Compute configuration reference

This article explains the configuration settings available in the Create Compute UI. Most users create compute resources using their assigned policies, which limits the configurable settings. If you don’t see a particular setting in your UI, it’s because the policy you’ve selected does not allow you to configure that setting.

The configurations and management tools described in this article apply to both all-purpose and job compute. For more considerations on configuring job compute, see Configure compute for jobs.

Create a new all-purpose compute resource

To create a new all-purpose compute resource:

1. In the workspace sidebar, click Compute.
2. Click the Create compute button.
3. Configure the compute resource.
4. Click Create compute.

You new compute resource will automatically start spinning up and be ready to use shortly.

Policies

Policies are a set of rules used to limit the configuration options available to users when they create compute resources. If a user doesn’t have the Unrestricted cluster creation entitlement, then they can only create compute resources using their granted policies.

To create compute resources according to a policy, select a policy from the Policy drop-down menu.

By default, all users have access to the Personal Compute policy, allowing them to create single-machine compute resources. If you need access to Personal Compute or any additional policies, reach out to your workspace admin.

Single-node or multi-node compute

Depending on the policy, you can select between creating a Single node compute resource or a Multi node compute resource.

Single node compute is intended for jobs that use small amounts of data or non-distributed workloads such as single-node machine learning libraries. Multi-node compute should be used for larger jobs with distributed workloads.

Single node properties

A single node compute resource has the following properties:

• Runs Spark locally.
• Driver acts as both master and worker, with no worker nodes.
• Spawns one executor thread per logical core in the compute resource, minus 1 core for the driver.
• Saves all stderr, stdout, and log4j log outputs in the driver log.
• Can’t be converted to a multi-node compute resource.

Selecting single or multi node

Consider your use case when deciding between single or multi-node compute:

• Large-scale data processing will exhaust the resources on a single node compute resource. For these workloads, Databricks recommends using multi-node compute.

• Single-node compute is not designed to be shared. To avoid resource conflicts, Databricks recommends using a multi-node compute resource when the compute must be shared.

• A multi-node compute resource can’t be scaled to 0 workers. Use single node compute instead.

• Single-node compute is not compatible with process isolation.

• GPU scheduling is not enabled on single node compute.

• On single-node compute, Spark cannot read Parquet files with a UDT column. The following error message results:

  The Spark driver has stopped unexpectedly and is restarting. Your notebook will be automatically reattached.
  

  To work around this problem, disable the native Parquet reader:

  spark.conf.set("spark.databricks.io.parquet.nativeReader.enabled", False)
  

Access modes

Access mode is a security feature that determines who can use the compute resource and the data they can access using the compute resource. Every compute resource in Azure Databricks has an access mode.

Databricks recommends that you use shared access mode for all workloads. Use the single user access mode only if your required functionality is not supported by shared access mode.

You can upgrade an existing compute resource to meet the requirements of Unity Catalog by setting its access mode to Single User or Shared. For detailed information about the functionality that is supported by each of these access modes in Unity Catalog-enabled workspaces, see Compute access mode limitations for Unity Catalog.

Databricks Runtime versions

Databricks Runtime is the set of core components that run on your compute. Select the runtime using the Databricks Runtime Version drop-down menu. For details on specific Databricks Runtime versions, see Databricks Runtime release notes versions and compatibility. All versions include Apache Spark. Databricks recommends the following:

• For all-purpose compute, use the most current version to ensure you have the latest optimizations and the most up-to-date compatibility between your code and preloaded packages.
• For job compute running operational workloads, consider using the Long Term Support (LTS) Databricks Runtime version. Using the LTS version will ensure you don’t run into compatibility issues and can thoroughly test your workload before upgrading.
• For data science and machine learning use cases, consider Databricks Runtime ML version.

Use Photon acceleration

Photon is enabled by default on compute running Databricks Runtime 9.1 LTS and above.

To enable or disable Photon acceleration, select the Use Photon Acceleration checkbox. To learn more about Photon, see What is Photon?.

Worker and driver node types

A compute resource consists of one driver node and zero or more worker nodes. You can pick separate cloud provider instance types for the driver and worker nodes, although by default the driver node uses the same instance type as the worker node. Different families of instance types fit different use cases, such as memory-intensive or compute-intensive workloads.

You can also select a pool to use as the worker or driver node. Only use a pool with spot instances as your worker type. Select a separate on-demand driver type to prevent your driver from being reclaimed. See Connect to pools.

Worker type

In multi-node compute, worker nodes run the Spark executors and other services required for a properly functioning compute resource. When you distribute your workload with Spark, all of the distributed processing happens on worker nodes. Azure Databricks runs one executor per worker node. Therefore, the terms executor and worker are used interchangeably in the context of the Databricks architecture.

Worker node IP addresses

Azure Databricks launches worker nodes with two private IP addresses each. The node’s primary private IP address hosts Azure Databricks internal traffic. The secondary private IP address is used by the Spark container for intra-cluster communication. This model allows Azure Databricks to provide isolation between multiple compute resources in the same workspace.

Driver type

The driver node maintains state information of all notebooks attached to the compute resource. The driver node also maintains the SparkContext, interprets all the commands you run from a notebook or a library on the compute resource, and runs the Apache Spark master that coordinates with the Spark executors.

The default value of the driver node type is the same as the worker node type. You can choose a larger driver node type with more memory if you are planning to collect() a lot of data from Spark workers and analyze them in the notebook.

GPU instance types

For computationally challenging tasks that demand high performance, like those associated with deep learning, Azure Databricks supports compute resources that are accelerated with graphics processing units (GPUs). For more information, see GPU-enabled compute.

Azure confidential computing VMs

Azure confidential computing VM types prevent unauthorized access to data while it’s in use, including from the cloud operator. This VM type is beneficial to highly regulated industries and regions, as well as businesses with sensitive data in the cloud. For more information on Azure’s confidential computing, see Azure confidential computing.

To run your workloads using Azure confidential computing VMs, select from the DC or EC series VM types in the worker and driver node dropdowns. See Azure Confidential VM options.

Spot instances

To save cost, you can choose to use spot instances, also known as Azure Spot VMs by checking the Spot instances checkbox.

The first instance will always be on-demand (the driver node is always on-demand) and subsequent instances will be spot instances.

If instances are evicted due to unavailability, Azure Databricks will attempt to acquire new spot instances to replace the evicted instances. If spot instances can’t be acquired, on-demand instances are deployed to replace the evicted instances. This on-demand failback is only supported for spot instances that have been fully acquired and are running. Spot instances that fail during setup are not automatically replaced.

Additionally, when new nodes are added to existing compute resources, Azure Databricks attempts to acquire spot instances for those nodes.

Enable autoscaling

When Enable autoscaling is checked, you can provide a minimum and maximum number of workers for the compute resource. Databricks then chooses the appropriate number of workers required to run your job.

To set the minimum and the maximum number of workers your compute resource will autoscale between, use the Min workers and Max workers fields next to the Worker type dropdown.

If you don’t enable autoscaling, you must enter a fixed number of workers in the Workers field next to the Worker type dropdown.

Benefits of autoscaling

With autoscaling, Azure Databricks dynamically reallocates workers to account for the characteristics of your job. Certain parts of your pipeline may be more computationally demanding than others, and Databricks automatically adds additional workers during these phases of your job (and removes them when they’re no longer needed).

Autoscaling makes it easier to achieve high utilization because you don’t need to provision the compute to match a workload. This applies especially to workloads whose requirements change over time (like exploring a dataset during the course of a day), but it can also apply to a one-time shorter workload whose provisioning requirements are unknown. Autoscaling thus offers two advantages:

• Workloads can run faster compared to a constant-sized under-provisioned compute resource.
• Autoscaling can reduce overall costs compared to a statically-sized compute resource.

Depending on the constant size of the compute resource and the workload, autoscaling gives you one or both of these benefits at the same time. The compute size can go below the minimum number of workers selected when the cloud provider terminates instances. In this case, Azure Databricks continuously retries to re-provision instances in order to maintain the minimum number of workers.

How autoscaling behaves

Workspace on the Premium plan use optimized autoscaling. Workspaces on the standard pricing plan use standard autoscaling.

Optimized autoscaling has the following characteristics:

• Scales up from min to max in 2 steps.
• Can scale down, even if the compute resource is not idle, by looking at the shuffle file state.
• Scales down based on a percentage of current nodes.
• On job compute, scales down if the compute resource is underutilized over the last 40 seconds.
• On all-purpose compute, scales down if the compute resource is underutilized over the last 150 seconds.
• The spark.databricks.aggressiveWindowDownS Spark configuration property specifies in seconds how often the compute makes down-scaling decisions. Increasing the value causes the compute to scale down more slowly. The maximum value is 600.

Standard autoscaling is used in standard plan workspaces. Standard autoscaling has the following characteristics:

• Starts by adding 8 nodes. Then scales up exponentially, taking as many steps as required to reach the max.
• Scales down when 90% of the nodes are not busy for 10 minutes and the compute has been idle for at least 30 seconds.
• Scales down exponentially, starting with 1 node.

Autoscaling with pools

If you are attaching your compute resource to a pool, consider the following:

• Make sure the compute size requested is less than or equal to the minimum number of idle instances in the pool. If it is larger, compute startup time will be equivalent to compute that doesn’t use a pool.

• Make sure the maximum compute size is less than or equal to the maximum capacity of the pool. If it is larger, the compute creation will fail.

Autoscaling example

If you reconfigure a static compute resource to autoscale, Azure Databricks immediately resizes the compute resource within the minimum and maximum bounds and then starts autoscaling. As an example, the following table demonstrates what happens to a compute resource with a certain initial size if you reconfigure the compute resource to autoscale between 5 and 10 nodes.

Enable autoscaling local storage

It can often be difficult to estimate how much disk space a particular job will take. To save you from having to estimate how many gigabytes of managed disk to attach to your compute at creation time, Azure Databricks automatically enables autoscaling local storage on all Azure Databricks compute.

With autoscaling local storage, Azure Databricks monitors the amount of free disk space available on your compute’s Spark workers. If a worker begins to run too low on disk, Databricks automatically attaches a new managed disk to the worker before it runs out of disk space. Disks are attached up to a limit of 5 TB of total disk space per virtual machine (including the virtual machine’s initial local storage).

The managed disks attached to a virtual machine are detached only when the virtual machine is returned to Azure. That is, managed disks are never detached from a virtual machine as long as they are part of a running compute. To scale down managed disk usage, Azure Databricks recommends using this feature in compute configured with autoscaling compute or automatic termination.

Local disk encryption

Some instance types you use to run compute may have locally attached disks. Azure Databricks may store shuffle data or ephemeral data on these locally attached disks. To ensure that all data at rest is encrypted for all storage types, including shuffle data that is stored temporarily on your compute resource’s local disks, you can enable local disk encryption.

When local disk encryption is enabled, Azure Databricks generates an encryption key locally that is unique to each compute node and is used to encrypt all data stored on local disks. The scope of the key is local to each compute node and is destroyed along with the compute node itself. During its lifetime, the key resides in memory for encryption and decryption and is stored encrypted on the disk.

To enable local disk encryption, you must use the Clusters API. During compute creation or edit, set enable_local_disk_encryption to true.

Automatic termination

You can set auto termination for compute. During compute creation, specify an inactivity period in minutes after which you want the compute resource to terminate.

If the difference between the current time and the last command run on the compute resource is more than the inactivity period specified, Azure Databricks automatically terminates that compute. resource For more information on compute termination, see Terminate a compute.

Tags

Tags allow you to easily monitor the cost of cloud resources used by various groups in your organization. Specify tags as key-value pairs when you create compute, and Azure Databricks applies these tags to cloud resources like VMs and disk volumes, as well as DBU usage reports.

For compute launched from pools, the custom tags are only applied to DBU usage reports and do not propagate to cloud resources.

For detailed information about how pool and compute tag types work together, see Monitor usage using tags

To add tags to your compute resource:

1. In the Tags section, add a key-value pair for each custom tag.
2. Click Add.

Spark configuration

To fine-tune Spark jobs, you can provide custom Spark configuration properties.

1. On the compute configuration page, click the Advanced Options toggle.

2. Click the Spark tab.

   

   In Spark config, enter the configuration properties as one key-value pair per line.

When you configure compute using the Clusters API, set Spark properties in the spark_conf field in the create cluster API or Update cluster API.

To enforce Spark configurations on compute, workspace admins can use compute policies.

Retrieve a Spark configuration property from a secret

Databricks recommends storing sensitive information, such as passwords, in a secret instead of plaintext. To reference a secret in the Spark configuration, use the following syntax:

spark.<property-name> {{secrets/<scope-name>/<secret-name>}}

For example, to set a Spark configuration property called password to the value of the secret stored in secrets/acme_app/password:

spark.password {{secrets/acme-app/password}}

For more information, see Manage secrets.

SSH access to compute

For security reasons, in Azure Databricks the SSH port is closed by default. If you want to enable SSH access to your Spark clusters, see SSH to the driver node.

Environment variables

Configure custom environment variables that you can access from init scripts running on the compute resource. Databricks also provides predefined environment variables that you can use in init scripts. You cannot override these predefined environment variables.

1. On the compute configuration page, click the Advanced Options toggle.

2. Click the Spark tab.

3. Set the environment variables in the Environment Variables field.

   

You can also set environment variables using the spark_env_vars field in the Create cluster API or Update cluster API.

Compute log delivery

When you create compute, you can specify a location to deliver the logs for the Spark driver node, worker nodes, and events. Logs are delivered every five minutes and archived hourly in your chosen destination. When a compute resource is terminated, Azure Databricks guarantees to deliver all logs generated up until the compute resource was terminated.

The destination of the logs depends on the compute resource’s cluster_id. If the specified destination is dbfs:/cluster-log-delivery, compute logs for 0630-191345-leap375 are delivered to dbfs:/cluster-log-delivery/0630-191345-leap375.

To configure the log delivery location:

1. On the compute page, click the Advanced Options toggle.
2. Click the Logging tab.
3. Select a destination type.
4. Enter the compute log path.