If we decide for Akka, then we can choose the language to use. But with both bindings (Java and Scala), we would add a Scala dependency to the project, because Akka is implemented in Scala.
On Sun, Aug 31, 2014 at 5:44 PM, Stephan Ewen <[email protected]> wrote: > Here is one more: Akka has facilities that help creating hot standby > actors, that help with eliminating the JobManager as the > single-point-of-failure. > > > On Sun, Aug 31, 2014 at 5:43 PM, Stephan Ewen <[email protected]> wrote: > > > Hi! > > > > The Java vs Scala discussion is orthogonal to the actors discussion. We > > can use Akka actors in Java. And I think that makes a lot of sense, for > the > > reasons that Till mentioned, plus the following reasons: > > > > - Akka has made a lot of effort to combine message throughput (multiple > > actor calls in one message) with low message latency. I don't think we > > could do much better with something else > > > > - I am currently working on the ExecutionGraph and Scheduler to unify > > lazy computation / recovery / dynamic resource assignment. > > The Actor paradigm (order of calls, queuing invocations in the actor > > mailboxes) makes it much simpler to get concurrent situations right (such > > as certain calls overtaking each other, like for example deploy/cancel, > etc) > > > > - Actors work with thread pools be themselves, so we can get rid of all > > the inner runnables sent to executor services. makes the code much more > > readable > > > > Stephan > > > > > > > > On Sun, Aug 31, 2014 at 1:31 PM, Till Rohrmann <[email protected]> > > wrote: > > > >> Hi Daniel, > >> > >> the RPC rework is discussed in > >> https://issues.apache.org/jira/browse/FLINK-1019. Jira is currently > down > >> due to maintenance reasons. > >> > >> The ideas to use akka are the following. Akka allows us to reduce the > code > >> base which has to be maintained. Especially, we get rid of all the > >> multi-threading programming of the rpc service which is always hard to > >> work > >> with. With Akka we would get the heartbeat signal for free, because Akka > >> can detect dead actors. Akka uses supervision to handle fault tolerance > as > >> well as recovery and it allows an easy forwarding of remote exceptions. > At > >> the same time it offers a nice rpc abstraction which easily allows to > >> implement asynchronous services. Furthermore, it scales rather well to > >> large numbers of nodes and hopefully we get the latencies of Flink a > >> little > >> bit down. > >> > >> Bests, > >> > >> Till > >> > >> > >> On Sun, Aug 31, 2014 at 11:35 AM, Daniel Warneke <[email protected]> > >> wrote: > >> > >> > Hi, > >> > > >> > will akka just be used for RPC or are there any plans to expand the > >> > actor-based model to further parts of the runtime system? If so, could > >> you > >> > please point me to the discussion thread? > >> > > >> > Spontaneously, I would say that adding a hard dependency on Scala just > >> for > >> > the sake of having a hip RPC service sounds like a pretty dodgy deal. > >> > Therefore, I would like understand how much value akka could bring to > >> Flink > >> > in the long run. The discussion whether to reimplement core components > >> of > >> > the system in Scala should be the second step in my opinion. > >> > > >> > Bests, > >> > > >> > Daniel > >> > > >> > > >> > Am 29.08.2014 11:33, schrieb Asterios Katsifodimos: > >> > > >> > I agree that using Akka's actors from Java results in very ugly code. > >> >> Hiding the internals of Akka behind Java reflection looks better but > >> >> breaks > >> >> the principles of actors. For me it is kind of a deal breaker for > using > >> >> Akka from Java. I think that Till has more reasons to believe that > >> Scala > >> >> would be a more appropriate for building a new Job/Task Manager. > >> >> > >> >> I think that this discussion should focus on 4 main aspects: > >> >> 1. Performance > >> >> 2. Implementability > >> >> 3. Maintainability > >> >> 4. Available Tools > >> >> > >> >> 1. Performance: Since that the job of the JobManager and the > >> TaskManager > >> >> is > >> >> to 1) exchange messages in order to maintain a distributed state > >> machine > >> >> and 2) setup connections between task managers, 3) detect failures > >> etc.. > >> >> In > >> >> these basic operations, performance should not be an issue. Akka was > >> >> proven > >> >> to scale quite well with very low latency. I guess that the low level > >> >> "plumbing" (serialization, connections, etc.) will continue in Java > in > >> >> order to guarantee high performance. I have no clue on what's > happening > >> >> with memory management and whether this will be implemented in Java > or > >> >> Scala and the respective consequences. Please comment. > >> >> > >> >> 2. Since the Job/Task Manager is going to be essentially implemented > >> from > >> >> scratch, given the power of Akka, it seems to me that the > >> implementation > >> >> will be easier, shorter and less verbose in Scala, given that Till > is > >> >> comfortable enough with Scala. > >> >> > >> >> 3. Given #2, maintaining the code and trying out new ideas in Scala > >> would > >> >> take less time and effort. But maintaining low level plumbing in Java > >> and > >> >> high level logic in Scala scares me. Anyone that has done this before > >> >> could > >> >> comment on this? > >> >> > >> >> 4. Tools: Robert has raised some issues already but I think that > tools > >> >> will > >> >> get better with time. > >> >> > >> >> Given the above, I would focus on #3 to be honest. Apart from this, > >> going > >> >> the Scala way sounds like a great idea. I really second Kostas' > opinion > >> >> that if large changes are going to happen, this is the best moment. > >> >> > >> >> Cheers, > >> >> Asterios > >> >> > >> >> > >> >> > >> >> On Fri, Aug 29, 2014 at 1:02 AM, Till Rohrmann < > >> [email protected]> > >> >> wrote: > >> >> > >> >> I also agree with Robert and Kostas that it has to be a community > >> >>> decision. > >> >>> I understand the problems with Eclipse and the Scala IDE which is a > >> pain > >> >>> in > >> >>> the ass. But eventually these things will be fixed. Maybe we could > >> also > >> >>> talk to the typesafe guy and tell him that this problem bothers us a > >> lot. > >> >>> > >> >>> I also believe that the project is not about a specific programming > >> >>> language but a problem we want to tackle with Flink. From time to > >> time it > >> >>> might be necessary to adapt the tools in order to reach the goal. In > >> >>> fact, > >> >>> I don't believe that Scala parts would drive people away from the > >> >>> project. > >> >>> Instead, Scala enthusiasts would be motivated to join us. > >> >>> > >> >>> Actually I stumbled across a quote of Leibniz which put's my point > of > >> >>> view > >> >>> quite accurately in a nutshell: > >> >>> > >> >>> In symbols one observes an advantage in discovery which is greatest > >> when > >> >>> they express the exact nature of a thing briefly and, as it were, > >> picture > >> >>> it; then indeed the labor of thought is wonderfully diminished -- > >> >>> Gottfried > >> >>> Wilhelm Leibniz > >> >>> > >> >>> > >> >>> On Thu, Aug 28, 2014 at 12:57 PM, Kostas Tzoumas < > [email protected] > >> > > >> >>> wrote: > >> >>> > >> >>> On Thu, Aug 28, 2014 at 11:49 AM, Robert Metzger < > >> [email protected]> > >> >>>> wrote: > >> >>>> > >> >>>> Changing the programming language of a very important system > >> component > >> >>>>> > >> >>>> is > >> >>> > >> >>>> something we should carefully discuss. > >> >>>>> > >> >>>>> Definitely agree, I think the community should discuss this very > >> >>>> > >> >>> carefully. > >> >>> > >> >>>> > >> >>>> I understand that Akka is written in Scala and that it will be > much > >> >>>>> > >> >>>> more > >> >>> > >> >>>> natural to implement the actor based system using Scala. > >> >>>>> I see the following issues that we should consider: > >> >>>>> Until now, Flink is clearly a project implemented only in Java. > The > >> >>>>> > >> >>>> Scala > >> >>> > >> >>>> API basically sits on top of the Java-based runtime. We do not > really > >> >>>>> depend on Scala (we could easily remove the Scala API if we want > >> to). > >> >>>>> Having code written in Scala in the main system will add a hard > >> >>>>> > >> >>>> dependency > >> >>>> > >> >>>>> on a scala version. > >> >>>>> Being a pure Java project has some advantages: I think its a fact > >> that > >> >>>>> there are more Java programmers than Scala programmers. So our > >> chances > >> >>>>> > >> >>>> of > >> >>> > >> >>>> attracting new contributors are higher when being a Java project. > >> >>>>> On the other hand, we could maybe attract Scala developers to our > >> >>>>> > >> >>>> project. > >> >>>> > >> >>>>> But that has not happened (for contributors, not users!) so far > for > >> our > >> >>>>> Scala API, so I don't see any reason for that to happen. > >> >>>>> > >> >>>>> > >> >>>>> This is definitely an issue to consider. We need to carefully > >> weight > >> >>>> how > >> >>>> important this issue is. If we want to break things, incubation is > >> the > >> >>>> right time to do it. Below are some arguments in favor of breaking > >> >>>> > >> >>> things, > >> >>> > >> >>>> but do keep in mind that I am undecided, and I would really like to > >> see > >> >>>> > >> >>> the > >> >>> > >> >>>> community weighing in. > >> >>>> > >> >>>> First, I would dare say that the primary reason for someone to > >> >>>> contribute > >> >>>> to Flink so far has not been that the code is written in Java, but > >> more > >> >>>> > >> >>> the > >> >>> > >> >>>> content and nature of the project. Most contributors are Big Data > >> >>>> enthusiasts in some way or another. > >> >>>> > >> >>>> Second, Scala projects have attracted contributors in the past. > >> >>>> > >> >>>> Third, it should not be too hard for someone that does not know > >> Scala to > >> >>>> contribute to a different component if the interfaces are clear. > >> >>>> > >> >>>> > >> >>>> Another issue is tooling: There are a lot of problems with Scala > and > >> >>>>> Eclipse: I've recently switched to Eclipse Luna. It seems to be > >> >>>>> > >> >>>> impossible > >> >>>> > >> >>>>> to compile Scala code with Luna because ScalaIDE does not properly > >> cope > >> >>>>> with it. > >> >>>>> Even with Eclipse versions that are supported by ScalaIDE, you > have > >> to > >> >>>>> manually install 3 plugins, some of them are not available in the > >> >>>>> > >> >>>> Eclipse > >> >>> > >> >>>> Marketplace. So with a JobManager written in Scala, users can not > >> just > >> >>>>> import our project as a Maven project into Eclipse and start > >> >>>>> > >> >>>> developing. > >> >>> > >> >>>> The support for Maven is probably also limited. For example, I > don't > >> >>>>> > >> >>>> know > >> >>> > >> >>>> if there is a checkstyle plugin for Scala. > >> >>>>> > >> >>>>> I'm looking forward to hearing other opinions on this issue. As I > >> said > >> >>>>> > >> >>>> in > >> >>> > >> >>>> the beginning, we should exchange arguments on this and think about > >> it > >> >>>>> > >> >>>> for > >> >>>> > >> >>>>> some time before we decide on this. > >> >>>>> > >> >>>>> Best, > >> >>>> > >> >>>>> Robert > >> >>>>> > >> >>>>> > >> >>>>> > >> >>>>> On Thu, Aug 28, 2014 at 1:08 AM, Till Rohrmann < > >> [email protected]> > >> >>>>> wrote: > >> >>>>> > >> >>>>> Hi guys, > >> >>>>>> > >> >>>>>> I currently working on replacing the old rpc infrastructure with > an > >> >>>>>> > >> >>>>> akka > >> >>>> > >> >>>>> based actor system. In the wake of this change I will reimplement > >> the > >> >>>>>> JobManager and TaskManager which will then be actors. Akka > offers a > >> >>>>>> > >> >>>>> Java > >> >>>> > >> >>>>> API but the implementation turns out to be very verbose and > >> >>>>>> > >> >>>>> laborious, > >> >>> > >> >>>> because Java 6 and 7 do not support lambdas and pattern matching. > >> >>>>>> > >> >>>>> Using > >> >>> > >> >>>> Scala instead, would allow a far more succinct and clear > >> >>>>>> > >> >>>>> implementation > >> >>> > >> >>>> of > >> >>>>> > >> >>>>>> the JobManager and TaskManager. Instead of a lot of if statements > >> >>>>>> > >> >>>>> using > >> >>> > >> >>>> instanceof to figure out the message type, we could simply use > >> >>>>>> > >> >>>>> pattern > >> >>> > >> >>>> matching. Furthermore, the callback functions could simply be > Scala's > >> >>>>>> anonymous functions. Therefore I would propose to use Scala for > >> these > >> >>>>>> > >> >>>>> two > >> >>>> > >> >>>>> systems. > >> >>>>>> > >> >>>>>> The Akka system uses the slf4j library as logging interface. > >> >>>>>> > >> >>>>> Therefore > >> >>> > >> >>>> I > >> >>>> > >> >>>>> would also propose to replace the jcl logging system with the > slf4j > >> >>>>>> > >> >>>>> logging > >> >>>>> > >> >>>>>> system. Since we want to use Akka in many parts of the runtime > >> system > >> >>>>>> > >> >>>>> and > >> >>>> > >> >>>>> it recommends using logback as logging backend, I would also like > to > >> >>>>>> replace log4j with logback. But this change should inflict only > few > >> >>>>>> > >> >>>>> changes > >> >>>>> > >> >>>>>> once we established the slf4j logging interface everywhere. > >> >>>>>> > >> >>>>>> What do you guys think of that idea? > >> >>>>>> > >> >>>>>> Best regards, > >> >>>>>> > >> >>>>>> Till > >> >>>>>> > >> >>>>>> > >> > > >> > > > > >
