Spark, being a fast and distributed data processing engine, has taken over the traditional MapReduce in many use cases. With the continuous support from the community, the improvements in every release is taking Spark to new heights. Progress DataRPM leverages the computation power of Spark for our Cognitive Predictive Maintenance (CPdM) Platform. In this blog, I would like to share my experience of working with Spark for our CPdM Platform, which truly made our lives easier.
Every Spark application has dedicated independent set of executors. The same set of executors are used for any computation or caching of RDDs or DataFrames. But in our CPdM Platform, we have multiple applications running in parallel. For eg : DataLake data load, Machine Learning algorithms, number of Data Discovery and transformation applications etc. Concurrently using multiple applications, cluster resources (memory and cores) can get exhausted as each application keeps the executors running as long as the application runs. And that’s a big problem to tackle with.
Executor resources are always measured in terms of Cores and Memory. If we are sure about the type of applications running in our deployment, it is always possible to control the executor resources by configuring –spark.executor.memory, -spark.executor.cores. But this works only for a lab kind of deployment wherein only 1 or 2 users will be playing around Spark with small data. But in case of BigData, it is always necessary to handle such cases so that we don’t end up in DOS (Denial Of Service) for any application. Most of the distributed computation platforms run on YARN, where in you will be supporting many types of applications along with Spark ones. So putting all these things together, it is quite evident that managing the resources is equally important than coding a effective transformation function as it might lead to under/over utilization of resources and applications might starve for resources for a long time bringing the situation to STOP THE WORLD!!
Spark in version 1.2 introduces Dynamic Resource Allocation,where in there is an ability to scale the executors up and down based on the need. Until 1.5.2 version of Spark, this feature was only available with YARN. With Spark 1.6, this is made available for all types of deployments : Standalone, Mesos, YARN.
Let us see how simple it is to enable the same for YARN.
In addition to this, to make the external shuffle service(explained later) enabled for YARN,
Based on the –spark.executor.memory, –spark.executor.cores, –spark.tasks.cpus and total number of tasks, the number of executors required will be calculated. Once the number of executors are calculated, the same can be requested from the Application Master and in turn communicate with the YARN Resource Manager to request for desired number of containers to be allocated so that executors are run. This process happens in batches when there are pending tasks for spark.dynamicAllocation.schedulerBacklogTimeout seconds, and then triggered again every spark.dynamicAllocation.sustainedSchedulerBacklogTimeout seconds. Once the execution is done, the containers are released and the executors who are idle for spark.dynamicAllocation.executorIdleTimeout are shutdown explicitly by the Application Master, so that other running applications can request for more executors.
External Shuffle Service: A long running process that runs on each node independent of the application, serves the shuffle files to other executors, which means any executor’s shuffle state will be served even after its removal.
Knowing the rich set of configurations, with the combination of spark.dynamicAllocation.initialExecutors, spark.dynamicAllocation.maxExecutors, spark.dynamicAllocation.minExecutors we can divide the cluster resources among the applications by defining specific Spark Config properties for each application so that resources for the applications are not overlapped and continue to serve in parallel.
Dynamic Resource Allocation removes most of the headache, a user has to undergo to solve memory related issues in a distributed computation environment, so that he can focus mainly on the application development. Realizing the importance of such a hook provided by Spark, has truly made the lives of engineers easy.
Happy Sparking :)
Vinay Kashyap is an Associate Architect with Progress DataRPM with over 9 years of platform engineering experience. He has hands-on expertise in Big Data Analytics and related technologies. He is a strong believer of “Design for the future and implement for now” philosophy and is a champion contributor to the architecture, design and implementation of core features in the Machine Learning and Big Data platforms.
Subscribe to get all the news, info and tutorials you need to build better business apps and sites
Copyright © 2019 Progress Software Corporation and/or its subsidiaries or affiliates.All Rights Reserved.
Progress, Telerik, Ipswitch, and certain product names used herein are trademarks or registered trademarks of Progress Software Corporation and/or one of its subsidiaries or affiliates in the U.S. and/or other countries. See Trademarks for appropriate markings.