By the way, currently Dispatcher implements RestfulGateway and delegate resource request to ResourceManager. If we can, semantically, let WebMonitor implement RestfulGateway, and delegate job request to Dispatcher, resource request to ResourceManager, it seems reasonable that when WebMonitor receives a JarRun request, it spawns a process and run the main method of the main class of that jar.
Best, tison. Zili Chen <wander4...@gmail.com> 于2019年7月31日周三 下午7:10写道: > I don't think the `Program` interface could solve the problem. > > The launcher launches the job server which creates the job graph, > submits it and keeps monitoring. Even if user program implement > `Program` Flink still extracts the JobGraph from `getPlan` and > submits it, instead of really execute codes in main method of > user program, so that the launcher is not started. > > @Thomas, > > Here is an ongoing discussion on client refactoring[1] as Till > mentioned. However, I'm afraid that with current jar run semantic, > i.e., extract the job graph and submit it to the Dispatcher, it cannot > fits your requirement. The problem is that REST API directly > communicates with Dispatcher and thus it's strange to tell the > Dispatcher "just run a program in a process". > > As you mentioned in the document, with CLI in session mode the > whole program would be executed sequentially. I'll appreciate it > if you can participant the thread on client refactor[1]. In the > design document[2], we propose to provide rich interfaces for > downstream projects integration. You can customize your CLI for > executing your program arbitrarily. Any requirement or advise > would be help. > > Best, > tison. > > [1] > https://lists.apache.org/thread.html/ce99cba4a10b9dc40eb729d39910f315ae41d80ec74f09a356c73938@%3Cdev.flink.apache.org%3E > [2] > https://docs.google.com/document/d/1UWJE7eYWiMuZewBKS0YmdVO2LUTqXPd6-pbOCof9ddY/edit > > > > > Till Rohrmann <trohrm...@apache.org> 于2019年7月31日周三 下午4:50写道: > >> Are you looking for something similar to the `Program` interface? This >> interface, even though it is a bit outdated and might get removed in the >> future, offers a `getPlan` method which is called in order to generate the >> `JobGraph`. In the client refactoring discussion thread it is currently >> being discussed what to do with this interface. >> >> Cheers, >> Till >> >> On Wed, Jul 31, 2019 at 10:41 AM Chesnay Schepler <ches...@apache.org> >> wrote: >> >> > Couldn't the beam job server use the same work-around we're using in the >> > JarRunHandler to get access to the JobGraph? >> > >> > On 26/07/2019 17:38, Thomas Weise wrote: >> > > Hi Till, >> > > >> > > Thanks for taking a look! >> > > >> > > The Beam job server does not currently have the ability to just output >> > the >> > > job graph (and related artifacts) that could then be used with the >> > > JobSubmitHandler. It is itself using StreamExecutionEnvironment, >> which in >> > > turn will lead to a REST API submission. >> > > >> > > Here I'm looking at what happens before the Beam job server gets >> > involved: >> > > the interaction of the k8s operator with the Flink deployment. The jar >> > run >> > > endpoint (ignoring the current handler implementation) is generic and >> > > pretty much exactly matches what we would need for a uniform entry >> point. >> > > It's just that in the Beam case the jar file would itself be a >> "launcher" >> > > that doesn't provide the job graph itself, but the dependencies and >> > > mechanism to invoke the actual client. >> > > >> > > I could accomplish what I'm looking for by creating a separate REST >> > > endpoint that looks almost the same. But I would prefer to reuse the >> > Flink >> > > REST API interaction that is already implemented for the Flink Java >> jobs >> > to >> > > reduce the complexity of the deployment. >> > > >> > > Thomas >> > > >> > > >> > > >> > > >> > > On Fri, Jul 26, 2019 at 2:29 AM Till Rohrmann <trohrm...@apache.org> >> > wrote: >> > > >> > >> Hi Thomas, >> > >> >> > >> quick question: Why do you wanna use the JarRunHandler? If another >> > process >> > >> is building the JobGraph, then one could use the JobSubmitHandler >> which >> > >> expects a JobGraph and then starts executing it. >> > >> >> > >> Cheers, >> > >> Till >> > >> >> > >> On Thu, Jul 25, 2019 at 7:45 PM Thomas Weise <t...@apache.org> wrote: >> > >> >> > >>> Hi, >> > >>> >> > >>> While considering different options to launch Beam jobs through the >> > Flink >> > >>> REST API, I noticed that the implementation of JarRunHandler places >> > >> quite a >> > >>> few restrictions on how the entry point shall construct a Flink >> job, by >> > >>> extracting and manipulating the job graph. >> > >>> >> > >>> That's normally not a problem for Flink Java programs, but in the >> > >> scenario >> > >>> I'm looking at, the job graph would be constructed by a different >> > process >> > >>> and isn't available to the REST handler. Instead, I would like to be >> > able >> > >>> to just respond with the job ID of the already launched job. >> > >>> >> > >>> For context, please see: >> > >>> >> > >>> >> > >>> >> > >> >> > >> https://docs.google.com/document/d/1z3LNrRtr8kkiFHonZ5JJM_L4NWNBBNcqRc_yAf6G0VI/edit#heading=h.fh2f571kms4d >> > >>> The current JarRunHandler code is here: >> > >>> >> > >>> >> > >>> >> > >> >> > >> https://github.com/apache/flink/blob/f3c5dd960ff81a022ece2391ed3aee86080a352a/flink-runtime-web/src/main/java/org/apache/flink/runtime/webmonitor/handlers/JarRunHandler.java#L82 >> > >>> It would be nice if there was an option to delegate the >> responsibility >> > >> for >> > >>> job submission to the user code / entry point. That would be useful >> for >> > >>> Beam and other frameworks built on top of Flink that dynamically >> > create a >> > >>> job graph from a different representation. >> > >>> >> > >>> Possible ways to get there: >> > >>> >> > >>> * an interface that the main class can be implement end when >> present, >> > the >> > >>> jar run handler calls instead of main. >> > >>> >> > >>> * an annotated method >> > >>> >> > >>> Either way query parameters like savepoint path and parallelism >> would >> > be >> > >>> forwarded to the user code and the result would be the ID of the >> > launched >> > >>> job. >> > >>> >> > >>> Thougths? >> > >>> >> > >>> Thanks, >> > >>> Thomas >> > >>> >> > >> > >> >