Lessons Learned From Running Spark On Docker
23 likes15,545 views
Running Spark on Docker containers provides flexibility for data scientists and control for IT. Some key lessons learned include optimizing CPU and memory resources to avoid noisy neighbor problems, managing Docker images efficiently, using network plugins for multi-host connectivity, and addressing storage and security considerations. Performance testing showed Spark on Docker containers can achieve comparable performance to bare metal deployments for large-scale data processing workloads.
Ad
More Related Content
What's hot (20)
Similar to Lessons Learned From Running Spark On Docker (20)
Ad
More from Spark Summit (20)
Ad
Recently uploaded (20)
Lessons Learned From Running Spark On Docker
- 1. Lessons Learned From Running Spark On Docker Thomas Phelan Chief Architect, BlueData @tapbluedata
- 2. Outline • Spark Deployment Models • Spark on Docker • Lessons Learned • Performance • Demo • Key Takeaways
- 3. Data scientists want flexibility: • Different versions of Spark • Different sets of tools IT wants control: • Multi-tenancy - Data security - Network isolation Deploying Multiple Spark Clusters
- 4. Spark Deployment Models On-premises: • Hadoop distribution running Spark on YARN • Spark standalone mode • Spark using the Mesos container manager/resource scheduler • Spark (standalone or on YARN) deployed as a collection of services running within Docker containers Spark-as-a-Service in the cloud: • Databricks • AWS EMR, Google Dataproc, Microsoft Azure, IBM, others
- 5. Advantages of Docker Containers Property Description Hardware-Agnostic Using operating system primitives (e.g. LXC), containers can run consistently on any server or VM without modification Content-Agnostic Can encapsulate any payload and its dependencies Content Isolation Resource, network, and content isolation. Avoids dependency hell Automation Standard operations to run, start, stop, commit, etc. Perfect for DevOps Highly Efficient Lightweight, virtually no performance or start-up penalty
- 6. Running Spark on Docker • Docker containers provide a powerful option for greater agility and flexibility in application deployment on-premises • Running a complex, multi-service platform such as Spark in containers in a distributed enterprise-grade environment can be daunting • Here is how we did it ... while maintaining performance comparable to bare-metal
- 7. Spark on Docker: Design • Deploy Spark clusters as Docker containers spanning multiple physical hosts • Master container runs all Spark services (master, worker, jupyter, zeppelin) • Worker containers run Spark worker • Automate Spark service configuration inside containers to facilitate cluster cloning • Container storage is always ephemeral. Persistent storage is external
- 8. Spark Dockerfile # Spark-1.5.2 docker imagefor RHEL/CentOS 6.x FROM centos:centos6 # Download andextractspark RUN mkdir /usr/lib/spark;curl -s http://d3kbcqa49mib13.cloudfront.net/spark-1.5.2-bin-hadoop2.4.tgz | tar -xz -C /usr/lib/spark/ # Download andextractscala RUN mkdir /usr/lib/scala;curl -s http://www.scala-lang.org/files/archive/scala-2.10.3.tgz | tar xz -C /usr/lib/scala/ # Install zeppelin RUN mkdir /usr/lib/zeppelin;curl -s http://10.10.10.10:8080/build/thirdparty/zeppelin/zeppelin-0.6.0-incubating-SNAPSHOT- v2.tar.gz|tar xz -C /usr/lib/zeppelin RUN yum clean all && rm -rf /tmp/* /var/tmp/* /var/cache/yum/* ADD configure_spark_services.sh/root/configure_spark_services.sh RUN chmod -x /root/configure_spark_services.sh && /root/configure_spark_services.sh
- 9. Spark on Docker: Lessons Resource Utilization: • CPU cores vs. CPU shares • Over-provisioning of CPU recommended – noisy-neighbor problem • No over-provisioning of memory – swap Spark Image Management: • Utilize Docker’s open-source image repository • Author new Docker images using Dockerfiles • Tip: Docker images can get large. Use “docker squash” to save on size
- 10. Spark on Docker: Lessons Network: • Connect containers across hosts – Various network plugins available with Docker v1.10 • Persistence of IP address across container restart • DHCP/DNS service required for IP allocation and hostname resolution • Deploy VLANs and VxLAN tunnels for tenant level traffic isolation Storage: • Default size of a container’s /root needs to be tweaked • Resizing of storage inside an existing container is tricky • Mount /root and /data as block devices • Tip: Mounting block devices into a container does not support symbolic links (IOW: /dev/sdb will not work, /dm/… PCI device can change across host reboot)
- 11. Spark on Docker Architecture Container Orchestrator OVS OVS DHCP/DNS SparkMaster SparkWorrker SparkWorker SparkWorker SparkMaster SparkWorker Zeppelin VxLAN tunnel NIC NIC Tenant Networks
- 12. Docker Security Considerations • Security is essential since containers and host share their kernel – Non-privileged containers • Achieved through layered set of capabilities • Different capabilities provide different levels of isolation and protection • Add “capabilities” to a container based on what operations are permitted
- 13. Spark on Docker Performance • Spark 1.x on YARN • HiBench - Terasort – Data sizes: 100Gb, 500GB, 1TB • 10 node physical/virtual cluster • 36 cores and112GB memory per node • 2TB HDFS storage per node (SSDs) • 800GB ephemeral storage
- 14. Spark Performance Results 0 50 100 150 200 250 300 350 400 100GB 500GB 1TB Bare-Metal Containers (BlueData) MB/s
- 15. Demo
- 16. Key Takeaways of Spark in Docker • Value for a single cluster deployment – Significant benefits and savings for enterprise deployment • Get best of both worlds: – On-premises: security, governance, no data copying/moving – Spark-as-a-Service: multi-tenancy, elasticity, self-service • Performance is good