Hi Alexis, I'm not sure what "watermark" step refers to in you graph, but in general I'd say your intuition is correct.
For the "buffering" part, each sub-task needs to send data via data exchange (last operator in chain) has an output buffer and the operator that consumes this data (maybe on different machine) has an input buffer (buffer de-bloating feature can help to mitigate excessive buffering in case of back-pressure). but I’m not sure if this actually happens > How are you trying to verify this? Also can you check whether the operators are not back-pressured? Best, D. On Thu, Dec 2, 2021 at 12:27 PM Alexis Sarda-Espinosa < alexis.sarda-espin...@microfocus.com> wrote: > Hello, > > > > I have a use case with event-time processing that ends up with a DAG > roughly like this: > > > > source -> filter -> keyBy -> watermark -> window -> process -> keyBy_2 -> > connect (KeyedCoProcessFunction) -> sink > > | > / > > (side output) -> keyBy -> watermark > --------------------------------/ > > > > > > (In case the text gets mangled in the e-mail, the side output comes from > the filter and joins back with the connect operation) > > > > The filter takes all data and its main output is all _*valid*_ data with > state "open"; the side output is all _*valid*_ data regardless of state. > > > > The goal of the KeyedCoProcessFunction is to check the results of the > (sliding) window. The window only receives open states, but > KeyedCoProcessFunction receives all valid data and should discard results > from the main stream if states changed from "open" to something else before > the window was evaluated. > > > > I would have expected all data from the side output to be processed > roughly immediately in KeyedCoProcessFunction’s processElement2 because > there’s no windowing in the side stream, but I’m not sure if this actually > happens, maybe the side stream (or both streams) buffers some data before > passing it to the connected stream? If yes, is there any way I could tune > this? > > > > Regards, > > Alexis. > > >
