Good question. However, we ought to look at what options we have so to speak.
Let us consider Spark on Dataproc, Spark on Kubernetes and Spark on Dataflow Spark on DataProc <https://cloud.google.com/dataproc> is proven and it is in use at many organizations, I have deployed it extensively. It is infrastructure as a service provided including Spark, Hadoop and other artefacts. You have to manage cluster creation, automate cluster creation and tear down, submitting jobs etc. However, it is another stack that needs to be managed. It now has autoscaling <https://cloud.google.com/dataproc/docs/concepts/configuring-clusters/autoscaling> (enables cluster worker VM autoscaling ) policy as well. Spark on GKE <https://cloud.google.com/kubernetes-engine/docs/concepts/autopilot-overview> is something newer. Worth adding that the Spark DEV team are working hard to improve the performance of Spark on Kubernetes, for example, through Support for Customized Kubernetes Scheduler <https://docs.google.com/document/d/1xgQGRpaHQX6-QH_J9YV2C2Dh6RpXefUpLM7KGkzL6Fg>. As I explained in the first thread, Spark on Kubernetes relies on containerisation. Containers make applications more portable. Moreover, they simplify the packaging of dependencies, especially with PySpark and enable repeatable and reliable build workflows which is cost effective. They also reduce the overall devops load and allow one to iterate on the code faster. From a purely cost perspective it would be cheaper with Docker *as you can share resources* with your other services. You can create Spark docker with different versions of Spark, Scala, Java, OS etc. That docker file is portable. Can be used on Prem, AWS, GCP etc in container registries and devops and data science people can share it as well. Built once used by many. Kubernetes with autopilo <https://cloud.google.com/kubernetes-engine/docs/concepts/autopilot-overview#:~:text=Autopilot%20is%20a%20new%20mode,and%20yield%20higher%20workload%20availability.>t helps scale the nodes of the Kubernetes cluster depending on the load. *That is what I am currently looking into*. With regard to Dataflow <https://cloud.google.com/dataflow/docs>, which I believe is similar to AWS Glue <https://aws.amazon.com/glue/?whats-new-cards.sort-by=item.additionalFields.postDateTime&whats-new-cards.sort-order=desc>, it is a managed service for executing data processing patterns. Patterns or pipelines are built with the Apache Beam SDK <https://beam.apache.org/documentation/runners/spark/>, which is an open source programming model that supports Java, Python and GO. It enables batch and streaming pipelines. You create your pipelines with an Apache Beam program and then run them on the Dataflow service. The Apache Spark Runner <https://beam.apache.org/documentation/runners/spark/#:~:text=The%20Apache%20Spark%20Runner%20can,Beam%20pipelines%20using%20Apache%20Spark.&text=The%20Spark%20Runner%20executes%20Beam,same%20security%20features%20Spark%20provides.> can be used to execute Beam pipelines using Spark. When you run a job on Dataflow, it spins up a cluster of virtual machines, distributes the tasks in the job to the VMs, and dynamically scales the cluster based on how the job is performing. As I understand both iterative processing and notebooks plus Machine learning with Spark ML are not currently supported by Dataflow So we have three choices here. If you are migrating from on-prem Hadoop/spark/YARN set-up, you may go for Dataproc which will provide the same look and feel. If you want to use microservices and containers in your event driven architecture, you can adopt docker images that run on Kubernetes clusters, including Multi-Cloud Kubernetes Cluster. Dataflow is probably best suited for green-field projects. Less operational overhead, unified approach for batch and streaming pipelines. *So as ever your mileage varies*. If you want to migrate from your existing Hadoop/Spark cluster to GCP, or take advantage of your existing workforce, choose Dataproc or GKE. In many cases, a big consideration is that one already has a codebase written against a particular framework, and one just wants to deploy it on the GCP, so even if, say, the Beam programming mode/dataflow is superior to Hadoop, someone with a lot of Hadoop code might still choose Dataproc or GDE for the time being, rather than rewriting their code on Beam to run on Dataflow. HTH view my Linkedin profile <https://www.linkedin.com/in/mich-talebzadeh-ph-d-5205b2/> https://en.everybodywiki.com/Mich_Talebzadeh *Disclaimer:* Use it at your own risk. Any and all responsibility for any loss, damage or destruction of data or any other property which may arise from relying on this email's technical content is explicitly disclaimed. The author will in no case be liable for any monetary damages arising from such loss, damage or destruction. On Mon, 14 Feb 2022 at 05:46, Gourav Sengupta <gourav.sengu...@gmail.com> wrote: > Hi, > may be this is useful in case someone is testing SPARK in containers for > developing SPARK. > > *From a production scale work point of view:* > But if I am in AWS, I will just use GLUE if I want to use containers for > SPARK, without massively increasing my costs for operations unnecessarily. > > Also, in case I am not wrong, GCP already has SPARK running in serverless > mode. Personally I would never create the overhead of additional costs and > issues to my clients of deploying SPARK when those solutions are already > available by Cloud vendors. Infact, that is one of the precise reasons why > people use cloud - to reduce operational costs. > > Sorry, just trying to understand what is the scope of this work. > > > Regards, > Gourav Sengupta > > On Fri, Feb 11, 2022 at 8:35 PM Mich Talebzadeh <mich.talebza...@gmail.com> > wrote: > >> The equivalent of Google GKE autopilot >> <https://cloud.google.com/kubernetes-engine/docs/concepts/autopilot-overview> >> in >> AWS is AWS Fargate <https://aws.amazon.com/fargate/> >> >> >> I have not used the AWS Fargate so I can only mension Google's GKE >> Autopilot. >> >> >> This is developed from the concept of containerization and microservices. >> In the standard mode of creating a GKE cluster users can customize their >> configurations based on the requirements, GKE manages the control plane and >> users manually provision and manage their node infrastructure. So you >> choose your hardware type and memory/CPU where your spark containers will >> be running and they will be shown as VM hosts in your account. In GKE >> Autopilot mode, GKE manages the nodes, pre-configures the cluster with >> adds-on for auto-scaling, auto-upgrades, maintenance, Day 2 operations and >> security hardening. So there is a lot there. You don't choose your nodes >> and their sizes. You are effectively paying for the pods you use. >> >> >> Within spark-submit, you still need to specify the number of executors, >> driver and executor memory plus cores for each driver and executor when >> doing spark-submit. The theory is that the k8s cluster will deploy suitable >> nodes and will create enough pods on those nodes. With the standard k8s >> cluster you choose your nodes and you ensure that one core on each node is >> reserved for the OS itself. Otherwise if you allocate all cores to spark >> with --conf spark.executor.cores, you will receive this error >> >> >> kubctl describe pods -n spark >> >> ... >> >> Events: >> >> Type Reason Age From >> Message >> >> ---- ------ ---- ---- >> ------- >> >> Warning FailedScheduling 9s (x17 over 15m) default-scheduler >> 0/3 nodes are available: 3 Insufficient cpu. >> >> So with the standard k8s you have a choice of selecting your core sizes. >> With autopilot this node selection is left to autopilot to deploy suitable >> nodes and this will be a trial and error at the start (to get the >> configuration right). You may be lucky if the history of executions are >> kept current and the same job can be repeated. However, in my experience, >> to procedure the driver pod in "running state" is expensive timewise and >> without an executor in running state, there is no chance of spark job doing >> anything >> >> >> NAME READY STATUS RESTARTS >> AGE >> >> randomdatabigquery-cebab77eea6de971-exec-1 0/1 Pending 0 >> 31s >> >> randomdatabigquery-cebab77eea6de971-exec-2 0/1 Pending 0 >> 31s >> >> randomdatabigquery-cebab77eea6de971-exec-3 0/1 Pending 0 >> 31s >> >> randomdatabigquery-cebab77eea6de971-exec-4 0/1 Pending 0 >> 31s >> >> randomdatabigquery-cebab77eea6de971-exec-5 0/1 Pending 0 >> 31s >> >> randomdatabigquery-cebab77eea6de971-exec-6 0/1 Pending 0 >> 31s >> >> sparkbq-37405a7eea6b9468-driver 1/1 Running 0 >> 3m4s >> >> >> NAME READY STATUS >> RESTARTS AGE >> >> randomdatabigquery-cebab77eea6de971-exec-6 0/1 ContainerCreating >> 0 112s >> >> sparkbq-37405a7eea6b9468-driver 1/1 Running >> 0 4m25s >> >> NAME READY STATUS RESTARTS >> AGE >> >> randomdatabigquery-cebab77eea6de971-exec-6 1/1 Running 0 >> 114s >> >> sparkbq-37405a7eea6b9468-driver 1/1 Running 0 >> 4m27s >> >> Basically I told Spak to have 6 executors but could only bring into >> running state one executor after the driver pod spinning for 4 minutes. >> >> 22/02/11 20:16:18 INFO SparkKubernetesClientFactory: Auto-configuring K8S >> client using current context from users K8S config file >> >> 22/02/11 20:16:19 INFO Utils: Using initial executors = 6, max of >> spark.dynamicAllocation.initialExecutors, >> spark.dynamicAllocation.minExecutors and spark.executor.instances >> >> 22/02/11 20:16:19 INFO ExecutorPodsAllocator: Going to request 3 >> executors from Kubernetes for ResourceProfile Id: 0, target: 6 running: 0. >> >> 22/02/11 20:16:20 INFO BasicExecutorFeatureStep: Decommissioning not >> enabled, skipping shutdown script >> >> 22/02/11 20:16:20 INFO Utils: Successfully started service >> 'org.apache.spark.network.netty.NettyBlockTransferService' on port 7079. >> >> 22/02/11 20:16:20 INFO NettyBlockTransferService: Server created on >> sparkbq-37405a7eea6b9468-driver-svc.spark.svc:7079 >> >> 22/02/11 20:16:20 INFO BlockManager: Using >> org.apache.spark.storage.RandomBlockReplicationPolicy for block replication >> policy >> >> 22/02/11 20:16:20 INFO BlockManagerMaster: Registering BlockManager >> BlockManagerId(driver, sparkbq-37405a7eea6b9468-driver-svc.spark.svc, 7079, >> None) >> >> 22/02/11 20:16:20 INFO BlockManagerMasterEndpoint: Registering block >> manager sparkbq-37405a7eea6b9468-driver-svc.spark.svc:7079 with 366.3 MiB >> RAM, BlockManagerId(driver, >> sparkbq-37405a7eea6b9468-driver-svc.spark.svc, 7079, None) >> >> 22/02/11 20:16:20 INFO BlockManagerMaster: Registered BlockManager >> BlockManagerId(driver, sparkbq-37405a7eea6b9468-driver-svc.spark.svc, 7079, >> None) >> >> 22/02/11 20:16:20 INFO BlockManager: Initialized BlockManager: >> BlockManagerId(driver, sparkbq-37405a7eea6b9468-driver-svc.spark.svc, 7079, >> None) >> >> 22/02/11 20:16:20 INFO Utils: Using initial executors = 6, max of >> spark.dynamicAllocation.initialExecutors, >> spark.dynamicAllocation.minExecutors and spark.executor.instances >> >> 22/02/11 20:16:20 WARN ExecutorAllocationManager: Dynamic allocation >> without a shuffle service is an experimental feature. >> >> 22/02/11 20:16:20 INFO BasicExecutorFeatureStep: Decommissioning not >> enabled, skipping shutdown script >> >> 22/02/11 20:16:20 INFO BasicExecutorFeatureStep: Decommissioning not >> enabled, skipping shutdown script >> >> 22/02/11 20:16:20 INFO ExecutorPodsAllocator: Going to request 3 >> executors from Kubernetes for ResourceProfile Id: 0, target: 6 running: 3. >> >> 22/02/11 20:16:20 INFO BasicExecutorFeatureStep: Decommissioning not >> enabled, skipping shutdown script >> >> 22/02/11 20:16:20 INFO BasicExecutorFeatureStep: Decommissioning not >> enabled, skipping shutdown script >> >> 22/02/11 20:16:20 INFO BasicExecutorFeatureStep: Decommissioning not >> enabled, skipping shutdown script >> >> 22/02/11 20:16:49 INFO KubernetesClusterSchedulerBackend: >> SchedulerBackend is ready for scheduling beginning after waiting >> maxRegisteredResourcesWaitingTime: 30000000000(ns) >> >> 22/02/11 20:16:49 INFO SharedState: Setting hive.metastore.warehouse.dir >> ('null') to the value of spark.sql.warehouse.dir >> ('file:/opt/spark/work-dir/spark-warehouse'). >> >> 22/02/11 20:16:49 INFO SharedState: Warehouse path is >> 'file:/opt/spark/work-dir/spark-warehouse'. >> >> OK there is a lot to digest here and I appreciate feedback from other >> members that have experimented with GKE autopilot or AWS Fargate or are >> familiar with k8s. >> >> Thanks >> >> >> view my Linkedin profile >> <https://www.linkedin.com/in/mich-talebzadeh-ph-d-5205b2/> >> >> >> >> *Disclaimer:* Use it at your own risk. Any and all responsibility for >> any loss, damage or destruction of data or any other property which may >> arise from relying on this email's technical content is explicitly >> disclaimed. The author will in no case be liable for any monetary damages >> arising from such loss, damage or destruction. >> >> >> >