Re: Change for submitting to yarn in 1.3.1

[Kevin Markey]

This is an excellent discussion.  As mentioned in an earlier email, we agree with a number of Chester's suggestions, but we have yet other concerns.  I've researched this further in the past several days, and I've queried my team.  This email attempts to capture those other concerns. Making yarn.Client private has prevented us from moving from Spark 1.0.x to Spark 1.2 or 1.3 despite many alluring new features. The SparkLauncher, which provides “support for programmatically running Spark jobs” (SPARK-3733 and SPARK-4924) will not work in our environment or for our use case -- which requires programmatically initiating and monitoring Spark jobs on Yarn in cluster mode from a cloud-based application server. It is not just that the Yarn ApplicationId is no longer directly or indirectly available. More critically, it violates constraints imposed by any application server and additional constraints imposed by security, process, and dynamic resource allocation requirements in our cloud services environment. In Spark 1.0 and 1.1, with yarn.Client public, our applications' job scheduler marshalls configuration and environmental resources necessary for any Spark job, including cluster-, data- or job-specific parameters, makes the appropriate calls to initialize and run yarn.Client, which together with the other classes in the spark-yarn module requests the Yarn resource manager to start and monitor a job (see Figure 1) on the cluster. (Our job scheduler is not Yarn replacement; it leverages Yarn to coordinate a variety of different Spark analytic and data enrichment jobs.) More recent Spark versions make yarn.Client private and thus remove that capability, but the SparkLauncher, scheduled for Spark 1.4, replaces this simple programmatic solution with one considerably more complicated. Based on our understanding, in this scenario, our job scheduler marshalls configuration and environmental resources for the SparkLauncher much as it did for yarn.Client. It then calls launch() to initialize a new Linux process to execute the spark-submit shell script with the specified configuration and environment, which in turn starts a new JVM (with the Spark assembly jar in its class path) that executes launcher.Main. This ultimately calls yarn.Client (see Figure 2). This is more than an arm's-length transaction. There are three legs: job scheduler SparkLauncher.launch() call → spark-submit bash execution → launcher.Main call to yarn.Client → Yarn resource manager allocation and execution of job driver and executors. Not only is this scenario unnecessarily complicated, it will simply not work. The “programmatic” call to SparkLauncher.launch() starts a new JVM, which is not allowed in any application server, which must own all its JVMs. Perhaps, spark-submit and the launcher.Main JVM process could be hosted outside the application server, but in violation of security and multiple-tenant cloud architectural constraints. We appreciate that yarn.Client was perhaps never intended to be public. Configuring it is not for the faint-of-heart, and some of its methods should indeed be private. We wonder whether there is another option. In researching and discussing these issues with Cloudera and others, we've been told that only one mechanism is supported for starting Spark jobs: the spark-submit scripts. We also have gathered (perhaps mistakenly) from discussions reaching back 20 months that Spark's intention is to have a unified job submission interface for all supported platforms. Unfortunately this doesn't recognize the asymmetries among those platforms. Submitting a local Spark job or a job to a Spark master in cluster mode may indeed require initializing a separate process in order to pass configuration parameters via the environment and command line. But Spark's yarn.Client in cluster mode already has an arm's length relationship with the Yarn resource manager. Configuration may be passed from the job scheduling application to yarn.Client as Strings or property map variables and method parameters. Our request is for a public yarn.Client or some reasonable facsimile. Thanks. On 05/13/2015 08:22 PM, Patrick Wendell wrote: Hey Chester, Thanks for sending this. It's very helpful to have this list. The reason we made the Client API private was that it was never intended to be used by third parties programmatically and we don't intend to support it in its current form as a stable API. We thought the fact that it was for internal use would be obvious since it accepts arguments as a string array of CL args. It was always intended for command line use and the stable API was the command line. When we migrated the Launcher library we figured we covered most of the use cases in the off chance someone was using the Client. It appears we regressed one feature which was a clean way to get the app ID. The items you list here 2-6 all seem like new feature requests rather than a regression caused by us making that API private. I think the way to move forward is for someone to design a proper long-term stable API for the things you mentioned here. That could either be by extension of the Launcher library. Marcelo would be natural to help with this effort since he was heavily involved in both YARN support and the launcher. So I'm curious to hear his opinion on how best to move forward. I do see how apps that run Spark would benefit of having a control plane for querying status, both on YARN and elsewhere. - Patrick On Wed, May 13, 2015 at 5:44 AM, Chester At Work <ches...@alpinenow.com> wrote: Patrick There are several things we need, some of them already mentioned in the mailing list before. I haven't looked at the SparkLauncher code, but here are few things we need from our perspectives for Spark Yarn Client 1) client should not be private ( unless alternative is provided) so we can call it directly. 2) we need a way to stop the running yarn app programmatically ( the PR is already submitted) 3) before we start the spark job, we should have a call back to the application, which will provide the yarn container capacity (number of cores and max memory ), so spark program will not set values beyond max values (PR submitted) 4) call back could be in form of yarn app listeners, which call back based on yarn status changes ( start, in progress, failure, complete etc), application can react based on these events in PR) 5) yarn client passing arguments to spark program in the form of main program, we had experience problems when we pass a very large argument due the length limit. For example, we use json to serialize the argument and encoded, then parse them as argument. For wide columns datasets, we will run into limit. Therefore, an alternative way of passing additional larger argument is needed. We are experimenting with passing the args via a established akka messaging channel. 6) spark yarn client in yarn-cluster mode right now is essentially a batch job with no communication once it launched. Need to establish the communication channel so that logs, errors, status updates, progress bars, execution stages etc can be displayed on the application side. We added an akka communication channel for this (working on PR ). Combined with others items in this list, we are able to redirect print and error statement to application log (outside of the hadoop cluster), so spark UI equivalent progress bar via spark listener. We can show yarn progress via yarn app listener before spark started; and status can be updated during job execution. We are also experimenting with long running job with additional spark commands and interactions via this channel. Chester Sent from my iPad On May 12, 2015, at 20:54, Patrick Wendell <pwend...@gmail.com> wrote: Hey Kevin and Ron, So is the main shortcoming of the launcher library the inability to get an app ID back from YARN? Or are there other issues here that fundamentally regress things for you. It seems like adding a way to get back the appID would be a reasonable addition to the launcher. - Patrick On Tue, May 12, 2015 at 12:51 PM, Marcelo Vanzin <van...@cloudera.com> wrote: On Tue, May 12, 2015 at 11:34 AM, Kevin Markey <kevin.mar...@oracle.com> wrote: I understand that SparkLauncher was supposed to address these issues, but it really doesn't. Yarn already provides indirection and an arm's length transaction for starting Spark on a cluster. The launcher introduces yet another layer of indirection and dissociates the Yarn Client from the application that launches it. Well, not fully. The launcher was supposed to solve "how to launch a Spark app programatically", but in the first version nothing was added to actually gather information about the running app. It's also limited in the way it works because of Spark's limitations (one context per JVM, etc). Still, adding things like this is something that is definitely in the scope for the launcher library; information such as app id can be useful for the code launching the app, not just in yarn mode. We just have to find a clean way to provide that information to the caller. I am still reading the newest code, and we are still researching options to move forward. If there are alternatives, we'd like to know. Super hacky, but if you launch Spark as a child process you could parse the stderr and get the app ID. -- Marcelo --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org For additional commands, e-mail: dev-h...@spark.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@spark.apache.org For additional commands, e-mail: dev-h...@spark.apache.org