Thanks for the in depth explanation. These methods would require us to architect our Server around Spark and it is actually designed to be independent of the ML implementation. SparkML is an important algo source, to be sure, but so is TensorFlow, and Python non-spark libs among others. So Spark stays at arms length in a microservices pattern. Doing this with access to Job status and management is why Livy and the (Spark) Job Server exist. To us the ideal is treating Spark like a compute server that will respond to a service API for job submittal and control.
None of the above is solved by k8s Spark. Further we find that the Spark Programatic API does not support deploy mode = “cluster”. This means we have to take a simple part of our code and partition it into new Jars only to get spark-submit to work. To help with Job tracking and management when you are not using the Programatic API we look to Livy. I guess if you ask our opinion of spark-submit, we’d (selfishly) say it hides architectural issues that should be solved in the Spark Programatic API but the popularity of spark-submit is causing the community to avoid these or just not see or care about them. I guess we’ll see if Spark behind Livy gives us what we want. Maybe this is unusual but we see Spark as a service, not an integral platform. We also see Kubernetes as very important but optional for HA or when you want to scale horizontally, basically when vertical is not sufficient. Vertical scaling is more cost effective so Docker Compose is a nice solution for simpler, Kubernetes-less deployments. So if we are agnostic about the job master, and communicate through Livy, we are back to orchestrating services with Docker and Kubernetes. If k8s becomes a super duper job master, great! But it doesn’t solve todays question. From: Matt Cheah <mch...@palantir.com> <mch...@palantir.com> Reply: Matt Cheah <mch...@palantir.com> <mch...@palantir.com> Date: July 1, 2019 at 5:14:05 PM To: Pat Ferrel <p...@occamsmachete.com> <p...@occamsmachete.com>, user@spark.apache.org <user@spark.apache.org> <user@spark.apache.org> Subject: Re: k8s orchestrating Spark service > We’d like to deploy Spark Workers/Executors and Master (whatever master is easiest to talk about since we really don’t care) in pods as we do with the other services we use. Replace Spark Master with k8s if you insist. How do the executors get deployed? When running Spark against Kubernetes natively, the Spark library handles requesting executors from the API server. So presumably one would only need to know how to start the driver in the cluster – maybe spark-operator, spark-submit, or just starting the pod and making a Spark context in client mode with the right parameters. From there, the Spark scheduler code knows how to interface with the API server and request executor pods according to the resource requests configured in the app. > We have a machine Learning Server. It submits various jobs through the Spark Scala API. The Server is run in a pod deployed from a chart by k8s. It later uses the Spark API to submit jobs. I guess we find spark-submit to be a roadblock to our use of Spark and the k8s support is fine but how do you run our Driver and Executors considering that the Driver is part of the Server process? It depends on how the server runs the jobs: - If each job is meant to be a separate forked driver pod / process: The ML server code can use the SparkLauncher API <https://spark.apache.org/docs/latest/api/java/org/apache/spark/launcher/SparkLauncher.html> and configure the Spark driver through that API. Set the master to point to the Kubernetes API server and set the parameters for credentials according to your setup. SparkLauncher is a thin layer on top of spark-submit; a Spark distribution has to be packaged with the ML server image and SparkLauncher would point to the spark-submit script in said distribution. - If all jobs run inside the same driver, that being the ML server: One has to start the ML server with the right parameters to point to the Kubernetes master. Since the ML server is a driver, one has the option to use spark-submit or SparkLauncher to deploy the ML server itself. Alternatively one can use a custom script to start the ML server, then the ML server process has to create a SparkContext object parameterized against the Kubernetes server in question. I hope this helps! -Matt Cheah *From: *Pat Ferrel <p...@occamsmachete.com> *Date: *Monday, July 1, 2019 at 5:05 PM *To: *"user@spark.apache.org" <user@spark.apache.org>, Matt Cheah < mch...@palantir.com> *Subject: *Re: k8s orchestrating Spark service We have a machine Learning Server. It submits various jobs through the Spark Scala API. The Server is run in a pod deployed from a chart by k8s. It later uses the Spark API to submit jobs. I guess we find spark-submit to be a roadblock to our use of Spark and the k8s support is fine but how do you run our Driver and Executors considering that the Driver is part of the Server process? Maybe we are talking past each other with some mistaken assumptions (on my part perhaps). From: Pat Ferrel <p...@occamsmachete.com> <p...@occamsmachete.com> Reply: Pat Ferrel <p...@occamsmachete.com> <p...@occamsmachete.com> Date: July 1, 2019 at 4:57:20 PM To: user@spark.apache.org <user@spark.apache.org> <user@spark.apache.org>, Matt Cheah <mch...@palantir.com> <mch...@palantir.com> Subject: Re: k8s orchestrating Spark service k8s as master would be nice but doesn’t solve the problem of running the full cluster and is an orthogonal issue. We’d like to deploy Spark Workers/Executors and Master (whatever master is easiest to talk about since we really don’t care) in pods as we do with the other services we use. Replace Spark Master with k8s if you insist. How do the executors get deployed? We have our own containers that almost work for 2.3.3. We have used this before with older Spark so we are reasonably sure it makes sense. We just wonder if our own image builds and charts are the best starting point. Does anyone have something they like? From: Matt Cheah <mch...@palantir.com> <mch...@palantir.com> Reply: Matt Cheah <mch...@palantir.com> <mch...@palantir.com> Date: July 1, 2019 at 4:45:55 PM To: Pat Ferrel <p...@occamsmachete.com> <p...@occamsmachete.com>, user@spark.apache.org <user@spark.apache.org> <user@spark.apache.org> Subject: Re: k8s orchestrating Spark service Sorry, I don’t quite follow – why use the Spark standalone cluster as an in-between layer when one can just deploy the Spark application directly inside the Helm chart? I’m curious as to what the use case is, since I’m wondering if there’s something we can improve with respect to the native integration with Kubernetes here. Deploying on Spark standalone mode in Kubernetes is, to my understanding, meant to be superseded by the native integration introduced in Spark 2.4. *From: *Pat Ferrel <p...@occamsmachete.com> *Date: *Monday, July 1, 2019 at 4:40 PM *To: *"user@spark.apache.org" <user@spark.apache.org>, Matt Cheah < mch...@palantir.com> *Subject: *Re: k8s orchestrating Spark service Thanks Matt, Actually I can’t use spark-submit. We submit the Driver programmatically through the API. But this is not the issue and using k8s as the master is also not the issue though you may be right about it being easier, it doesn’t quite get to the heart. We want to orchestrate a bunch of services including Spark. The rest work, we are asking if anyone has seen a good starting point for adding Spark as a k8s managed service. From: Matt Cheah <mch...@palantir.com> <mch...@palantir.com> Reply: Matt Cheah <mch...@palantir.com> <mch...@palantir.com> Date: July 1, 2019 at 3:26:20 PM To: Pat Ferrel <p...@occamsmachete.com> <p...@occamsmachete.com>, user@spark.apache.org <user@spark.apache.org> <user@spark.apache.org> Subject: Re: k8s orchestrating Spark service I would recommend looking into Spark’s native support for running on Kubernetes. One can just start the application against Kubernetes directly using spark-submit in cluster mode or starting the Spark context with the right parameters in client mode. See https://spark.apache.org/docs/latest/running-on-kubernetes.html [spark.apache.org] <https://urldefense.proofpoint.com/v2/url?u=https-3A__spark.apache.org_docs_latest_running-2Don-2Dkubernetes.html&d=DwMFaQ&c=izlc9mHr637UR4lpLEZLFFS3Vn2UXBrZ4tFb6oOnmz8&r=hzwIMNQ9E99EMYGuqHI0kXhVbvX3nU3OSDadUnJxjAs&m=4XyH4cxucBNQAlSaHyR4gXJbHIo9g9vcur4VzBCYkwk&s=Q6mv_pZUq3UmxJU6EiJYJvG8L44WBeWJyAnw3vG0GBw&e=> I would think that building Helm around this architecture of running Spark applications would be easier than running a Spark standalone cluster. But admittedly I’m not very familiar with the Helm technology – we just use spark-submit. -Matt Cheah *From: *Pat Ferrel <p...@occamsmachete.com> *Date: *Sunday, June 30, 2019 at 12:55 PM *To: *"user@spark.apache.org" <user@spark.apache.org> *Subject: *k8s orchestrating Spark service We're trying to setup a system that includes Spark. The rest of the services have good Docker containers and Helm charts to start from. Spark on the other hand is proving difficult. We forked a container and have tried to create our own chart but are having several problems with this. So back to the community… Can anyone recommend a Docker Container + Helm Chart for use with Kubernetes to orchestrate: - Spark standalone Master - several Spark Workers/Executors This not a request to use k8s to orchestrate Spark Jobs, but the service cluster itself. Thanks
