  the return array by ChannelSelector, and the only difference will be
>  returning an array instead of an integer, and accessing the first
>  element
>  of the returned array instead of reading the integer directly.
>
>  Best,
>  Yun
>
>
>  ------------------------------------------------------------------
>  From:Piotr Nowojski <[hidden email] <mailto:
> [hidden email]
>
>  Send Time:2019 Aug. 23 (Fri.) 15:20
>  To:dev <[hidden email] <mailto:[hidden email]>>
>  Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
>  Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
>  Pattern
>
>  Hi,
>
>  Yun:
>
>  Thanks for proposing the idea. I have checked the document and
>  left
>  couple of questions there, but it might be better to answer them
>  here.
>
>  What is the exact motivation and what problems do you want to
>  solve?
>  We have dropped multicast support from the network stack [1] for two
>  reasons:
>  1. Performance
>  2. Code simplicity
>
>  The proposal to re introduce `int[]
>  ChannelSelector#selectChannels()`
>  would revert those changes. At that time we were thinking about a
>  way
>  how
>  to keep the multicast support on the network level, while keeping
>  the
>  performance and simplicity for non multicast cases and there are
>  ways
>  to
>  achieve that. However they would add extra complexity to Flink,
>  which
>  it
>  would be better to avoid.
>
>  On the other hand, supporting dual pattern: standard partitioning
>  or
>  broadcasting is easy to do, as LatencyMarkers are doing exactly
>  that.
>  It
>  would be just a matter of exposing this to the user in some way. So
>  before
>  we go any further, can you describe your use cases/motivation? Isn’t
>  mix of
>  standard partitioning and broadcasting enough? Do we need
>  multicasting?
>
>  Zhu:
>
>  Could you rephrase your example? I didn’t quite understand it.
>
>  Piotrek
>
>  [1] https://issues.apache.org/jira/browse/FLINK-10662 <
> https://issues.apache.org/jira/browse/FLINK-10662> <
> https://issues.apache.org/jira/browse/FLINK-10662 <
> https://issues.apache.org/jira/browse/FLINK-10662>>
>
>  On 23 Aug 2019, at 09:17, Zhu Zhu <[hidden email] <mailto:
> [hidden email]> <mailto:[hidden email] <mailto:
> [hidden email]
>
>  wrote:
>
>  Thanks Yun for starting this discussion.
>  I think the multicasting can be very helpful in certain cases.
>
>  I have received requirements from users that they want to do
>  broadcast
>  join, while the data set to broadcast is too large to fit in one
>  task.
>  Thus the requirement turned out to be to support cartesian
>  product
>  of
>  2
>  data set(one of which can be infinite stream).
>  For example, A(parallelism=2) broadcast join B(parallelism=2) in
>  JobVertex
>  C.
>  The idea to is have 4 C subtasks to deal with different
>  combinations
>  of A/B
>  partitions, like C1(A1,B1), C2(A1,B2), C3(A2,B1), C4(A2,B2).
>  This requires one record to be sent to multiple downstream
>  subtasks,
>  but
>  not to all subtasks.
>
>  With current interface this is not supported, as one record can
>  only
>  be
>  sent to one subtask, or to all subtasks of a JobVertex.
>  And the user had to split the broadcast data set manually to
>  several
>  different JobVertices, which is hard to maintain and extend.
>
>  Thanks,
>  Zhu Zhu
>
>  Yun Gao <[hidden email] <mailto:
>  [hidden email] <mailto:[hidden email]
>
>  于2019年8月22日周四 下午8:42写道：
>
>  Hi everyone,
>     In some scenarios we met a requirement that some operators
>  want
>  to
>  send records to theirs downstream operators with an multicast
>  communication
>  pattern. In detail, for some records, the operators want to send
>  them
>  according to the partitioner (for example, Rebalance), and for
>  some
>  other
>  records, the operators want to send them to all the connected
>  operators and
>  tasks. Such a communication pattern could be viewed as a kind of
>  multicast:
>  it does not broadcast every record, but some record will indeed
>  be
>  sent to
>  multiple downstream operators.
>
>  However, we found that this kind of communication pattern seems
>  could
>  not
>  be implemented rightly if the operators have multiple consumers
>  with
>  different parallelism, using the customized partitioner. To solve
>  the
>  above
>  problem, we propose to enhance the support for such kind of
>  irregular
>  communication pattern. We think there may be two options:
>
>    1. Support a kind of customized operator events, which share
>  much
>  similarity with Watermark, and these events can be broadcasted to
>  the
>  downstream operators separately.
>    2. Let the channel selector supports multicast, and also add
>  the
>  separate RecordWriter implementation to avoid impacting the
>  performance of
>  the channel selector that does not need multicast.
>
>  The problem and options are detailed in
>
>
>
>
>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>  <
>
>
>
>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>
>  <
>
>
>
>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>  <
>
>
>
>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>
>
>  We are also wondering if there are other methods to implement
>  this
>  requirement with or without changing Runtime. Very thanks for any
>  feedbacks
>  !
>
>
>  Best,
>  Yun
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>

[Piotr Nowojski-3]

Hi,

Before starting a work on the design doc, I would suggest to find someone to shepherd this project. Otherwise this effort might drown among other parallel things. I could take care of that from the runtime perspective, however most of the changes are about the API and changes, which are outside of my area of expertise.

Regarding the multicast, before we start working on that, I would also prefer to see a motivation design doc, how that feature would be used for example for cross or theta joins in the Table API, since very similar questions would apply to that as well.

Piotrek

> On 27 Aug 2019, at 08:10, SHI Xiaogang <[hidden email]> wrote:
>
> Hi Yun Gao,
>
> Thanks a lot for your clarification.
>
> Now that the notification of broadcast events requires alignment whose
> implementation, in my opinion, will affect the correctness of synchronous
> iterations, I prefer to postpone the discussion until you have completed
> the design of the new iteration library, or at least the progress tracking
> part. Otherwise, the discussion for broadcasting events may become an empty
> talk if it does not fit in with the final design.
>
> What do you think?
>
> Regards,
> Xiaogang
>
> Yun Gao <[hidden email]> 于2019年8月27日周二 上午11:33写道：
>
>>     Hi Xiaogang,
>>
>>      Very thanks for also considering the iteration case! :) These points
>> are really important for iteration. As a whole, we are implementing a new
>> iteration library on top of Stream API. As a library, most of its
>> implementation does not need to touch Runtime layer, but it really has some
>> new requirements on the API, like the one for being able to broadcast the
>> progressive events. To be more detail, these events indeed carry the
>> sender's index and the downstream operators need to do alignment the events
>> from all the upstream operators. It works very similar to watermark, thus
>> these events do not need to be contained in checkpoints.
>>
>> Some other points are also under implementation. However, since some part
>> of the design is still under discussion internally, we may not be able to
>> start a new discussion on iteration immediately. Besides, we should also
>> need to fix the problems that may have new requirements on the Runtime,
>> like broadcasting events, to have a complete design. Therefore, I think we
>> may still first have the broadcasting problem settled in this thread? Based
>> on the points learned in the discussion, now I think that we might be able
>> to decouple the broadcasting events requirements and more generalized
>> multicasting mechanism. :)
>>
>> Best,
>> Yun
>>
>>
>>
>> ------------------------------------------------------------------
>> From:SHI Xiaogang <[hidden email]>
>> Send Time:2019 Aug. 27 (Tue.) 09:16
>> To:dev <[hidden email]>; Yun Gao <[hidden email]>
>> Cc:Piotr Nowojski <[hidden email]>
>> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern
>>
>> Hi, Yun Gao
>>
>> The discussion seems to move in a different direction, changing from
>> supporting multicasting to implementing new iteration libraries on data
>> streams.
>>
>> Regarding the broadcast events in iterations, many details of new
>> iteration libraries are unclear,
>> 1. How the iteration progress is determined and notified? The iterations
>> are synchronous or asynchronous? As far as i know, progress tracking for
>> asynchronous iterations is very difficult.
>> 2. Do async I/O operators allowed in the iterations? If so, how the
>> broadcast events are checkpointed and restored? How broadcast events are
>> distributed when the degree of parallelism changes?
>> 3. Do the emitted broadcast events carry the sender's index? Will they be
>> aligned in a similar way to checkpoint barriers in downstream operators?
>> 4. In the case of synchronous iterations, do we need something similar to
>> barrier buffers to guarantee the correctness of iterations?
>> 5. Will checkpointing be enabled in iterations? If checkpointing is
>> enabled, how will checkpoint barriers interact with broadcast events?
>>
>> I think a detailed design document for iterations will help understand
>> these problems, hencing improving the discussion.
>>
>> I also suggest a new thread for the discussion on iterations.
>> This thread should focus on multicasting and discuss those problems
>> related to multicasting, including how data is delivered and states are
>> partitioned.
>>
>> Regards,
>> Xiaogang
>> Yun Gao <[hidden email]> 于2019年8月26日周一 下午11:35写道：
>>
>> Hi,
>>
>> Very thanks for all the points raised !
>>
>> @Piotr For using another edge to broadcast the event, I think it may not
>> be able to address the iteration case. The primary problem is that with
>> two edges we cannot ensure the order of records. However, In the iteration
>> case, the broadcasted event is used to mark the progress of the iteration
>> and it works like watermark, thus its position relative to the normal
>> records can not change.
>> And @Piotr, @Xiaogang, for the requirements on the state, I think
>> different options seems vary. The first option is to allow Operator<T> to
>> broadcast a separate event and have a separate process method for this
>> event. To be detail, we may add a new type of StreamElement called Event
>> and allow Operator<T> to broadcastEmit Event. Then in the received side, we
>> could add a new `processEvent` method to the (Keyed)ProcessFunction.
>> Similar to the broadcast side of KeyedBroadcastProcessFunction, in this new
>> method users cannot access keyed state with specific key, but can register
>> a state function to touch all the elements in the keyed state. This option
>> needs to modify the runtime to support the new type of StreamElement, but
>> it does not affect the semantics of states and thus it has no requirements
>> on state.
>> The second option is to allow Operator<T> to broadcastEmit T and in the
>> receiver side, user can process the broadcast element with the existing
>> process method. This option is consistent with the OperatorState, but for
>> keyedState we may send a record to tasks that do not containing the
>> corresponding keyed state, thus it should require some changes on the State.
>> The third option is to support the generic Multicast. For keyedState it
>> also meets the problem of inconsistency between network partitioner and
>> keyed state partitioner, and if we want to rely on it to implement the
>> non-key join, it should be also meet the problem of cannot control the
>> partitioning of operator state. Therefore, it should also require some
>> changes on the State.
>> Then for the different scenarios proposed, the iteration case in fact
>> requires exactly the ability to broadcast a different event type. In the
>> iteration the fields of the progress event are in fact different from that
>> of normal records. It does not contain actual value but contains some
>> fields for the downstream operators to align the events and track the
>> progress. Therefore, broadcasting a different event type is able to solve
>> the iteration case without the requirements on the state. Besides, allowing
>> the operator to broadcast a separate event may also facilitate some other
>> user cases, for example, users may notify the downstream operators to
>> change logic if some patterns are matched. The notification might be
>> different from the normal records and users do not need to uniform them
>> with a wrapper type manually if the operators are able to broadcast a
>> separate event. However, it truly cannot address the non-key join
>> scenarios.
>> Since allowing broadcasting a separate event seems to be able to serve as
>> a standalone functionality, and it does not require change on the state, I
>> am thinking that is it possible for us to partition to multiple steps and
>> supports broadcasting events first ? At the same time we could also
>> continue working on other options to support more scenarios like non-key
>> join and they seems to requires more thoughts.
>>
>> Best,
>> Yun
>>
>>
>>
>> ------------------------------------------------------------------
>> From:Piotr Nowojski <[hidden email]>
>> Send Time:2019 Aug. 26 (Mon.) 18:59
>> To:dev <[hidden email]>
>> Cc:Yun Gao <[hidden email]>
>> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern
>>
>> Hi,
>>
>> Xiaogang, those things worry me the most.
>> 1. Regarding the broadcasting, doesn’t the BroadcastState [1] cover our
>> issues? Can not we construct a job graph, where one operator has two
>> outputs, one keyed another broadcasted, which are wired together back to
>> the KeyedBroadcastProcessFunction or BroadcastProcessFunction?
>>
>> 2. Multicast on keyed streams, might be done by iterating over all of the
>> keys. However I have a feeling that might not be the feature which
>> distributed cross/theta joins would want, since they would probably need a
>> guarantee to have only a single key per operator instance.
>>
>> Kurt, by broadcast optimisation do you mean [2]?
>>
>> I’m not sure if we should split the discussion yet. Most of the changes
>> required by either multicast or broadcast will be in the API/state layers.
>> Runtime changes for broadcast would be almost none (just exposing existing
>> features) and for multicast they shouldn't be huge as well. However maybe
>> we should consider those two things together at the API level, so that we
>> do not make wrong decisions when just looking at the simpler/more narrow
>> broadcast support?
>>
>> Piotrek
>>
>> [1]
>> https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/state/broadcast_state.html
>> [2] https://github.com/apache/flink/pull/7713
>>
>>
>> On 26 Aug 2019, at 09:35, Kurt Young <[hidden email]> wrote:
>> From SQL's perspective, distributed cross join is a valid feature but not
>> very
>> urgent. Actually this discuss reminds me about another useful feature
>> (sorry
>> for the distraction):
>>
>> when doing broadcast in batch shuffle mode, we can make each producer only
>> write one copy of the output data, but not for every consumer. Broadcast
>> join
>> is much more useful, and this is a very important optimization. Not sure
>> if
>> we
>> have already consider this.
>>
>> Best,
>> Kurt
>>
>>
>> On Mon, Aug 26, 2019 at 12:16 PM Guowei Ma <[hidden email]> wrote:
>>
>> Thanks Yun for bringing up this discussion and very thanks for all the
>> deep
>> thoughts!
>>
>> For now, I think this discussion contains two scenarios: one if for
>> iteration library support and the other is for SQL join support. I think
>> both of the two scenarios are useful but they seem to have different best
>> suitable solutions. For making the discussion more clear, I would suggest
>> to split the discussion into two threads.
>>
>> And I agree with Piotr that it is very tricky that a keyed stream received
>> a "broadcast element". So we may add some new interfaces, which could
>> broadcast or process some special "broadcast event". In that way
>> "broadcast
>> event" will not be sent with the normal process.
>>
>> Best,
>> Guowei
>>
>>
>> SHI Xiaogang <[hidden email]> 于2019年8月26日周一 上午9:27写道：
>>
>> Hi all,
>>
>> I also think that multicasting is a necessity in Flink, but more details
>> are needed to be considered.
>>
>> Currently network is tightly coupled with states in Flink to achieve
>> automatic scaling. We can only access keyed states in keyed streams and
>> operator states in all streams.
>> In the concrete example of theta-joins implemented with mutlticasting,
>> the
>> following questions exist:
>>
>>    - In which type of states will the data be stored? Do we need another
>>    type of states which is coupled with multicasting streams?
>>    - How to ensure the consistency between network and states when jobs
>>    scale out or scale in?
>>
>> Regards,
>> Xiaogang
>>
>> Xingcan Cui <[hidden email]> 于2019年8月25日周日 上午10:03写道：
>>
>> Hi all,
>>
>> Sorry for joining this thread late. Basically, I think enabling
>> multicast
>> pattern could be the right direction, but more detailed implementation
>> policies need to be discussed.
>>
>> Two years ago, I filed an issue [1] about the multicast API. However,
>> due
>> to some reasons, it was laid aside. After that, when I tried to
>> cherry-pick
>> the change for experimental use, I found the return type of
>> `selectChannels()` method had changed from `int[]` to `int`, which
>> makes
>> the old implementation not work anymore.
>>
>> From my side, the multicast has always been used for theta-join. As far
>> as
>> I know, it’s an essential requirement for some sophisticated joining
>> algorithms. Until now, the Flink non-equi joins can still only be
>> executed
>> single-threaded. If we'd like to make some improvements on this, we
>> should
>> first take some measures to support multicast pattern.
>>
>> Best,
>> Xingcan
>>
>> [1] https://issues.apache.org/jira/browse/FLINK-6936
>>
>> On Aug 24, 2019, at 5:54 AM, Zhu Zhu <[hidden email]> wrote:
>>
>> Hi Piotr,
>>
>> Thanks for the explanation.
>> Agreed that the broadcastEmit(record) is a better choice for
>> broadcasting
>> for the iterations.
>> As broadcasting for the iterations is the first motivation, let's
>> support
>> it first.
>>
>> Thanks,
>> Zhu Zhu
>>
>> Yun Gao <[hidden email]> 于2019年8月23日周五 下午11:56写道：
>>
>>     Hi Piotr,
>>
>>      Very thanks for the suggestions!
>>
>>     Totally agree with that we could first focus on the broadcast
>> scenarios and exposing the broadcastEmit method first considering
>> the
>> semantics and performance.
>>
>>     For the keyed stream, I also agree with that broadcasting keyed
>> records to all the tasks may be confused considering the semantics
>> of
>> keyed
>> partitioner. However, in the iteration case supporting broadcast
>> over
>> keyed
>> partitioner should be required since users may create any subgraph
>> for
>> the
>> iteration body, including the operators with key. I think a possible
>> solution to this issue is to introduce another data type for
>> 'broadcastEmit'. For example, for an operator Operator<T>, it may
>> broadcast
>> emit another type E instead of T, and the transmitting E will bypass
>> the
>> partitioner and setting keyed context. This should result in the
>> design
>> to
>> introduce customized operator event (option 1 in the document). The
>> cost of
>> this method is that we need to introduce a new type of StreamElement
>> and
>> new interface for this type, but it should be suitable for both
>> keyed
>> or
>> non-keyed partitioner.
>>
>> Best,
>> Yun
>>
>>
>>
>> ------------------------------------------------------------------
>> From:Piotr Nowojski <[hidden email]>
>> Send Time:2019 Aug. 23 (Fri.) 22:29
>> To:Zhu Zhu <[hidden email]>
>> Cc:dev <[hidden email]>; Yun Gao <[hidden email]>
>> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
>> Pattern
>>
>> Hi,
>>
>> If the primary motivation is broadcasting (for the iterations) and
>> we
>> have
>> no immediate need for multicast (cross join), I would prefer to
>> first
>> expose broadcast via the DataStream API and only later, once we
>> finally
>> need it, support multicast. As I wrote, multicast would be more
>> challenging
>> to implement, with more complicated runtime and API. And re-using
>> multicast
>> just to support broadcast doesn’t have much sense:
>>
>> 1. It’s a bit obfuscated. It’s easier to understand
>> collectBroadcast(record) or broadcastEmit(record) compared to some
>> multicast channel selector that just happens to return all of the
>> channels.
>> 2. There are performance benefits of explicitly calling
>> `RecordWriter#broadcastEmit`.
>>
>>
>> On a different note, what would be the semantic of such broadcast
>> emit
>> on
>> KeyedStream? Would it be supported? Or would we limit support only
>> to
>> the
>> non-keyed streams?
>>
>> Piotrek
>>
>> On 23 Aug 2019, at 12:48, Zhu Zhu <[hidden email]> wrote:
>>
>> Thanks Piotr,
>>
>> Users asked for this feature sometimes ago when they migrating
>> batch
>> jobs to Flink(Blink).
>> It's not very urgent as they have taken some workarounds to solve
>> it.(like partitioning data set to different job vertices)
>> So it's fine to not make it top priority.
>>
>> Anyway, as a commonly known scenario, I think users can benefit
>> from
>> cross join sooner or later.
>>
>> Thanks,
>> Zhu Zhu
>>
>> Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
>> 于2019年8月23日周五 下午6:19写道：
>> Hi,
>>
>> Thanks for the answers :) Ok I understand the full picture now. +1
>> from
>> my side on solving this issue somehow. But before we start
>> discussing
>> how
>> to solve it one last control question:
>>
>> I guess this multicast is intended to be used in blink planner,
>> right?
>> Assuming that we implement the multicast support now, when would it
>> be
>> used
>> by the blink? I would like to avoid a scenario, where we implement
>> an
>> unused feature and we keep maintaining it for a long period of time.
>>
>> Piotrek
>>
>> PS, try to include motivating examples, including concrete ones in
>> the
>> proposals/design docs, for example in the very first paragraph.
>> Especially
>> if it’s a commonly known feature like cross join :)
>>
>> On 23 Aug 2019, at 11:38, Yun Gao <[hidden email]>
>> wrote:
>>
>>    Hi Piotr,
>>
>>       Thanks a lot for sharing the thoughts!
>>
>>       For the iteration, agree with that multicasting is not
>> necessary. Exploring the broadcast interface to Output of the
>> operators
>> in
>> some way should also solve this issue, and I think it should be even
>> more
>> convenient to have the broadcast method for the iteration.
>>
>>       Also thanks Zhu Zhu for the cross join case!
>> Best,
>>  Yun
>>
>>
>>
>> ------------------------------------------------------------------
>> From:Zhu Zhu <[hidden email] <mailto:[hidden email]>>
>> Send Time:2019 Aug. 23 (Fri.) 17:25
>> To:dev <[hidden email] <mailto:[hidden email]>>
>> Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
>> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
>> Pattern
>>
>> Hi Piotr,
>>
>> Yes you are right it's a distributed cross join requirement.
>> Broadcast join can help with cross join cases. But users cannot
>> use
>> it
>> if the data set to join is too large to fit into one subtask.
>>
>> Sorry for left some details behind.
>>
>> Thanks,
>> Zhu Zhu
>> Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
>> 于2019年8月23日周五 下午4:57写道：
>> Hi Yun and Zhu Zhu,
>>
>> Thanks for the more detailed example Zhu Zhu.
>>
>> As far as I understand for the iterations example we do not need
>> multicasting. Regarding the Join example, I don’t fully understand
>> it.
>> The
>> example that Zhu Zhu presented has a drawback of sending both tables
>> to
>> multiple nodes. What’s the benefit of using broadcast join over a
>> hash
>> join
>> in such case? As far as I know, the biggest benefit of using
>> broadcast
>> join
>> instead of hash join is that we can avoid sending the larger table
>> over
>> the
>> network, because we can perform the join locally. In this example we
>> are
>> sending both of the tables to multiple nodes, which should defeat
>> the
>> purpose.
>>
>> Is it about implementing cross join or near cross joins in a
>> distributed fashion?
>>
>> if we introduce a new MulticastRecordWriter
>>
>> That’s one of the solutions. It might have a drawback of 3 class
>> virtualisation problem (We have RecordWriter and
>> BroadcastRecordWriter
>> already). With up to two implementations, JVM is able to
>> devirtualise
>> the
>> calls.
>>
>> Previously I was also thinking about just providing two different
>> ChannelSelector interfaces. One with `int[]` and
>> `SingleChannelSelector`
>> with plain `int` and based on that, RecordWriter could perform some
>> magic
>> (worst case scenario `instaceof` checks).
>>
>> Another solution might be to change `ChannelSelector` interface
>> into
>> an iterator.
>>
>> But let's discuss the details after we agree on implementing this.
>>
>> Piotrek
>>
>> On 23 Aug 2019, at 10:20, Yun Gao <[hidden email] <mailto:
>> [hidden email]>> wrote:
>>
>>  Hi Piotr,
>>
>>       Thanks a lot for the suggestions!
>>
>>       The core motivation of this discussion is to implement a
>> new
>> iteration library on the DataStream, and it requires to insert
>> special
>> records in the stream to notify the progress of the iteration. The
>> mechanism of such records is very similar to the current Watermark,
>> and
>> we
>> meet the problem of sending normal records according to the
>> partition
>> (Rebalance, etc..) and also be able to broadcast the inserted
>> progress
>> records to all the connected records. I have read the notes in the
>> google
>> doc and I totally agree with that exploring the broadcast interface
>> in
>> RecordWriter in some way is able to solve this issue.
>>
>>      Regarding to `int[] ChannelSelector#selectChannels()`, I'm
>> wondering if we introduce a new MulticastRecordWriter and left the
>> current
>> RecordWriter untouched, could we avoid the performance degradation ?
>> Since
>> with such a modification the normal RecordWriter does not need to
>> iterate
>> the return array by ChannelSelector, and the only difference will be
>> returning an array instead of an integer, and accessing the first
>> element
>> of the returned array instead of reading the integer directly.
>>
>> Best,
>> Yun
>>
>>
>> ------------------------------------------------------------------
>> From:Piotr Nowojski <[hidden email] <mailto:
>> [hidden email]
>>
>> Send Time:2019 Aug. 23 (Fri.) 15:20
>> To:dev <[hidden email] <mailto:[hidden email]>>
>> Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
>> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
>> Pattern
>>
>> Hi,
>>
>> Yun:
>>
>> Thanks for proposing the idea. I have checked the document and
>> left
>> couple of questions there, but it might be better to answer them
>> here.
>>
>> What is the exact motivation and what problems do you want to
>> solve?
>> We have dropped multicast support from the network stack [1] for two
>> reasons:
>> 1. Performance
>> 2. Code simplicity
>>
>> The proposal to re introduce `int[]
>> ChannelSelector#selectChannels()`
>> would revert those changes. At that time we were thinking about a
>> way
>> how
>> to keep the multicast support on the network level, while keeping
>> the
>> performance and simplicity for non multicast cases and there are
>> ways
>> to
>> achieve that. However they would add extra complexity to Flink,
>> which
>> it
>> would be better to avoid.
>>
>> On the other hand, supporting dual pattern: standard partitioning
>> or
>> broadcasting is easy to do, as LatencyMarkers are doing exactly
>> that.
>> It
>> would be just a matter of exposing this to the user in some way. So
>> before
>> we go any further, can you describe your use cases/motivation? Isn’t
>> mix of
>> standard partitioning and broadcasting enough? Do we need
>> multicasting?
>>
>> Zhu:
>>
>> Could you rephrase your example? I didn’t quite understand it.
>>
>> Piotrek
>>
>> [1] https://issues.apache.org/jira/browse/FLINK-10662 <
>> https://issues.apache.org/jira/browse/FLINK-10662> <
>> https://issues.apache.org/jira/browse/FLINK-10662 <
>> https://issues.apache.org/jira/browse/FLINK-10662>>
>>
>> On 23 Aug 2019, at 09:17, Zhu Zhu <[hidden email] <mailto:
>> [hidden email]> <mailto:[hidden email] <mailto:
>> [hidden email]
>>
>> wrote:
>>
>> Thanks Yun for starting this discussion.
>> I think the multicasting can be very helpful in certain cases.
>>
>> I have received requirements from users that they want to do
>> broadcast
>> join, while the data set to broadcast is too large to fit in one
>> task.
>> Thus the requirement turned out to be to support cartesian
>> product
>> of
>> 2
>> data set(one of which can be infinite stream).
>> For example, A(parallelism=2) broadcast join B(parallelism=2) in
>> JobVertex
>> C.
>> The idea to is have 4 C subtasks to deal with different
>> combinations
>> of A/B
>> partitions, like C1(A1,B1), C2(A1,B2), C3(A2,B1), C4(A2,B2).
>> This requires one record to be sent to multiple downstream
>> subtasks,
>> but
>> not to all subtasks.
>>
>> With current interface this is not supported, as one record can
>> only
>> be
>> sent to one subtask, or to all subtasks of a JobVertex.
>> And the user had to split the broadcast data set manually to
>> several
>> different JobVertices, which is hard to maintain and extend.
>>
>> Thanks,
>> Zhu Zhu
>>
>> Yun Gao <[hidden email] <mailto:
>> [hidden email] <mailto:[hidden email]
>>
>> 于2019年8月22日周四 下午8:42写道：
>>
>> Hi everyone,
>>    In some scenarios we met a requirement that some operators
>> want
>> to
>> send records to theirs downstream operators with an multicast
>> communication
>> pattern. In detail, for some records, the operators want to send
>> them
>> according to the partitioner (for example, Rebalance), and for
>> some
>> other
>> records, the operators want to send them to all the connected
>> operators and
>> tasks. Such a communication pattern could be viewed as a kind of
>> multicast:
>> it does not broadcast every record, but some record will indeed
>> be
>> sent to
>> multiple downstream operators.
>>
>> However, we found that this kind of communication pattern seems
>> could
>> not
>> be implemented rightly if the operators have multiple consumers
>> with
>> different parallelism, using the customized partitioner. To solve
>> the
>> above
>> problem, we propose to enhance the support for such kind of
>> irregular
>> communication pattern. We think there may be two options:
>>
>>   1. Support a kind of customized operator events, which share
>> much
>> similarity with Watermark, and these events can be broadcasted to
>> the
>> downstream operators separately.
>>   2. Let the channel selector supports multicast, and also add
>> the
>> separate RecordWriter implementation to avoid impacting the
>> performance of
>> the channel selector that does not need multicast.
>>
>> The problem and options are detailed in
>>
>>
>>
>>
>>
>> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>> <
>>
>>
>>
>>
>> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>>
>> <
>>
>>
>>
>>
>> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>> <
>>
>>
>>
>>
>> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
>>
>>
>> We are also wondering if there are other methods to implement
>> this
>> requirement with or without changing Runtime. Very thanks for any
>> feedbacks
>> !
>>
>>
>> Best,
>> Yun
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>

[Jark Wu-2]

Hi all,

Thanks Yun for bringing this topic. I missed this discussion because of the
"multicast" title.
After reading the design, if I understand correctly, it is proposing a
custom event mach mechanism, i.e. broadcasting custom event.
It is a orthogonality topic with multicasting. So I would suggest to start
a new thread to discuss about it.

Regarding to broadcasting custom event:

I would +1 for motivation, because we also encountered similar requirements
when improving Table API & SQL before.

For example, the mini-batch mechanism in blink planner will emit a special
mini-batch event to the data stream to indicate this is a start of a new
mini-batch.
The downstream aggregation operator will buffer the data records until it
receive the mini-batch event, and then process the buffer at once. This
will reduce a lot of state access.
However, we don't have a proper custom event mechanism currently, so we
leverage the watermark as the mini-batch event (which is a little hack in
my opinion).

Another case is joining a huge dimension table which is stored/produced in
hive daily. We can scan the hive table and shuffle to the JOIN operators by
the join key to join with the main stream.
Note that the dimension table is changed every day, we want to join the
latest version of the hive table. Then we need to re-scan and re-shuffle
the hive table once a new daily partition is produced.
However, we need some special events to distinguish the boundary of
different version of the dimension table. The events will be used to notify
downstream operators (mainly the JOIN operator)
 to know "ok, I will receive a new version of the dimension table", "ok, I
received the all the data of this version."

From my understanding, in order to support this feature, we might need to:
 1) expose collectEvent(CustomEvent) or broadcastEvent(CustomEvent) API to
users.
 2) support to register the serialization and deserialization of the custom
event
 3) expose processEvent(int channel, CustomEvent) API to StreamOperator


Regards,
Jark


On Tue, 27 Aug 2019 at 15:18, Piotr Nowojski <[hidden email]> wrote:

> Hi,
>
> Before starting a work on the design doc, I would suggest to find someone
> to shepherd this project. Otherwise this effort might drown among other
> parallel things. I could take care of that from the runtime perspective,
> however most of the changes are about the API and changes, which are
> outside of my area of expertise.
>
> Regarding the multicast, before we start working on that, I would also
> prefer to see a motivation design doc, how that feature would be used for
> example for cross or theta joins in the Table API, since very similar
> questions would apply to that as well.
>
> Piotrek
>
> > On 27 Aug 2019, at 08:10, SHI Xiaogang <[hidden email]> wrote:
> >
> > Hi Yun Gao,
> >
> > Thanks a lot for your clarification.
> >
> > Now that the notification of broadcast events requires alignment whose
> > implementation, in my opinion, will affect the correctness of synchronous
> > iterations, I prefer to postpone the discussion until you have completed
> > the design of the new iteration library, or at least the progress
> tracking
> > part. Otherwise, the discussion for broadcasting events may become an
> empty
> > talk if it does not fit in with the final design.
> >
> > What do you think?
> >
> > Regards,
> > Xiaogang
> >
> > Yun Gao <[hidden email]> 于2019年8月27日周二 上午11:33写道：
> >
> >>     Hi Xiaogang,
> >>
> >>      Very thanks for also considering the iteration case! :) These
> points
> >> are really important for iteration. As a whole, we are implementing a
> new
> >> iteration library on top of Stream API. As a library, most of its
> >> implementation does not need to touch Runtime layer, but it really has
> some
> >> new requirements on the API, like the one for being able to broadcast
> the
> >> progressive events. To be more detail, these events indeed carry the
> >> sender's index and the downstream operators need to do alignment the
> events
> >> from all the upstream operators. It works very similar to watermark,
> thus
> >> these events do not need to be contained in checkpoints.
> >>
> >> Some other points are also under implementation. However, since some
> part
> >> of the design is still under discussion internally, we may not be able
> to
> >> start a new discussion on iteration immediately. Besides, we should also
> >> need to fix the problems that may have new requirements on the Runtime,
> >> like broadcasting events, to have a complete design. Therefore, I think
> we
> >> may still first have the broadcasting problem settled in this thread?
> Based
> >> on the points learned in the discussion, now I think that we might be
> able
> >> to decouple the broadcasting events requirements and more generalized
> >> multicasting mechanism. :)
> >>
> >> Best,
> >> Yun
> >>
> >>
> >>
> >> ------------------------------------------------------------------
> >> From:SHI Xiaogang <[hidden email]>
> >> Send Time:2019 Aug. 27 (Tue.) 09:16
> >> To:dev <[hidden email]>; Yun Gao <[hidden email]>
> >> Cc:Piotr Nowojski <[hidden email]>
> >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
> Pattern
> >>
> >> Hi, Yun Gao
> >>
> >> The discussion seems to move in a different direction, changing from
> >> supporting multicasting to implementing new iteration libraries on data
> >> streams.
> >>
> >> Regarding the broadcast events in iterations, many details of new
> >> iteration libraries are unclear,
> >> 1. How the iteration progress is determined and notified? The iterations
> >> are synchronous or asynchronous? As far as i know, progress tracking for
> >> asynchronous iterations is very difficult.
> >> 2. Do async I/O operators allowed in the iterations? If so, how the
> >> broadcast events are checkpointed and restored? How broadcast events are
> >> distributed when the degree of parallelism changes?
> >> 3. Do the emitted broadcast events carry the sender's index? Will they
> be
> >> aligned in a similar way to checkpoint barriers in downstream operators?
> >> 4. In the case of synchronous iterations, do we need something similar
> to
> >> barrier buffers to guarantee the correctness of iterations?
> >> 5. Will checkpointing be enabled in iterations? If checkpointing is
> >> enabled, how will checkpoint barriers interact with broadcast events?
> >>
> >> I think a detailed design document for iterations will help understand
> >> these problems, hencing improving the discussion.
> >>
> >> I also suggest a new thread for the discussion on iterations.
> >> This thread should focus on multicasting and discuss those problems
> >> related to multicasting, including how data is delivered and states are
> >> partitioned.
> >>
> >> Regards,
> >> Xiaogang
> >> Yun Gao <[hidden email]> 于2019年8月26日周一 下午11:35写道：
> >>
> >> Hi,
> >>
> >> Very thanks for all the points raised !
> >>
> >> @Piotr For using another edge to broadcast the event, I think it may not
> >> be able to address the iteration case. The primary problem is that with
> >> two edges we cannot ensure the order of records. However, In the
> iteration
> >> case, the broadcasted event is used to mark the progress of the
> iteration
> >> and it works like watermark, thus its position relative to the normal
> >> records can not change.
> >> And @Piotr, @Xiaogang, for the requirements on the state, I think
> >> different options seems vary. The first option is to allow Operator<T>
> to
> >> broadcast a separate event and have a separate process method for this
> >> event. To be detail, we may add a new type of StreamElement called Event
> >> and allow Operator<T> to broadcastEmit Event. Then in the received
> side, we
> >> could add a new `processEvent` method to the (Keyed)ProcessFunction.
> >> Similar to the broadcast side of KeyedBroadcastProcessFunction, in this
> new
> >> method users cannot access keyed state with specific key, but can
> register
> >> a state function to touch all the elements in the keyed state. This
> option
> >> needs to modify the runtime to support the new type of StreamElement,
> but
> >> it does not affect the semantics of states and thus it has no
> requirements
> >> on state.
> >> The second option is to allow Operator<T> to broadcastEmit T and in the
> >> receiver side, user can process the broadcast element with the existing
> >> process method. This option is consistent with the OperatorState, but
> for
> >> keyedState we may send a record to tasks that do not containing the
> >> corresponding keyed state, thus it should require some changes on the
> State.
> >> The third option is to support the generic Multicast. For keyedState it
> >> also meets the problem of inconsistency between network partitioner and
> >> keyed state partitioner, and if we want to rely on it to implement the
> >> non-key join, it should be also meet the problem of cannot control the
> >> partitioning of operator state. Therefore, it should also require some
> >> changes on the State.
> >> Then for the different scenarios proposed, the iteration case in fact
> >> requires exactly the ability to broadcast a different event type. In the
> >> iteration the fields of the progress event are in fact different from
> that
> >> of normal records. It does not contain actual value but contains some
> >> fields for the downstream operators to align the events and track the
> >> progress. Therefore, broadcasting a different event type is able to
> solve
> >> the iteration case without the requirements on the state. Besides,
> allowing
> >> the operator to broadcast a separate event may also facilitate some
> other
> >> user cases, for example, users may notify the downstream operators to
> >> change logic if some patterns are matched. The notification might be
> >> different from the normal records and users do not need to uniform them
> >> with a wrapper type manually if the operators are able to broadcast a
> >> separate event. However, it truly cannot address the non-key join
> >> scenarios.
> >> Since allowing broadcasting a separate event seems to be able to serve
> as
> >> a standalone functionality, and it does not require change on the
> state, I
> >> am thinking that is it possible for us to partition to multiple steps
> and
> >> supports broadcasting events first ? At the same time we could also
> >> continue working on other options to support more scenarios like non-key
> >> join and they seems to requires more thoughts.
> >>
> >> Best,
> >> Yun
> >>
> >>
> >>
> >> ------------------------------------------------------------------
> >> From:Piotr Nowojski <[hidden email]>
> >> Send Time:2019 Aug. 26 (Mon.) 18:59
> >> To:dev <[hidden email]>
> >> Cc:Yun Gao <[hidden email]>
> >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
> Pattern
> >>
> >> Hi,
> >>
> >> Xiaogang, those things worry me the most.
> >> 1. Regarding the broadcasting, doesn’t the BroadcastState [1] cover our
> >> issues? Can not we construct a job graph, where one operator has two
> >> outputs, one keyed another broadcasted, which are wired together back to
> >> the KeyedBroadcastProcessFunction or BroadcastProcessFunction?
> >>
> >> 2. Multicast on keyed streams, might be done by iterating over all of
> the
> >> keys. However I have a feeling that might not be the feature which
> >> distributed cross/theta joins would want, since they would probably
> need a
> >> guarantee to have only a single key per operator instance.
> >>
> >> Kurt, by broadcast optimisation do you mean [2]?
> >>
> >> I’m not sure if we should split the discussion yet. Most of the changes
> >> required by either multicast or broadcast will be in the API/state
> layers.
> >> Runtime changes for broadcast would be almost none (just exposing
> existing
> >> features) and for multicast they shouldn't be huge as well. However
> maybe
> >> we should consider those two things together at the API level, so that
> we
> >> do not make wrong decisions when just looking at the simpler/more narrow
> >> broadcast support?
> >>
> >> Piotrek
> >>
> >> [1]
> >>
> https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/state/broadcast_state.html
> >> [2] https://github.com/apache/flink/pull/7713
> >>
> >>
> >> On 26 Aug 2019, at 09:35, Kurt Young <[hidden email]> wrote:
> >> From SQL's perspective, distributed cross join is a valid feature but
> not
> >> very
> >> urgent. Actually this discuss reminds me about another useful feature
> >> (sorry
> >> for the distraction):
> >>
> >> when doing broadcast in batch shuffle mode, we can make each producer
> only
> >> write one copy of the output data, but not for every consumer. Broadcast
> >> join
> >> is much more useful, and this is a very important optimization. Not sure
> >> if
> >> we
> >> have already consider this.
> >>
> >> Best,
> >> Kurt
> >>
> >>
> >> On Mon, Aug 26, 2019 at 12:16 PM Guowei Ma <[hidden email]>
> wrote:
> >>
> >> Thanks Yun for bringing up this discussion and very thanks for all the
> >> deep
> >> thoughts!
> >>
> >> For now, I think this discussion contains two scenarios: one if for
> >> iteration library support and the other is for SQL join support. I think
> >> both of the two scenarios are useful but they seem to have different
> best
> >> suitable solutions. For making the discussion more clear, I would
> suggest
> >> to split the discussion into two threads.
> >>
> >> And I agree with Piotr that it is very tricky that a keyed stream
> received
> >> a "broadcast element". So we may add some new interfaces, which could
> >> broadcast or process some special "broadcast event". In that way
> >> "broadcast
> >> event" will not be sent with the normal process.
> >>
> >> Best,
> >> Guowei
> >>
> >>
> >> SHI Xiaogang <[hidden email]> 于2019年8月26日周一 上午9:27写道：
> >>
> >> Hi all,
> >>
> >> I also think that multicasting is a necessity in Flink, but more details
> >> are needed to be considered.
> >>
> >> Currently network is tightly coupled with states in Flink to achieve
> >> automatic scaling. We can only access keyed states in keyed streams and
> >> operator states in all streams.
> >> In the concrete example of theta-joins implemented with mutlticasting,
> >> the
> >> following questions exist:
> >>
> >>    - In which type of states will the data be stored? Do we need another
> >>    type of states which is coupled with multicasting streams?
> >>    - How to ensure the consistency between network and states when jobs
> >>    scale out or scale in?
> >>
> >> Regards,
> >> Xiaogang
> >>
> >> Xingcan Cui <[hidden email]> 于2019年8月25日周日 上午10:03写道：
> >>
> >> Hi all,
> >>
> >> Sorry for joining this thread late. Basically, I think enabling
> >> multicast
> >> pattern could be the right direction, but more detailed implementation
> >> policies need to be discussed.
> >>
> >> Two years ago, I filed an issue [1] about the multicast API. However,
> >> due
> >> to some reasons, it was laid aside. After that, when I tried to
> >> cherry-pick
> >> the change for experimental use, I found the return type of
> >> `selectChannels()` method had changed from `int[]` to `int`, which
> >> makes
> >> the old implementation not work anymore.
> >>
> >> From my side, the multicast has always been used for theta-join. As far
> >> as
> >> I know, it’s an essential requirement for some sophisticated joining
> >> algorithms. Until now, the Flink non-equi joins can still only be
> >> executed
> >> single-threaded. If we'd like to make some improvements on this, we
> >> should
> >> first take some measures to support multicast pattern.
> >>
> >> Best,
> >> Xingcan
> >>
> >> [1] https://issues.apache.org/jira/browse/FLINK-6936
> >>
> >> On Aug 24, 2019, at 5:54 AM, Zhu Zhu <[hidden email]> wrote:
> >>
> >> Hi Piotr,
> >>
> >> Thanks for the explanation.
> >> Agreed that the broadcastEmit(record) is a better choice for
> >> broadcasting
> >> for the iterations.
> >> As broadcasting for the iterations is the first motivation, let's
> >> support
> >> it first.
> >>
> >> Thanks,
> >> Zhu Zhu
> >>
> >> Yun Gao <[hidden email]> 于2019年8月23日周五 下午11:56写道：
> >>
> >>     Hi Piotr,
> >>
> >>      Very thanks for the suggestions!
> >>
> >>     Totally agree with that we could first focus on the broadcast
> >> scenarios and exposing the broadcastEmit method first considering
> >> the
> >> semantics and performance.
> >>
> >>     For the keyed stream, I also agree with that broadcasting keyed
> >> records to all the tasks may be confused considering the semantics
> >> of
> >> keyed
> >> partitioner. However, in the iteration case supporting broadcast
> >> over
> >> keyed
> >> partitioner should be required since users may create any subgraph
> >> for
> >> the
> >> iteration body, including the operators with key. I think a possible
> >> solution to this issue is to introduce another data type for
> >> 'broadcastEmit'. For example, for an operator Operator<T>, it may
> >> broadcast
> >> emit another type E instead of T, and the transmitting E will bypass
> >> the
> >> partitioner and setting keyed context. This should result in the
> >> design
> >> to
> >> introduce customized operator event (option 1 in the document). The
> >> cost of
> >> this method is that we need to introduce a new type of StreamElement
> >> and
> >> new interface for this type, but it should be suitable for both
> >> keyed
> >> or
> >> non-keyed partitioner.
> >>
> >> Best,
> >> Yun
> >>
> >>
> >>
> >> ------------------------------------------------------------------
> >> From:Piotr Nowojski <[hidden email]>
> >> Send Time:2019 Aug. 23 (Fri.) 22:29
> >> To:Zhu Zhu <[hidden email]>
> >> Cc:dev <[hidden email]>; Yun Gao <[hidden email]>
> >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
> >> Pattern
> >>
> >> Hi,
> >>
> >> If the primary motivation is broadcasting (for the iterations) and
> >> we
> >> have
> >> no immediate need for multicast (cross join), I would prefer to
> >> first
> >> expose broadcast via the DataStream API and only later, once we
> >> finally
> >> need it, support multicast. As I wrote, multicast would be more
> >> challenging
> >> to implement, with more complicated runtime and API. And re-using
> >> multicast
> >> just to support broadcast doesn’t have much sense:
> >>
> >> 1. It’s a bit obfuscated. It’s easier to understand
> >> collectBroadcast(record) or broadcastEmit(record) compared to some
> >> multicast channel selector that just happens to return all of the
> >> channels.
> >> 2. There are performance benefits of explicitly calling
> >> `RecordWriter#broadcastEmit`.
> >>
> >>
> >> On a different note, what would be the semantic of such broadcast
> >> emit
> >> on
> >> KeyedStream? Would it be supported? Or would we limit support only
> >> to
> >> the
> >> non-keyed streams?
> >>
> >> Piotrek
> >>
> >> On 23 Aug 2019, at 12:48, Zhu Zhu <[hidden email]> wrote:
> >>
> >> Thanks Piotr,
> >>
> >> Users asked for this feature sometimes ago when they migrating
> >> batch
> >> jobs to Flink(Blink).
> >> It's not very urgent as they have taken some workarounds to solve
> >> it.(like partitioning data set to different job vertices)
> >> So it's fine to not make it top priority.
> >>
> >> Anyway, as a commonly known scenario, I think users can benefit
> >> from
> >> cross join sooner or later.
> >>
> >> Thanks,
> >> Zhu Zhu
> >>
> >> Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
> >> 于2019年8月23日周五 下午6:19写道：
> >> Hi,
> >>
> >> Thanks for the answers :) Ok I understand the full picture now. +1
> >> from
> >> my side on solving this issue somehow. But before we start
> >> discussing
> >> how
> >> to solve it one last control question:
> >>
> >> I guess this multicast is intended to be used in blink planner,
> >> right?
> >> Assuming that we implement the multicast support now, when would it
> >> be
> >> used
> >> by the blink? I would like to avoid a scenario, where we implement
> >> an
> >> unused feature and we keep maintaining it for a long period of time.
> >>
> >> Piotrek
> >>
> >> PS, try to include motivating examples, including concrete ones in
> >> the
> >> proposals/design docs, for example in the very first paragraph.
> >> Especially
> >> if it’s a commonly known feature like cross join :)
> >>
> >> On 23 Aug 2019, at 11:38, Yun Gao <[hidden email]>
> >> wrote:
> >>
> >>    Hi Piotr,
> >>
> >>       Thanks a lot for sharing the thoughts!
> >>
> >>       For the iteration, agree with that multicasting is not
> >> necessary. Exploring the broadcast interface to Output of the
> >> operators
> >> in
> >> some way should also solve this issue, and I think it should be even
> >> more
> >> convenient to have the broadcast method for the iteration.
> >>
> >>       Also thanks Zhu Zhu for the cross join case!
> >> Best,
> >>  Yun
> >>
> >>
> >>
> >> ------------------------------------------------------------------
> >> From:Zhu Zhu <[hidden email] <mailto:[hidden email]>>
> >> Send Time:2019 Aug. 23 (Fri.) 17:25
> >> To:dev <[hidden email] <mailto:[hidden email]>>
> >> Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
> >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
> >> Pattern
> >>
> >> Hi Piotr,
> >>
> >> Yes you are right it's a distributed cross join requirement.
> >> Broadcast join can help with cross join cases. But users cannot
> >> use
> >> it
> >> if the data set to join is too large to fit into one subtask.
> >>
> >> Sorry for left some details behind.
> >>
> >> Thanks,
> >> Zhu Zhu
> >> Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
> >> 于2019年8月23日周五 下午4:57写道：
> >> Hi Yun and Zhu Zhu,
> >>
> >> Thanks for the more detailed example Zhu Zhu.
> >>
> >> As far as I understand for the iterations example we do not need
> >> multicasting. Regarding the Join example, I don’t fully understand
> >> it.
> >> The
> >> example that Zhu Zhu presented has a drawback of sending both tables
> >> to
> >> multiple nodes. What’s the benefit of using broadcast join over a
> >> hash
> >> join
> >> in such case? As far as I know, the biggest benefit of using
> >> broadcast
> >> join
> >> instead of hash join is that we can avoid sending the larger table
> >> over
> >> the
> >> network, because we can perform the join locally. In this example we
> >> are
> >> sending both of the tables to multiple nodes, which should defeat
> >> the
> >> purpose.
> >>
> >> Is it about implementing cross join or near cross joins in a
> >> distributed fashion?
> >>
> >> if we introduce a new MulticastRecordWriter
> >>
> >> That’s one of the solutions. It might have a drawback of 3 class
> >> virtualisation problem (We have RecordWriter and
> >> BroadcastRecordWriter
> >> already). With up to two implementations, JVM is able to
> >> devirtualise
> >> the
> >> calls.
> >>
> >> Previously I was also thinking about just providing two different
> >> ChannelSelector interfaces. One with `int[]` and
> >> `SingleChannelSelector`
> >> with plain `int` and based on that, RecordWriter could perform some
> >> magic
> >> (worst case scenario `instaceof` checks).
> >>
> >> Another solution might be to change `ChannelSelector` interface
> >> into
> >> an iterator.
> >>
> >> But let's discuss the details after we agree on implementing this.
> >>
> >> Piotrek
> >>
> >> On 23 Aug 2019, at 10:20, Yun Gao <[hidden email] <mailto:
> >> [hidden email]>> wrote:
> >>
> >>  Hi Piotr,
> >>
> >>       Thanks a lot for the suggestions!
> >>
> >>       The core motivation of this discussion is to implement a
> >> new
> >> iteration library on the DataStream, and it requires to insert
> >> special
> >> records in the stream to notify the progress of the iteration. The
> >> mechanism of such records is very similar to the current Watermark,
> >> and
> >> we
> >> meet the problem of sending normal records according to the
> >> partition
> >> (Rebalance, etc..) and also be able to broadcast the inserted
> >> progress
> >> records to all the connected records. I have read the notes in the
> >> google
> >> doc and I totally agree with that exploring the broadcast interface
> >> in
> >> RecordWriter in some way is able to solve this issue.
> >>
> >>      Regarding to `int[] ChannelSelector#selectChannels()`, I'm
> >> wondering if we introduce a new MulticastRecordWriter and left the
> >> current
> >> RecordWriter untouched, could we avoid the performance degradation ?
> >> Since
> >> with such a modification the normal RecordWriter does not need to
> >> iterate
> >> the return array by ChannelSelector, and the only difference will be
> >> returning an array instead of an integer, and accessing the first
> >> element
> >> of the returned array instead of reading the integer directly.
> >>
> >> Best,
> >> Yun
> >>
> >>
> >> ------------------------------------------------------------------
> >> From:Piotr Nowojski <[hidden email] <mailto:
> >> [hidden email]
> >>
> >> Send Time:2019 Aug. 23 (Fri.) 15:20
> >> To:dev <[hidden email] <mailto:[hidden email]>>
> >> Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
> >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
> >> Pattern
> >>
> >> Hi,
> >>
> >> Yun:
> >>
> >> Thanks for proposing the idea. I have checked the document and
> >> left
> >> couple of questions there, but it might be better to answer them
> >> here.
> >>
> >> What is the exact motivation and what problems do you want to
> >> solve?
> >> We have dropped multicast support from the network stack [1] for two
> >> reasons:
> >> 1. Performance
> >> 2. Code simplicity
> >>
> >> The proposal to re introduce `int[]
> >> ChannelSelector#selectChannels()`
> >> would revert those changes. At that time we were thinking about a
> >> way
> >> how
> >> to keep the multicast support on the network level, while keeping
> >> the
> >> performance and simplicity for non multicast cases and there are
> >> ways
> >> to
> >> achieve that. However they would add extra complexity to Flink,
> >> which
> >> it
> >> would be better to avoid.
> >>
> >> On the other hand, supporting dual pattern: standard partitioning
> >> or
> >> broadcasting is easy to do, as LatencyMarkers are doing exactly
> >> that.
> >> It
> >> would be just a matter of exposing this to the user in some way. So
> >> before
> >> we go any further, can you describe your use cases/motivation? Isn’t
> >> mix of
> >> standard partitioning and broadcasting enough? Do we need
> >> multicasting?
> >>
> >> Zhu:
> >>
> >> Could you rephrase your example? I didn’t quite understand it.
> >>
> >> Piotrek
> >>
> >> [1] https://issues.apache.org/jira/browse/FLINK-10662 <
> >> https://issues.apache.org/jira/browse/FLINK-10662> <
> >> https://issues.apache.org/jira/browse/FLINK-10662 <
> >> https://issues.apache.org/jira/browse/FLINK-10662>>
> >>
> >> On 23 Aug 2019, at 09:17, Zhu Zhu <[hidden email] <mailto:
> >> [hidden email]> <mailto:[hidden email] <mailto:
> >> [hidden email]
> >>
> >> wrote:
> >>
> >> Thanks Yun for starting this discussion.
> >> I think the multicasting can be very helpful in certain cases.
> >>
> >> I have received requirements from users that they want to do
> >> broadcast
> >> join, while the data set to broadcast is too large to fit in one
> >> task.
> >> Thus the requirement turned out to be to support cartesian
> >> product
> >> of
> >> 2
> >> data set(one of which can be infinite stream).
> >> For example, A(parallelism=2) broadcast join B(parallelism=2) in
> >> JobVertex
> >> C.
> >> The idea to is have 4 C subtasks to deal with different
> >> combinations
> >> of A/B
> >> partitions, like C1(A1,B1), C2(A1,B2), C3(A2,B1), C4(A2,B2).
> >> This requires one record to be sent to multiple downstream
> >> subtasks,
> >> but
> >> not to all subtasks.
> >>
> >> With current interface this is not supported, as one record can
> >> only
> >> be
> >> sent to one subtask, or to all subtasks of a JobVertex.
> >> And the user had to split the broadcast data set manually to
> >> several
> >> different JobVertices, which is hard to maintain and extend.
> >>
> >> Thanks,
> >> Zhu Zhu
> >>
> >> Yun Gao <[hidden email] <mailto:
> >> [hidden email] <mailto:[hidden email]
> >>
> >> 于2019年8月22日周四 下午8:42写道：
> >>
> >> Hi everyone,
> >>    In some scenarios we met a requirement that some operators
> >> want
> >> to
> >> send records to theirs downstream operators with an multicast
> >> communication
> >> pattern. In detail, for some records, the operators want to send
> >> them
> >> according to the partitioner (for example, Rebalance), and for
> >> some
> >> other
> >> records, the operators want to send them to all the connected
> >> operators and
> >> tasks. Such a communication pattern could be viewed as a kind of
> >> multicast:
> >> it does not broadcast every record, but some record will indeed
> >> be
> >> sent to
> >> multiple downstream operators.
> >>
> >> However, we found that this kind of communication pattern seems
> >> could
> >> not
> >> be implemented rightly if the operators have multiple consumers
> >> with
> >> different parallelism, using the customized partitioner. To solve
> >> the
> >> above
> >> problem, we propose to enhance the support for such kind of
> >> irregular
> >> communication pattern. We think there may be two options:
> >>
> >>   1. Support a kind of customized operator events, which share
> >> much
> >> similarity with Watermark, and these events can be broadcasted to
> >> the
> >> downstream operators separately.
> >>   2. Let the channel selector supports multicast, and also add
> >> the
> >> separate RecordWriter implementation to avoid impacting the
> >> performance of
> >> the channel selector that does not need multicast.
> >>
> >> The problem and options are detailed in
> >>
> >>
> >>
> >>
> >>
> >>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
> >> <
> >>
> >>
> >>
> >>
> >>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
> >>
> >> <
> >>
> >>
> >>
> >>
> >>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
> >> <
> >>
> >>
> >>
> >>
> >>
> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
> >>
> >>
> >> We are also wondering if there are other methods to implement
> >> this
> >> requirement with or without changing Runtime. Very thanks for any
> >> feedbacks
> >> !
> >>
> >>
> >> Best,
> >> Yun
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
> >>
>
>

[Yun Gao]

Hi all,

Very thanks Jark for the new scenarios. Based on the these new scenarios, I think these scenarios and iteration should be able to represent a type of scenarios that requires broadcasting events.

I also totally agree with Piotr that all the scenarios we have discussed should be clearly motivated. From what we learned from the discussion, now we think that broadcasting events seems to be most suitable for iteration and also some other scenarios, therefore, we would rewrite a motivation design doc for broadcasting events first and reinitiate a separate discussion for that. The current discussion would be then continue for scenarios require actual multicasting. Very thanks for all the valuable points raised, and I think now the comparison of different methods and scenarios are more clear. :)

Best,
Yun


------------------------------------------------------------------
From:Jark Wu <[hidden email]>
Send Time:2019 Aug. 27 (Tue.) 16:27
To:dev <[hidden email]>
Cc:Yun Gao <[hidden email]>
Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern

Hi all,

Thanks Yun for bringing this topic. I missed this discussion because of the "multicast" title.
After reading the design, if I understand correctly, it is proposing a custom event mach mechanism, i.e. broadcasting custom event.
It is a orthogonality topic with multicasting. So I would suggest to start a new thread to discuss about it.

Regarding to broadcasting custom event:

I would +1 for motivation, because we also encountered similar requirements when improving Table API & SQL before.

For example, the mini-batch mechanism in blink planner will emit a special mini-batch event to the data stream to indicate this is a start of a new mini-batch.
The downstream aggregation operator will buffer the data records until it receive the mini-batch event, and then process the buffer at once. This will reduce a lot of state access.
However, we don't have a proper custom event mechanism currently, so we leverage the watermark as the mini-batch event (which is a little hack in my opinion).

Another case is joining a huge dimension table which is stored/produced in hive daily. We can scan the hive table and shuffle to the JOIN operators by the join key to join with the main stream.
Note that the dimension table is changed every day, we want to join the latest version of the hive table. Then we need to re-scan and re-shuffle the hive table once a new daily partition is produced.
However, we need some special events to distinguish the boundary of different version of the dimension table. The events will be used to notify downstream operators (mainly the JOIN operator)
 to know "ok, I will receive a new version of the dimension table", "ok, I received the all the data of this version."

From my understanding, in order to support this feature, we might need to:
 1) expose collectEvent(CustomEvent) or broadcastEvent(CustomEvent) API to users.
 2) support to register the serialization and deserialization of the custom event
 3) expose processEvent(int channel, CustomEvent) API to StreamOperator


Regards,
Jark


On Tue, 27 Aug 2019 at 15:18, Piotr Nowojski <[hidden email]> wrote:
Hi,

 Before starting a work on the design doc, I would suggest to find someone to shepherd this project. Otherwise this effort might drown among other parallel things. I could take care of that from the runtime perspective, however most of the changes are about the API and changes, which are outside of my area of expertise.

 Regarding the multicast, before we start working on that, I would also prefer to see a motivation design doc, how that feature would be used for example for cross or theta joins in the Table API, since very similar questions would apply to that as well.

 Piotrek

 > On 27 Aug 2019, at 08:10, SHI Xiaogang <[hidden email]> wrote:
 >
 > Hi Yun Gao,
 >
 > Thanks a lot for your clarification.
 >
 > Now that the notification of broadcast events requires alignment whose
 > implementation, in my opinion, will affect the correctness of synchronous
 > iterations, I prefer to postpone the discussion until you have completed
 > the design of the new iteration library, or at least the progress tracking
 > part. Otherwise, the discussion for broadcasting events may become an empty
 > talk if it does not fit in with the final design.
 >
 > What do you think?
 >
 > Regards,
 > Xiaogang
 >
 > Yun Gao <[hidden email]> 于2019年8月27日周二 上午11:33写道：
 >
 >>     Hi Xiaogang,
 >>
 >>      Very thanks for also considering the iteration case! :) These points
 >> are really important for iteration. As a whole, we are implementing a new
 >> iteration library on top of Stream API. As a library, most of its
 >> implementation does not need to touch Runtime layer, but it really has some
 >> new requirements on the API, like the one for being able to broadcast the
 >> progressive events. To be more detail, these events indeed carry the
 >> sender's index and the downstream operators need to do alignment the events
 >> from all the upstream operators. It works very similar to watermark, thus
 >> these events do not need to be contained in checkpoints.
 >>
 >> Some other points are also under implementation. However, since some part
 >> of the design is still under discussion internally, we may not be able to
 >> start a new discussion on iteration immediately. Besides, we should also
 >> need to fix the problems that may have new requirements on the Runtime,
 >> like broadcasting events, to have a complete design. Therefore, I think we
 >> may still first have the broadcasting problem settled in this thread? Based
 >> on the points learned in the discussion, now I think that we might be able
 >> to decouple the broadcasting events requirements and more generalized
 >> multicasting mechanism. :)
 >>
 >> Best,
 >> Yun
 >>
 >>
 >>
 >> ------------------------------------------------------------------
 >> From:SHI Xiaogang <[hidden email]>
 >> Send Time:2019 Aug. 27 (Tue.) 09:16
 >> To:dev <[hidden email]>; Yun Gao <[hidden email]>
 >> Cc:Piotr Nowojski <[hidden email]>
 >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern
 >>
 >> Hi, Yun Gao
 >>
 >> The discussion seems to move in a different direction, changing from
 >> supporting multicasting to implementing new iteration libraries on data
 >> streams.
 >>
 >> Regarding the broadcast events in iterations, many details of new
 >> iteration libraries are unclear,
 >> 1. How the iteration progress is determined and notified? The iterations
 >> are synchronous or asynchronous? As far as i know, progress tracking for
 >> asynchronous iterations is very difficult.
 >> 2. Do async I/O operators allowed in the iterations? If so, how the
 >> broadcast events are checkpointed and restored? How broadcast events are
 >> distributed when the degree of parallelism changes?
 >> 3. Do the emitted broadcast events carry the sender's index? Will they be
 >> aligned in a similar way to checkpoint barriers in downstream operators?
 >> 4. In the case of synchronous iterations, do we need something similar to
 >> barrier buffers to guarantee the correctness of iterations?
 >> 5. Will checkpointing be enabled in iterations? If checkpointing is
 >> enabled, how will checkpoint barriers interact with broadcast events?
 >>
 >> I think a detailed design document for iterations will help understand
 >> these problems, hencing improving the discussion.
 >>
 >> I also suggest a new thread for the discussion on iterations.
 >> This thread should focus on multicasting and discuss those problems
 >> related to multicasting, including how data is delivered and states are
 >> partitioned.
 >>
 >> Regards,
 >> Xiaogang
 >> Yun Gao <[hidden email]> 于2019年8月26日周一 下午11:35写道：
 >>
 >> Hi,
 >>
 >> Very thanks for all the points raised !
 >>
 >> @Piotr For using another edge to broadcast the event, I think it may not
 >> be able to address the iteration case. The primary problem is that with
 >> two edges we cannot ensure the order of records. However, In the iteration
 >> case, the broadcasted event is used to mark the progress of the iteration
 >> and it works like watermark, thus its position relative to the normal
 >> records can not change.
 >> And @Piotr, @Xiaogang, for the requirements on the state, I think
 >> different options seems vary. The first option is to allow Operator<T> to
 >> broadcast a separate event and have a separate process method for this
 >> event. To be detail, we may add a new type of StreamElement called Event
 >> and allow Operator<T> to broadcastEmit Event. Then in the received side, we
 >> could add a new `processEvent` method to the (Keyed)ProcessFunction.
 >> Similar to the broadcast side of KeyedBroadcastProcessFunction, in this new
 >> method users cannot access keyed state with specific key, but can register
 >> a state function to touch all the elements in the keyed state. This option
 >> needs to modify the runtime to support the new type of StreamElement, but
 >> it does not affect the semantics of states and thus it has no requirements
 >> on state.
 >> The second option is to allow Operator<T> to broadcastEmit T and in the
 >> receiver side, user can process the broadcast element with the existing
 >> process method. This option is consistent with the OperatorState, but for
 >> keyedState we may send a record to tasks that do not containing the
 >> corresponding keyed state, thus it should require some changes on the State.
 >> The third option is to support the generic Multicast. For keyedState it
 >> also meets the problem of inconsistency between network partitioner and
 >> keyed state partitioner, and if we want to rely on it to implement the
 >> non-key join, it should be also meet the problem of cannot control the
 >> partitioning of operator state. Therefore, it should also require some
 >> changes on the State.
 >> Then for the different scenarios proposed, the iteration case in fact
 >> requires exactly the ability to broadcast a different event type. In the
 >> iteration the fields of the progress event are in fact different from that
 >> of normal records. It does not contain actual value but contains some
 >> fields for the downstream operators to align the events and track the
 >> progress. Therefore, broadcasting a different event type is able to solve
 >> the iteration case without the requirements on the state. Besides, allowing
 >> the operator to broadcast a separate event may also facilitate some other
 >> user cases, for example, users may notify the downstream operators to
 >> change logic if some patterns are matched. The notification might be
 >> different from the normal records and users do not need to uniform them
 >> with a wrapper type manually if the operators are able to broadcast a
 >> separate event. However, it truly cannot address the non-key join
 >> scenarios.
 >> Since allowing broadcasting a separate event seems to be able to serve as
 >> a standalone functionality, and it does not require change on the state, I
 >> am thinking that is it possible for us to partition to multiple steps and
 >> supports broadcasting events first ? At the same time we could also
 >> continue working on other options to support more scenarios like non-key
 >> join and they seems to requires more thoughts.
 >>
 >> Best,
 >> Yun
 >>
 >>
 >>
 >> ------------------------------------------------------------------
 >> From:Piotr Nowojski <[hidden email]>
 >> Send Time:2019 Aug. 26 (Mon.) 18:59
 >> To:dev <[hidden email]>
 >> Cc:Yun Gao <[hidden email]>
 >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication Pattern
 >>
 >> Hi,
 >>
 >> Xiaogang, those things worry me the most.
 >> 1. Regarding the broadcasting, doesn’t the BroadcastState [1] cover our
 >> issues? Can not we construct a job graph, where one operator has two
 >> outputs, one keyed another broadcasted, which are wired together back to
 >> the KeyedBroadcastProcessFunction or BroadcastProcessFunction?
 >>
 >> 2. Multicast on keyed streams, might be done by iterating over all of the
 >> keys. However I have a feeling that might not be the feature which
 >> distributed cross/theta joins would want, since they would probably need a
 >> guarantee to have only a single key per operator instance.
 >>
 >> Kurt, by broadcast optimisation do you mean [2]?
 >>
 >> I’m not sure if we should split the discussion yet. Most of the changes
 >> required by either multicast or broadcast will be in the API/state layers.
 >> Runtime changes for broadcast would be almost none (just exposing existing
 >> features) and for multicast they shouldn't be huge as well. However maybe
 >> we should consider those two things together at the API level, so that we
 >> do not make wrong decisions when just looking at the simpler/more narrow
 >> broadcast support?
 >>
 >> Piotrek
 >>
 >> [1]
 >> https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/state/broadcast_state.html
 >> [2] https://github.com/apache/flink/pull/7713
 >>
 >>
 >> On 26 Aug 2019, at 09:35, Kurt Young <[hidden email]> wrote:
 >> From SQL's perspective, distributed cross join is a valid feature but not
 >> very
 >> urgent. Actually this discuss reminds me about another useful feature
 >> (sorry
 >> for the distraction):
 >>
 >> when doing broadcast in batch shuffle mode, we can make each producer only
 >> write one copy of the output data, but not for every consumer. Broadcast
 >> join
 >> is much more useful, and this is a very important optimization. Not sure
 >> if
 >> we
 >> have already consider this.
 >>
 >> Best,
 >> Kurt
 >>
 >>
 >> On Mon, Aug 26, 2019 at 12:16 PM Guowei Ma <[hidden email]> wrote:
 >>
 >> Thanks Yun for bringing up this discussion and very thanks for all the
 >> deep
 >> thoughts!
 >>
 >> For now, I think this discussion contains two scenarios: one if for
 >> iteration library support and the other is for SQL join support. I think
 >> both of the two scenarios are useful but they seem to have different best
 >> suitable solutions. For making the discussion more clear, I would suggest
 >> to split the discussion into two threads.
 >>
 >> And I agree with Piotr that it is very tricky that a keyed stream received
 >> a "broadcast element". So we may add some new interfaces, which could
 >> broadcast or process some special "broadcast event". In that way
 >> "broadcast
 >> event" will not be sent with the normal process.
 >>
 >> Best,
 >> Guowei
 >>
 >>
 >> SHI Xiaogang <[hidden email]> 于2019年8月26日周一 上午9:27写道：
 >>
 >> Hi all,
 >>
 >> I also think that multicasting is a necessity in Flink, but more details
 >> are needed to be considered.
 >>
 >> Currently network is tightly coupled with states in Flink to achieve
 >> automatic scaling. We can only access keyed states in keyed streams and
 >> operator states in all streams.
 >> In the concrete example of theta-joins implemented with mutlticasting,
 >> the
 >> following questions exist:
 >>
 >>    - In which type of states will the data be stored? Do we need another
 >>    type of states which is coupled with multicasting streams?
 >>    - How to ensure the consistency between network and states when jobs
 >>    scale out or scale in?
 >>
 >> Regards,
 >> Xiaogang
 >>
 >> Xingcan Cui <[hidden email]> 于2019年8月25日周日 上午10:03写道：
 >>
 >> Hi all,
 >>
 >> Sorry for joining this thread late. Basically, I think enabling
 >> multicast
 >> pattern could be the right direction, but more detailed implementation
 >> policies need to be discussed.
 >>
 >> Two years ago, I filed an issue [1] about the multicast API. However,
 >> due
 >> to some reasons, it was laid aside. After that, when I tried to
 >> cherry-pick
 >> the change for experimental use, I found the return type of
 >> `selectChannels()` method had changed from `int[]` to `int`, which
 >> makes
 >> the old implementation not work anymore.
 >>
 >> From my side, the multicast has always been used for theta-join. As far
 >> as
 >> I know, it’s an essential requirement for some sophisticated joining
 >> algorithms. Until now, the Flink non-equi joins can still only be
 >> executed
 >> single-threaded. If we'd like to make some improvements on this, we
 >> should
 >> first take some measures to support multicast pattern.
 >>
 >> Best,
 >> Xingcan
 >>
 >> [1] https://issues.apache.org/jira/browse/FLINK-6936
 >>
 >> On Aug 24, 2019, at 5:54 AM, Zhu Zhu <[hidden email]> wrote:
 >>
 >> Hi Piotr,
 >>
 >> Thanks for the explanation.
 >> Agreed that the broadcastEmit(record) is a better choice for
 >> broadcasting
 >> for the iterations.
 >> As broadcasting for the iterations is the first motivation, let's
 >> support
 >> it first.
 >>
 >> Thanks,
 >> Zhu Zhu
 >>
 >> Yun Gao <[hidden email]> 于2019年8月23日周五 下午11:56写道：
 >>
 >>     Hi Piotr,
 >>
 >>      Very thanks for the suggestions!
 >>
 >>     Totally agree with that we could first focus on the broadcast
 >> scenarios and exposing the broadcastEmit method first considering
 >> the
 >> semantics and performance.
 >>
 >>     For the keyed stream, I also agree with that broadcasting keyed
 >> records to all the tasks may be confused considering the semantics
 >> of
 >> keyed
 >> partitioner. However, in the iteration case supporting broadcast
 >> over
 >> keyed
 >> partitioner should be required since users may create any subgraph
 >> for
 >> the
 >> iteration body, including the operators with key. I think a possible
 >> solution to this issue is to introduce another data type for
 >> 'broadcastEmit'. For example, for an operator Operator<T>, it may
 >> broadcast
 >> emit another type E instead of T, and the transmitting E will bypass
 >> the
 >> partitioner and setting keyed context. This should result in the
 >> design
 >> to
 >> introduce customized operator event (option 1 in the document). The
 >> cost of
 >> this method is that we need to introduce a new type of StreamElement
 >> and
 >> new interface for this type, but it should be suitable for both
 >> keyed
 >> or
 >> non-keyed partitioner.
 >>
 >> Best,
 >> Yun
 >>
 >>
 >>
 >> ------------------------------------------------------------------
 >> From:Piotr Nowojski <[hidden email]>
 >> Send Time:2019 Aug. 23 (Fri.) 22:29
 >> To:Zhu Zhu <[hidden email]>
 >> Cc:dev <[hidden email]>; Yun Gao <[hidden email]>
 >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
 >> Pattern
 >>
 >> Hi,
 >>
 >> If the primary motivation is broadcasting (for the iterations) and
 >> we
 >> have
 >> no immediate need for multicast (cross join), I would prefer to
 >> first
 >> expose broadcast via the DataStream API and only later, once we
 >> finally
 >> need it, support multicast. As I wrote, multicast would be more
 >> challenging
 >> to implement, with more complicated runtime and API. And re-using
 >> multicast
 >> just to support broadcast doesn’t have much sense:
 >>
 >> 1. It’s a bit obfuscated. It’s easier to understand
 >> collectBroadcast(record) or broadcastEmit(record) compared to some
 >> multicast channel selector that just happens to return all of the
 >> channels.
 >> 2. There are performance benefits of explicitly calling
 >> `RecordWriter#broadcastEmit`.
 >>
 >>
 >> On a different note, what would be the semantic of such broadcast
 >> emit
 >> on
 >> KeyedStream? Would it be supported? Or would we limit support only
 >> to
 >> the
 >> non-keyed streams?
 >>
 >> Piotrek
 >>
 >> On 23 Aug 2019, at 12:48, Zhu Zhu <[hidden email]> wrote:
 >>
 >> Thanks Piotr,
 >>
 >> Users asked for this feature sometimes ago when they migrating
 >> batch
 >> jobs to Flink(Blink).
 >> It's not very urgent as they have taken some workarounds to solve
 >> it.(like partitioning data set to different job vertices)
 >> So it's fine to not make it top priority.
 >>
 >> Anyway, as a commonly known scenario, I think users can benefit
 >> from
 >> cross join sooner or later.
 >>
 >> Thanks,
 >> Zhu Zhu
 >>
 >> Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
 >> 于2019年8月23日周五 下午6:19写道：
 >> Hi,
 >>
 >> Thanks for the answers :) Ok I understand the full picture now. +1
 >> from
 >> my side on solving this issue somehow. But before we start
 >> discussing
 >> how
 >> to solve it one last control question:
 >>
 >> I guess this multicast is intended to be used in blink planner,
 >> right?
 >> Assuming that we implement the multicast support now, when would it
 >> be
 >> used
 >> by the blink? I would like to avoid a scenario, where we implement
 >> an
 >> unused feature and we keep maintaining it for a long period of time.
 >>
 >> Piotrek
 >>
 >> PS, try to include motivating examples, including concrete ones in
 >> the
 >> proposals/design docs, for example in the very first paragraph.
 >> Especially
 >> if it’s a commonly known feature like cross join :)
 >>
 >> On 23 Aug 2019, at 11:38, Yun Gao <[hidden email]>
 >> wrote:
 >>
 >>    Hi Piotr,
 >>
 >>       Thanks a lot for sharing the thoughts!
 >>
 >>       For the iteration, agree with that multicasting is not
 >> necessary. Exploring the broadcast interface to Output of the
 >> operators
 >> in
 >> some way should also solve this issue, and I think it should be even
 >> more
 >> convenient to have the broadcast method for the iteration.
 >>
 >>       Also thanks Zhu Zhu for the cross join case!
 >> Best,
 >>  Yun
 >>
 >>
 >>
 >> ------------------------------------------------------------------
 >> From:Zhu Zhu <[hidden email] <mailto:[hidden email]>>
 >> Send Time:2019 Aug. 23 (Fri.) 17:25
 >> To:dev <[hidden email] <mailto:[hidden email]>>
 >> Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
 >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
 >> Pattern
 >>
 >> Hi Piotr,
 >>
 >> Yes you are right it's a distributed cross join requirement.
 >> Broadcast join can help with cross join cases. But users cannot
 >> use
 >> it
 >> if the data set to join is too large to fit into one subtask.
 >>
 >> Sorry for left some details behind.
 >>
 >> Thanks,
 >> Zhu Zhu
 >> Piotr Nowojski <[hidden email] <mailto:[hidden email]>>
 >> 于2019年8月23日周五 下午4:57写道：
 >> Hi Yun and Zhu Zhu,
 >>
 >> Thanks for the more detailed example Zhu Zhu.
 >>
 >> As far as I understand for the iterations example we do not need
 >> multicasting. Regarding the Join example, I don’t fully understand
 >> it.
 >> The
 >> example that Zhu Zhu presented has a drawback of sending both tables
 >> to
 >> multiple nodes. What’s the benefit of using broadcast join over a
 >> hash
 >> join
 >> in such case? As far as I know, the biggest benefit of using
 >> broadcast
 >> join
 >> instead of hash join is that we can avoid sending the larger table
 >> over
 >> the
 >> network, because we can perform the join locally. In this example we
 >> are
 >> sending both of the tables to multiple nodes, which should defeat
 >> the
 >> purpose.
 >>
 >> Is it about implementing cross join or near cross joins in a
 >> distributed fashion?
 >>
 >> if we introduce a new MulticastRecordWriter
 >>
 >> That’s one of the solutions. It might have a drawback of 3 class
 >> virtualisation problem (We have RecordWriter and
 >> BroadcastRecordWriter
 >> already). With up to two implementations, JVM is able to
 >> devirtualise
 >> the
 >> calls.
 >>
 >> Previously I was also thinking about just providing two different
 >> ChannelSelector interfaces. One with `int[]` and
 >> `SingleChannelSelector`
 >> with plain `int` and based on that, RecordWriter could perform some
 >> magic
 >> (worst case scenario `instaceof` checks).
 >>
 >> Another solution might be to change `ChannelSelector` interface
 >> into
 >> an iterator.
 >>
 >> But let's discuss the details after we agree on implementing this.
 >>
 >> Piotrek
 >>
 >> On 23 Aug 2019, at 10:20, Yun Gao <[hidden email] <mailto:
 >> [hidden email]>> wrote:
 >>
 >>  Hi Piotr,
 >>
 >>       Thanks a lot for the suggestions!
 >>
 >>       The core motivation of this discussion is to implement a
 >> new
 >> iteration library on the DataStream, and it requires to insert
 >> special
 >> records in the stream to notify the progress of the iteration. The
 >> mechanism of such records is very similar to the current Watermark,
 >> and
 >> we
 >> meet the problem of sending normal records according to the
 >> partition
 >> (Rebalance, etc..) and also be able to broadcast the inserted
 >> progress
 >> records to all the connected records. I have read the notes in the
 >> google
 >> doc and I totally agree with that exploring the broadcast interface
 >> in
 >> RecordWriter in some way is able to solve this issue.
 >>
 >>      Regarding to `int[] ChannelSelector#selectChannels()`, I'm
 >> wondering if we introduce a new MulticastRecordWriter and left the
 >> current
 >> RecordWriter untouched, could we avoid the performance degradation ?
 >> Since
 >> with such a modification the normal RecordWriter does not need to
 >> iterate
 >> the return array by ChannelSelector, and the only difference will be
 >> returning an array instead of an integer, and accessing the first
 >> element
 >> of the returned array instead of reading the integer directly.
 >>
 >> Best,
 >> Yun
 >>
 >>
 >> ------------------------------------------------------------------
 >> From:Piotr Nowojski <[hidden email] <mailto:
 >> [hidden email]
 >>
 >> Send Time:2019 Aug. 23 (Fri.) 15:20
 >> To:dev <[hidden email] <mailto:[hidden email]>>
 >> Cc:Yun Gao <[hidden email] <mailto:[hidden email]>>
 >> Subject:Re: [DISCUSS] Enhance Support for Multicast Communication
 >> Pattern
 >>
 >> Hi,
 >>
 >> Yun:
 >>
 >> Thanks for proposing the idea. I have checked the document and
 >> left
 >> couple of questions there, but it might be better to answer them
 >> here.
 >>
 >> What is the exact motivation and what problems do you want to
 >> solve?
 >> We have dropped multicast support from the network stack [1] for two
 >> reasons:
 >> 1. Performance
 >> 2. Code simplicity
 >>
 >> The proposal to re introduce `int[]
 >> ChannelSelector#selectChannels()`
 >> would revert those changes. At that time we were thinking about a
 >> way
 >> how
 >> to keep the multicast support on the network level, while keeping
 >> the
 >> performance and simplicity for non multicast cases and there are
 >> ways
 >> to
 >> achieve that. However they would add extra complexity to Flink,
 >> which
 >> it
 >> would be better to avoid.
 >>
 >> On the other hand, supporting dual pattern: standard partitioning
 >> or
 >> broadcasting is easy to do, as LatencyMarkers are doing exactly
 >> that.
 >> It
 >> would be just a matter of exposing this to the user in some way. So
 >> before
 >> we go any further, can you describe your use cases/motivation? Isn’t
 >> mix of
 >> standard partitioning and broadcasting enough? Do we need
 >> multicasting?
 >>
 >> Zhu:
 >>
 >> Could you rephrase your example? I didn’t quite understand it.
 >>
 >> Piotrek
 >>
 >> [1] https://issues.apache.org/jira/browse/FLINK-10662 <
 >> https://issues.apache.org/jira/browse/FLINK-10662> <
 >> https://issues.apache.org/jira/browse/FLINK-10662 <
 >> https://issues.apache.org/jira/browse/FLINK-10662>>
 >>
 >> On 23 Aug 2019, at 09:17, Zhu Zhu <[hidden email] <mailto:
 >> [hidden email]> <mailto:[hidden email] <mailto:
 >> [hidden email]
 >>
 >> wrote:
 >>
 >> Thanks Yun for starting this discussion.
 >> I think the multicasting can be very helpful in certain cases.
 >>
 >> I have received requirements from users that they want to do
 >> broadcast
 >> join, while the data set to broadcast is too large to fit in one
 >> task.
 >> Thus the requirement turned out to be to support cartesian
 >> product
 >> of
 >> 2
 >> data set(one of which can be infinite stream).
 >> For example, A(parallelism=2) broadcast join B(parallelism=2) in
 >> JobVertex
 >> C.
 >> The idea to is have 4 C subtasks to deal with different
 >> combinations
 >> of A/B
 >> partitions, like C1(A1,B1), C2(A1,B2), C3(A2,B1), C4(A2,B2).
 >> This requires one record to be sent to multiple downstream
 >> subtasks,
 >> but
 >> not to all subtasks.
 >>
 >> With current interface this is not supported, as one record can
 >> only
 >> be
 >> sent to one subtask, or to all subtasks of a JobVertex.
 >> And the user had to split the broadcast data set manually to
 >> several
 >> different JobVertices, which is hard to maintain and extend.
 >>
 >> Thanks,
 >> Zhu Zhu
 >>
 >> Yun Gao <[hidden email] <mailto:
 >> [hidden email] <mailto:[hidden email]
 >>
 >> 于2019年8月22日周四 下午8:42写道：
 >>
 >> Hi everyone,
 >>    In some scenarios we met a requirement that some operators
 >> want
 >> to
 >> send records to theirs downstream operators with an multicast
 >> communication
 >> pattern. In detail, for some records, the operators want to send
 >> them
 >> according to the partitioner (for example, Rebalance), and for
 >> some
 >> other
 >> records, the operators want to send them to all the connected
 >> operators and
 >> tasks. Such a communication pattern could be viewed as a kind of
 >> multicast:
 >> it does not broadcast every record, but some record will indeed
 >> be
 >> sent to
 >> multiple downstream operators.
 >>
 >> However, we found that this kind of communication pattern seems
 >> could
 >> not
 >> be implemented rightly if the operators have multiple consumers
 >> with
 >> different parallelism, using the customized partitioner. To solve
 >> the
 >> above
 >> problem, we propose to enhance the support for such kind of
 >> irregular
 >> communication pattern. We think there may be two options:
 >>
 >>   1. Support a kind of customized operator events, which share
 >> much
 >> similarity with Watermark, and these events can be broadcasted to
 >> the
 >> downstream operators separately.
 >>   2. Let the channel selector supports multicast, and also add
 >> the
 >> separate RecordWriter implementation to avoid impacting the
 >> performance of
 >> the channel selector that does not need multicast.
 >>
 >> The problem and options are detailed in
 >>
 >>
 >>
 >>
 >>
 >> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
 >> <
 >>
 >>
 >>
 >>
 >> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
 >>
 >> <
 >>
 >>
 >>
 >>
 >> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
 >> <
 >>
 >>
 >>
 >>
 >> https://docs.google.com/document/d/1npi5c_SeP68KuT2lNdKd8G7toGR_lxQCGOnZm_hVMks/edit?usp=sharing
 >>
 >>
 >> We are also wondering if there are other methods to implement
 >> this
 >> requirement with or without changing Runtime. Very thanks for any
 >> feedbacks
 >> !
 >>
 >>
 >> Best,
 >> Yun
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>
 >>