Interesting! I agree with Luke that it seems not a great fit for Beam in the most rigorous sense. There are many considerations:
1. We assume ParDo has side effects by default. So the model actual *requires* eager evaluation, not lazy, in order to make all the side effects happen. But for your case let us assume somehow we know it is all @Pure. 2. Lazy evaluation and parallelism are in opposition. In pure computations like Haskell, literally everything (except monadic sequence) is parallel for free, but the problem is nothing starts until it is needed so parallelism requires forcing computations early. On the other hand, we can think about ways forward here. A first step is if the join is a "side lookup join" where we always process all of source 1 but try to process less of source 2. If source 2 is feeding into a map side input then this could be lazy in some way. When an element from source 1 calls the side input lookup it could be a blocking call that triggers reads from source 2 until a match is found. This computation strategy is consistent with the model and will read all of source 1 but only the prefix of source 2 needed to join all of source 1. I think you could implement this pattern with parallelism on both the main input and side input. Then, to read less of source 1 you need feedback from the sink to the source. We have nothing like that... This is all very abstract hypotheticals. If we get to practical implementation "today" then every runner pretty much reads all of a bounded source before even starting the next transform, no?. I wonder if it makes sense to convert them to unbounded (which is still allowed to terminate but does not support dynamic splits). Then you just terminate the pipeline when you have enough output. You will read more than you need but maybe that is not so bad, and anyhow hard to avoid. Also a vague idea... And I have to ask, though, can you build indices instead of brute force for the join? Kenn On Thu, Oct 10, 2019 at 10:47 AM Luke Cwik <lc...@google.com> wrote: > This doesn't seem like a good fit for Apache Beam but have you tried: > * using a StatefulDoFn that performs all the joining and signals the > service powering the sources to stop sending data once your criteria is met > (most services powering these sources won't have a way to be controlled > this way)? > * using a StatefulDoFn that performs all the joining and to write out the > data to the output directly and then shutdown the pipeline (you can't have > any transforms that are after the StatefulDoFn)? > > Both of these ideas remove a lot of the parallelism that Apache Beam > provides. > > > > On Thu, Oct 10, 2019 at 10:36 AM Alexey Romanenko < > aromanenko....@gmail.com> wrote: > >> Hello, >> >> We have a use case and it's not clear how it can be solved/implemented >> with Beam. I count on community help with this, maybe I miss something that >> lays on the surface. >> >> Let’s say, there are two different bounded sources and one join transform >> (say GBK) downstream. This Join transform is like INNER JOIN which joins >> elements of two collections only if they have common key (though, it could >> be any other join logic there, doesn’t matter). What matters is that this >> Join has to return only N records as output and then we have to stop >> pipeline after they have been processed. It means that, in the best case, >> we need to read only N records from every source, join them and move >> downstream and after pipeline should be stopped. In other cases, if some >> records don’t have common key in other collection, we need to read another >> bunch of records and see if it would be enough to have N joined records >> after Join. >> >> Below, here is a simple example of this. Say, every source contains 1M of >> records but after Join we need to have only 1K of joined records. So, we >> don’t want to read all two millions from 2 sources in case if we can have >> an output after reading much less records in the end. So, 1K of joined >> records is a stop condition. >> >> 1M >> ————— >> | Source 1 |———— >> ————— | ——— >> |———> | Join |———> Output 1K and stop >> 1M | ——— >> ————— | >> | Source 2 |———— >> ————— >> >> So, it looks like I need to have ability to read new portion of data "on >> demand” or like to have a back pressure mechanizm which signals from >> downstream to upstream that “please, give me only N elements and then wait >> until I ask for more”. I’m not sure that Beam supports something like this. >> >> As an idea, I was trying to split initial inputs into fixed Windows with >> trigger “AfterPane.elementCountAtLeast(N)” to read data by limited batches >> and use another “AfterPane.elementCountAtLeast(N)” after Join which should >> trigger only once. It doesn’t work and still, it won’t read data “on >> demand” and stop the whole pipeline, I guess. >> >> Do you think it can be feasible to do in Beam? >> Any ideas or advices are very welcomed! >> >> >>
