dynamically partitioned stream
Locked
 
|
Hi all,
I am trying to implement the following using Flink: I have 2 input message streams: 1. Data Stream: KEY VALUE TIME . . . C V6 6 B V6 6 A V5 5 A V4 4 C V3 3 A V3 3 B V3 3 B V2 2 A V1 1 2. Control Stream: Lambda ArgumentKeys TIME . . . f2 [A, C] 4 f1 [A, B, C] 1 I want to apply the lambdas coming in the control stream to the selection of keys that are coming in the data stream. Since we have 2 streams I naturally thought of connecting them using .connect. For this I need to key both of them by a certain criteria. And here lies the problem, how can I make sure the messages with keys A,B,C specified in the control stream end up in the same task as well as the control message (f1, [A, B, C]) itself. Basically I don't know how to key by to achieve this. I suspect a custom partitioner is required that partitions the data stream based on the messages in the control stream? Is this even possible? Any suggestions welcomed! Thanks, M |
|
|
Hi Martin, Let me understand your question first. You have two Stream: Data Stream and Control Stream and you want to select data in Data Stream based on the key set got from Control Stream. If I were not misunderstanding your question, I think SideInput is what you want. https://cwiki.apache.org/confluence/display/FLINK/FLIP-17+Side+Inputs+for+DataStream+API#FLIP-17SideInputsforDataStreamAPI-StoringSide-InputData It lets you to define one stream as a SideInput and can be assigned to the other stream, then the data in SideInput stream will be broadcasted. So far, I have no idea if there is any solution to solve this without SideInput. Best, Tony Wei 2017-08-31 16:10 GMT+08:00 Martin Eden <[hidden email]>:
|
Re: dynamically partitioned stream
|
|
Thanks for your reply Tony. So there are actually 2 problems to solve: 1. All control stream msgs need to be broadcasted to all tasks. 2. The data stream messages with the same keys as those specified in the control message need to go to the same task as well, so that all the values required for the lambda (i.e. functions f1, f2 ...) are there. In my understanding side inputs (which are actually not available in the current release) would address problem 1. To address problem 1 I also tried dataStream.keyBy(key).connect(controlStream.broadcast).flatMap(new RichCoFlatMapFunction) but I get a runtime exception telling me I still need to do a keyBy before the flatMap. So are the upcoming side inputs the only way to broadcast a control stream to all tasks of a coFlatMap? Or is there another way? As for problem 2, I am still pending a reply. Would appreciate if anyone has some suggestions. Thanks, M On Thu, Aug 31, 2017 at 9:59 AM, Tony Wei <[hidden email]> wrote:
|
|
|
Hi Martin, About problem 2. How were those lambda functions created? Pre-defined functions / operators or automatically generated based on the message from Control Stream? For the former, you could give each function one id and user flapMap to duplicate data with multiple ids. Then, you could use filter function and send them to the corresponding operators. For the general case like the latter, because you had broadcasted the messages to all tasks, it could always build a mapping table from argument keys to lambda functions in each sub-task and use the map to process the data. But I was wondering if it is possible to generate a completely new function in the runtime. Best, Tony Wei 2017-08-31 18:33 GMT+08:00 Martin Eden <[hidden email]>:
|
Re: dynamically partitioned stream
|
|
Thanks for your reply Tony, Yes we are in the latter case, where the functions/lambdas come in the control stream. Think of them as strings containing the logic of the function. The values for each of the arguments to the function come from the data stream. That is why we need to co-locate the data stream messages for the corresponding keys with the control message that has the function to be applied. We have a way of interpreting the logic described in the string and executing it on the incoming values from the data stream. This is kicked off from within the Flink runtime (synchronous to a flatMap of the RichCoFlatMapFunct So yeah I see your point about mapping the arguments but the problem is not really that, the problem is making sure that the values in the control stream are in the same instance of the task/ keyed managed state as a the actual control stream message. Once they are we can pass them in. Any other thoughts? M On Thu, Aug 31, 2017 at 12:06 PM, Tony Wei <[hidden email]> wrote:
|
|
|
Hi Martin, So the problem is that you want to group those arguments in Data Stream and pass them to the lambda function from Control Stream at the same time. Am I right? If right, then you could give each lambda function an id as well. Use these ids to tag those arguments to which they belong. After that, keyBy function could be used to group those arguments belonging to the same lambda function. Joining this stream with Control Stream by function id could make arguments and function be in the same instance. What do you think? Could this solution solve your problem? Best, Tony Wei 2017-08-31 20:43 GMT+08:00 Martin Eden <[hidden email]>:
|
Re: dynamically partitioned stream
|
|
Hi Martin,
In your original example, what does this syntax mean exactly: f1 [A, B, C] 1 Does it mean that f1 needs one A, one B and one C from the main stream? If yes, which ones, because there are multiple As and Bs and so on. Or does it mean that f1 can apply to an A or a B or a C? If it's the first, then I think it's quite hard to find a partitioning such that both f1, f2, and all A, B, and C go to the same machine. Best, Aljoscha
|
Re: dynamically partitioned stream
|
|
Hi Aljoscha, Tony,
Aljoscha: Yes it's the first option you mentioned. Yes, the stream has multiple values in flight for A, B, C. f1 needs to be applied each time a new value for either A, B or C comes in. So we need to use state to cache the latest values. So using the example data stream in my first msg the emitted stream should be: 1. Data Stream: KEY VALUE TIME . . . C V6 6 B V6 6 A V5 5 A V4 4 C V3 3 A V3 3 B V3 3 B V2 2 A V1 1 2. Control Stream: Lambda ArgumentKeys TIME . . . f2 [A, C] 4 f1 [A, B, C] 1 3. Expected emitted stream: TIME VALUE . . . 6 f1(V5, V6, V3) f1(V5, V6, V6) f2(V5, V6) 5 f1(V5, V3, V3) f2(V5, V3) 4 f1(V4, V3, V3) f2(V4, V3) 3 f1(V3, V3, V3) 2 - 1 - So essentially as soon as the argument list fills up then we apply the function/lambda at each new arriving message in the data stream for either argument key. Tony: Yes we need to group by and pass to the lambda. Ok, so what you are proposing might work. So your solution assumes that we have to connect with the control stream twice? Once for the tagging and another time re-connect-ing the control stream with the tagged stream for the actual application of the function/lambda? Thanks, Alex On Thu, Aug 31, 2017 at 2:57 PM, Aljoscha Krettek <[hidden email]> wrote:
|
|
|
Hi Martin, Yes, that is exactly what I thought. But the first step also needs to be fulfilled by SideInput. I'm not sure how to achieve this in the current release. Best, Tony Wei Martin Eden <[hidden email]>於 2017年8月31日 週四，下午11:32寫道：
|
Re: dynamically partitioned stream
|
|
This might be a way forward but since side inputs are not there I will try and key the control stream by the keys in the first co flat map. I'll see how it goes. Thanks guys, M On Thu, Aug 31, 2017 at 5:16 PM, Tony Wei <[hidden email]> wrote:
|
Re: dynamically partitioned stream
|
|
Hi Martin,
I think with those requirements this is very hard (or maybe impossible) to do efficiently in a distributed setting. It might be that I'm misunderstanding things but let's look at an example. Assume that initially, we don't have any lambdas, so data can be sent to any machine because it doesn't matter where they go. Now, we get a new lambda f2 [A, C]. Say this gets routed to machine "0", now this means that messages with key A and C also need to be router to machine "0". Now, we get a new lambda f1 [D, E], say this gets routed to machine "2", meaning that messages with key D and E are also routed to machine "2". Then, we get a new lambda f3 [C, D]. Do we now re-route all previous lambdas and inputs to different machines? They all have to go to the same machine, but which one? I'm currently thinking that there would need to be some component that does the routing, but this has to be global, so it's hard to do in a distributed setting. What do you think? Best, Aljoscha
|
|
|
Hi Martin, Aljoscha I think Aljoscha is right. My origin thought was to keep the state only after a lambda function coming. Use Aljoscha's scenario as example, initially, all data will be discarded because there is no any lambdas. When lambda f1 [D, E] and f2 [A, C] comes, A, C begin to be routed to machine "0" and D, E begin to be routed to machine "1". Then, when we get a new lambda f3 [C, D], we can duplicate C, D and route these copies to machine "2". However, after reading your example again, I found what you want is a whole picture for all variables' state in a global view, so that no matter what time a new lambda comes it can always get its variables' state immediately. In that case, I have the same opinion as Aljoscha. Best, Tony Wei 2017-09-01 23:59 GMT+08:00 Aljoscha Krettek <[hidden email]>:
|
Re: dynamically partitioned stream
|
|
Hi Aljoscha, Tony, We actually do not need all the keys to be on all nodes where lambdas are. We just need the keys that represent the data for the lambda arguments to be routed to the same node as the lambda, whichever one it might be. Essentially in the solution we emit the data multiple times and by doing that we roughly multiply the input rate by the average number of lambdas a key is a part of (X). In terms of memory this is O(X * N) where N is the number of keys int the data. N is the large bit. If X ~ N then we have O (N^2) complexity for the Flink state. And in that case yes I see your point about performance Aljoscha. But if X << N, as is our case, then we have O(N) which should be manageable by Flink's distributed state mechanism right? Do you see any gotchas in this new light? Are my assumptions correct? Thanks, M On Sat, Sep 2, 2017 at 3:38 AM, Tony Wei <[hidden email]> wrote:
|
|
|
Hi Martin, The performance is an issue, but in your case, yes, it might not be a problem if X << N. However, the other problem is where data should go in the beginning if there is no lambda been received. This problem doesn't associate with performance, but instead with correctness. If you want to keep the value state for the incoming lambda you should broadcast it to all nodes, because you would never know where the next lambda that requires this data would be routed to. Of course, you can send this data to a pre-defined node and route the lambda to this node, but this will lead to all data in the same node to let all lambda can get all required data. It is not a good solution because of a lack of scalability. In my origin thought, it is based on only storing state of data after you receive at least one lambda that requires it, so that data has its destination node to go. Can this assumption be acceptable in your case? What do you think? Best, Tony Wei 2017-09-06 22:41 GMT+08:00 Martin Eden <[hidden email]>:
|
Re: dynamically partitioned stream
|
|
Hi Tony, Yes exactly I am assuming the lambda emits a value only after it has been published to the control topic (t1) and at least 1 value arrives in the data topic for each of it's arguments. This will happen at a time t2 > t1. So yes, there is uncertainty with regards to when t2 will happen. Ideally t2 - t1 ~ 0 but for our use case it is fine. Is this the correctness that you are talking about? Do I have the right picture of what happens? Thanks M On Thu, Sep 7, 2017 at 3:11 AM, Tony Wei <[hidden email]> wrote:
|
|
|
Hi Martin, What I was talking is about how to store the arguments' state. In the example you explained your use case to Aljoscha. 4 f1(V4, V3, V3) f2(V4, V3) 3 f1(V3, V3, V3) 2 - 1 - You showed that when lambda f2 came, it would emit f2(V4, V3) immediately. However, the second argument (B, V3, 3) came before f2. You couldn't know how to route it at that time. If you didn't store this data's state, then f2(V4, V3) won't happen and the problem is easily. Otherwise, you had to route all data and all lambda to the same node to guarantee that every lambda won't lose any their arguments' state. Best, Tony Wei 2017-09-07 14:31 GMT+08:00 Martin Eden <[hidden email]>:
|
Re: dynamically partitioned stream
|
|
Hi Tony,
Ah I see. Yes you are right. What I was saying in my last message is that I relaxed that requirement after realising that it works how you just described it (and Aljoscha previously) and global state is not really feasible/possible. Here is a re-worked example. Please let me know if it makes sense? Basically f2 is only emitted at time 7 when we have the first values for A and C emitted after time 4 when f2 itself was emitted. 1. Data Stream: KEY VALUE TIME . . . C V6 7 B V6 7 A V5 6 A V4 5 C V3 3 A V3 3 B V3 3 B V2 2 A V1 1 2. Control Stream: Lambda ArgumentKeys TIME . . . f2 [A, C] 4 f1 [A, B, C] 0 3. Expected emitted stream: TIME VALUE . . . 7 f1(V5, V6, V3) f1(V5, V6, V6) f2(V5, V6) 6 f1(V5, V3, V3) 5 f1(V4, V3, V3) 4 - 3 f1(V3, V3, V3) - or f1(V1,V2,V3) if C,V3 arrives before A V3 or B V3 2 - 1 - 0 - I guess to wrap up this discussion there is only one more thing to clarify about this. There is some non-determinism as to what results are emitted. I have highlighted this in red above. Basically the example assumes the order of events perceived in the operators is as it in the incoming data stream and is maintained throughout the Flink dag, regardless of keyBys and flatMaps, connect etc. However if the order of the events is changed by Flink after the source receives and before it reaches the first flatMap1 in the first connected operator (i.e. the 3 events for time 3 are re-ordered) then the lambda emits different combinations. Questions: 1. Is what I am saying correct? Can Flink change the order of events in an input stream as they arrive to the input of flatMap in a connected stream? Maybe because of the key by? 2. Can this be remedied by using a different time dimension like event time or ingest time? Would that work for a simple connected streams? Does it need to be somehow combined with a windowed operator? To clarify, in my mind, in the example above the TIME column was really both event time and ingest time for the data stream and control stream (in the same order as written) and processing time for the emitted output stream. So the data stream was really this: KEY VALUE EV_TIME INGEST_TIME PROCESSING_TIME . . . C V3 3 33 43 A V3 2 32 42 B V3 1 31 41 Thanks, M On Thu, Sep 7, 2017 at 9:24 AM, Tony Wei <[hidden email]> wrote:
|
|
|
Hi Martin, For the first question, as far as I know, Flink guarantees that the order of records from the same sub-task of consumer won't be changed. If A, B and C came from different sub tasks, the result might be like your concern. After all, you can't have all sub tasks process in the same speed or synchronize them. For the second question, I think it is hard to remedy it, because sorting elements in a continuous stream without any constraint is impossible. Using window operator to buffer elements for a while and sort them is adding a constraint somehow. These are what I know so far. Maybe there are some mistakes. I think Aljoscha can explain more clearly than me. Best, Tony Wei. 2017-09-07 18:00 GMT+08:00 Martin Eden <[hidden email]>:
|
«
Return to (DEPRECATED) Apache Flink User Mailing List archive.
|
1 view|%1 views
| Free forum by Nabble | Edit this page |