Working with your cluster

Creating a cluster with workspaces

To create an on-demand Spark cluster attached to a Domino Workspace, click New Workspace from the Workspaces menu. On the Launch New Workspace dialog select the option to Attach Cluster. Specify the desired cluster settings and launch you workspace. Once the workspace is up, it will have access to the Spark cluster you configured.

The Hardware Tier for your workspace will determine the compute resources available to your Spark driver process.

new_workspace_with_spark.png

Creating a cluster with jobs

Similarly to workspaces, to create and on-demand Spark cluster attached to a Domino job, click on Run from the Jobs menu. One the Start a Job dialog select the option to Attach Cluster. Specify the desired cluster settings and launch your job. The job will have access to the Spark cluster you configured.

new_job_with_spark.png

As your command, you can use any Python script that contains a PySpark job.

You can also submit jobs using spark-submit but since it is not recognized automatically as one of the Domino supported job types you will need to wrap it with a shell script unless you included a copy as spark-submit.sh as part of preparing your compute environment.

Below is an example of a simple wrapper my-spark-submit.sh

#!/usr/bin/env bash

spark-submit $@

Understanding your cluster settings

Domino makes it simple to specify key settings when creating a Spark cluster.

cluster_settings_spark.png

  • Number of Executors

    Number of Executors that will be available to your Spark application.

    When you instantiate Spark context with the default settings, the spark.executor.instances Spark setting will be set to the number specified in the above dialog.

  • Quota Max

    The maximum number of executors that you can make available to your cluster is limited by the number of per-user executions that your Domino administrator has configured for your deployment.

    In addition to the number of Spark application executors, you will need 1 slot for your cluster master and 1 slot for your workspace or job.

  • Executor Hardware Tier

    The amount of compute resources (CPU and memory) that will be made available to Spark executors.

    When you instantiate Spark context with the default settings, the spark.executor.cores and spark.executor.memory Spark settings will be set to the values corresponding to the Hardware Tier.

    For cores, the number of cores will be rounded up to the nearest integer. For example, a HW Tier with 1.5 cores will result in spark.executor.cores set to 2 when creating the default context.

  • Master Hardware Tier

    Same mechanics as the Executor Hardware Tier, but applied to the resources that will be available for your Spark cluster master.

    The master in the Domino configuration has a pretty limited role in Spark application scheduling, and it does not need a significant amount of resources.

  • Cluster Compute Environment

    Designates your environment. Projects already using this environment
    will be allowed to continue to use it until a new project environment
    is set.
  • Dedicated local storage per executor

    The amount of dedicated storage in Gigibytes (2^30 bytes) that will be available to each executor.

    The storage will be automatically mounted to /tmp which is the default location for spark.local.dir that serves as the location for RDDs, when they need to be on disk, and for shuffle data.

    The storage will be automatically provisioned when the cluster is created and de-provisioned when it is shut down.

    Warning

    The local storage per executor should not be used for storing any data which needs to be available after the cluster is shut down.

Advanced cluster configuration topics

Spark executors vs workers

Domino gives you the ability to have flexible resource allocation through Hardware Tiers which makes it convenient to plan your cluster resources needed for your application in terms of executors as opposed to workers. When creating on-demand Spark clusters, Domino will automatically manage the relationship between Spark executors and workers (see Spark definitions). By default, there will be one worker for each executor.

Resource overhead

When setting up worker and executor memory limits, Domino ensures that the worker containers have additional memory on top of the Hardware Tier memory definition to account for the recommended overhead of 10% of executor memory with a minimum of 384MB. The usable executor memory will always be based on the value from the actual Hardware Tier.

You may need to take this into account when designing Hardware Tiers to achieve optimal packing of Spark worker containers on the physical instances of the Domino Kubernetes cluster.

Because of the worker memory overhead and the approach of using 1 worker per executor, it is possible that using lots of executors with smaller Hardware Tiers may result in excessive memory overhead.

If this is undesirable, an experienced data scientist can chose a large Hardware Tier when creating the cluster and then overwrite the executor core and memory settings to achieve the desired effect. See Adding and overwriting parameters for details on how to make such changes.

Connecting to your cluster

When provisioning your on-demand Spark cluster, Domino sets up key default cluster configuration parameters (e.g. spark.master, spark.driver.host) to the appropriate cluster URLs so that creating SparkSession or SparkContext with default context will properly connect the cluster to your workspace or job.

Creating SparkSession

You can create SparkSession using the following:

from pyspark.sql import SparkSession

spark = SparkSession \
        .builder \
        .appName("MyAppName") \
        .getOrCreate()

 # You can examine the full config
 spark.sparkContext.getConf().getAll()

Creating SparkContext

Alternatively, you can create SparkContext using the following:

from pyspark import SparkContext

conf = SparkConf().setAppName(appName).setMaster(master)
sc = SparkContext(conf=conf)

Adding and overwriting parameters

It is possible to add additional config parameters or overwrite any of the defaults by manipulating your context configuration.

from pyspark.sql import SparkSession

spark = SparkSession \
        .builder \
        .appName("MyAppName") \
        .config("spark.some.config.option", "some-value") \
        .getOrCreate()

You can also supply additional configuration settings that will apply for any clusters associated with a Domino project from Settings->Integrations->Domino managed on-demand cluster

project_cluster_settings_spark.png

Warning

You should not overwrite the default configuration values for spark.master, spark.driver.host since you may lose proper connectivity between your cluster and workspace or job.

Accessing the Spark UI

Spark provides a suite of web user interfaces that you can use to monitor the status and resource consumption of your Spark cluster.

Domino makes the Spark Web UI available for active on-demand clusters attached to both workspaces and jobs.

Spark UI from Workspaces

The Spark UI is available from a dedicated tab in your workspace.

access_ui_tabs_spark.png

Spark UI from Jobs

The Spark UI is also available for running jobs from Details tab.

access_ui_job_spark.png

Cluster lifecycle

On workspace or job startup, a Domino on-demand Spark cluster with the desired cluster settings is automatically provisioned and attached to the workspace or job as soon as the cluster becomes available.

On workspace or job termination, the on-demand Spark cluster and all associated resources are automatically terminated and de-provisioned. This includes any compute resources and storage allocated for the cluster.

Cluster network security

The on-demand Spark clusters created by Domino are not meant for sharing between multiple users. Each cluster is associated with a given workspace or a job instance. Access to the cluster and the Spark Web UI is restricted only to users who can access the workspace or the job attached to it. This restriction is enforced at the networking level and the cluster is only reachable from the execution that provisioned it.