Running Spark with Yarn on the CDP base cluster
The primary supported way to run Spark workloads on Cloudera Machine Learning uses Spark on Kubernetes. This is different from Cloudera Data Science Workbench, with uses Spark on Yarn to run Spark workloads.
For users who are migrating projects from CDSW to CML, or who have existing Yarn workloads, CML Private Cloud offers a way to run those Spark on Yarn workloads on the CDP base cluster. This is sometimes called "Spark pushdown." This allows the Spark workloads to run without needing to modify them to run on Kubernetes.
The CML Admin must enable this mode for a CML workspace, and each CML workload must enable this mode to run Spark workloads in the attached CDP base cluster.
When this mode is enabled, each newly launched CML workload has port forwarding rules set up in Kubernetes. Additionally, Spark configurations are set in the CML session to allow Spark applications launched in the CML session to run in client mode with Executors in Yarn in the attached base cluster.
- In CML, Spark on Yarn Pushdown workloads are only supported with ML Runtimes.
- Spark pushdown functionality only works with DEX 1.19.1 Spark Addons.
- Yarn Service configured and running in your CDP Base Cluster
- Spark On Yarn service configured and running in your CDP Base Cluster
- The CDP Base Cluster must have access to the Spark drivers that run on Data Service Hosts running CML workloads, these are launched on a set of randomized ports in the range: 30000-32768
- Python must be installed on all CDP Base Cluster YARN Node Manager nodes which should match the Python version of the selected ML Runtime (i.e. 3.7 or 3.8)
- The python binary available on Yarn Node Manager nodes must be specified in the
PYSPARK_PYTHON environment variable
- As an example for 3.7, one could specify the environment variable like this for the CML
project with Spark Pushdown enabled:
- PYSPARK_PYTHON - The location of python in executors running in Yarn Nodes
- Note: In CML PYSPARK_PYTHON is by default set to /usr/local/bin/python3
- This should be changed to the appropriate location in Yarn Nodes
- PYSPARK_DRIVER_PYTHON = The location of python in the driver running in a CML session
Note: For CML runtimes PYSPARK_DRIVER_PYTHON is set to /usr/local/bin/python3
- As an example for 3.7, one could specify the environment variable like this for the CML project with Spark Pushdown enabled:
Enabling Spark on the base cluster
Spark can be enabled on the base cluster both site-wide and project-specific.
- Site Administration > Settings
Select Allow users to enable Spark Pushdown Configuration for Projects.
- A project-specific setting to enable spark pushdown for all newly launched workloads in the
project. Each project that intends to use the CDP Base Cluster Yarn for spark workloads must
enable this setting.
In Project Settings, select Settings > Enable Spark Pushdown.
Spark Application Dependencies
Due to the unique running mode of Spark on Yarn in CML, how dependencies are handled differ greatly from running the same jobs while on the base cluster.
To determine which dependencies are required on the cluster, you must understand that Spark code applications run in Spark executor processes distributed throughout the cluster. If the Python code you are running uses any third-party libraries, Spark executors require access to those libraries when they run on remote executors.
Refer to the following Spark configurations to determine how dependencies can be made available to executors.
- By default, this is unset in a CML Project Spark Pushdown project to ensure that all spark jars loaded from the CML Spark Runtime Addon is made available to yarn executors.
- This configuration should not be overridden within your CML projects. Consider using spark.yarn.dist.jars to indicate external references to jars.
- (Add note about added transfer time at beginning of workloads)
- This is not configured by CML.
- By default, this is set to /opt/spark/python/lib/*.zip to ensure that the pyspark and py4j zips included in CML Spark Runtime Addons are available to executors.
- (Can be overridden, keeping original)
- spark.yarn.dist.archives - This is not configured by CML.
- spark.yarn.dist.files - This is not configured by CML.
User-Specified Spark Application Configurations
Multiple Spark configuration sources are appended to a single file for Spark Pushdown in CML PVC. This occurs in the following order (lower has higher precedence as the contents of /etc/spark/conf/spark-defaults.conf are loaded from top-down):
- Base Cluster Spark spark-defaults.conf Defaults and Safety valves are included here
- CML system-specific configurations injection
- CML Project spark-defaults.conf
Check the contents of /etc/spark/conf/spark-defaults.conf inside the CML Session for the final configuration used by the spark driver.
CML-Injected Spark Application Configurations
There are a number of Spark Configurations which are applied by CML in order to enable or simplify Spark on Basecluster Yarn workloads.
Spark Environment Variables
Multiple environment variable sources are considered when setting up the CML session which will run the interactive spark driver.
- Base Cluster Spark spark-env.sh Defaults and Safety valves are included here
- CML system-specific spark envs overriding
For CML Session Environment
- Contents of constructed spark-env.sh (see above)
- Workspace env vars
- Project env vars
- User env vars