Databricks SDK for Python (Beta)

Beta: This SDK is supported for production use cases, but we do expect future releases to have some interface changes; see Interface stability. We are keen to hear feedback from you on these SDKs. Please file issues, and we will address them. | See also the SDK for Java | See also the SDK for Go | See also the Terraform Provider | See also cloud-specific docs (AWS, Azure, GCP) | See also the API reference on readthedocs

The Databricks SDK for Python includes functionality to accelerate development with Python for the Databricks Lakehouse. It covers all public Databricks REST API operations. The SDK's internal HTTP client is robust and handles failures on different levels by performing intelligent retries.

Contents

• Getting started
• Code examples
• Authentication
• Long-running operations
• Paginated responses
• Single-sign-on with OAuth
• Error handling
• Logging
• Integration with dbutils
• Interface stability

Getting started

1. Please install Databricks SDK for Python via pip install databricks-sdk and instantiate WorkspaceClient:

from databricks.sdk import WorkspaceClient
w = WorkspaceClient()
for c in w.clusters.list():
    print(c.cluster_name)

Databricks SDK for Python is compatible with Python 3.7 (until June 2023), 3.8, 3.9, 3.10, and 3.11.
Note: Databricks Runtime starting from version 13.1 includes a bundled version of the Python SDK.
It is highly recommended to upgrade to the latest version which you can do by running the following in a notebook cell:

%pip install --upgrade databricks-sdk

followed by

dbutils.library.restartPython()

Code examples

The Databricks SDK for Python comes with a number of examples demonstrating how to use the library for various common use-cases, including

• Using the SDK with OAuth from a webserver
• Using long-running operations
• Authenticating a client app using OAuth

These examples and more are located in the examples/ directory of the Github repository.

Some other examples of using the SDK include:

• Unity Catalog Automated Migration heavily relies on Python SDK for working with Databricks APIs.
• ip-access-list-analyzer checks & prunes invalid entries from IP Access Lists.

Authentication

If you use Databricks configuration profiles or Databricks-specific environment variables for Databricks authentication, the only code required to start working with a Databricks workspace is the following code snippet, which instructs the Databricks SDK for Python to use its default authentication flow:

from databricks.sdk import WorkspaceClient
w = WorkspaceClient()
w. # press <TAB> for autocompletion

The conventional name for the variable that holds the workspace-level client of the Databricks SDK for Python is w, which is shorthand for workspace.

In this section

• Default authentication flow
• Databricks native authentication
• Azure native authentication
• Overriding .databrickscfg
• Additional authentication configuration options

Default authentication flow

If you run the Databricks Terraform Provider, the Databricks SDK for Go, the Databricks CLI, or applications that target the Databricks SDKs for other languages, most likely they will all interoperate nicely together. By default, the Databricks SDK for Python tries the following authentication methods, in the following order, until it succeeds:

1. Databricks native authentication
2. Azure native authentication
3. If the SDK is unsuccessful at this point, it returns an authentication error and stops running.

You can instruct the Databricks SDK for Python to use a specific authentication method by setting the auth_type argument as described in the following sections.

For each authentication method, the SDK searches for compatible authentication credentials in the following locations, in the following order. Once the SDK finds a compatible set of credentials that it can use, it stops searching:

1. Credentials that are hard-coded into configuration arguments.

   :warning: Caution: Databricks does not recommend hard-coding credentials into arguments, as they can be exposed in plain text in version control systems. Use environment variables or configuration profiles instead.

2. Credentials in Databricks-specific environment variables.

3. For Databricks native authentication, credentials in the .databrickscfg file's DEFAULT configuration profile from its default file location (~ for Linux or macOS, and %USERPROFILE% for Windows).

4. For Azure native authentication, the SDK searches for credentials through the Azure CLI as needed.

Depending on the Databricks authentication method, the SDK uses the following information. Presented are the WorkspaceClient and AccountClient arguments (which have corresponding .databrickscfg file fields), their descriptions, and any corresponding environment variables.

Databricks native authentication

By default, the Databricks SDK for Python initially tries Databricks token authentication (auth_type='pat' argument). If the SDK is unsuccessful, it then tries Databricks basic (username/password) authentication (auth_type="basic" argument).

• For Databricks token authentication, you must provide host and token; or their environment variable or .databrickscfg file field equivalents.
• For Databricks basic authentication, you must provide host, username, and password (for AWS workspace-level operations); or host, account_id, username, and password (for AWS, Azure, or GCP account-level operations); or their environment variable or .databrickscfg file field equivalents.

For example, to use Databricks token authentication:

from databricks.sdk import WorkspaceClient
w = WorkspaceClient(host=input('Databricks Workspace URL: '), token=input('Token: '))

Azure native authentication

By default, the Databricks SDK for Python first tries Azure client secret authentication (auth_type='azure-client-secret' argument). If the SDK is unsuccessful, it then tries Azure CLI authentication (auth_type='azure-cli' argument). See Manage service principals.

The Databricks SDK for Python picks up an Azure CLI token, if you've previously authenticated as an Azure user by running az login on your machine. See Get Azure AD tokens for users by using the Azure CLI.

To authenticate as an Azure Active Directory (Azure AD) service principal, you must provide one of the following. See also Add a service principal to your Azure Databricks account:

• azure_workspace_resource_id, azure_client_secret, azure_client_id, and azure_tenant_id; or their environment variable or .databrickscfg file field equivalents.
• azure_workspace_resource_id and azure_use_msi; or their environment variable or .databrickscfg file field equivalents.

For example, to use Azure client secret authentication:

from databricks.sdk import WorkspaceClient
w = WorkspaceClient(host=input('Databricks Workspace URL: '),
                    azure_workspace_resource_id=input('Azure Resource ID: '),
                    azure_tenant_id=input('AAD Tenant ID: '),
                    azure_client_id=input('AAD Client ID: '),
                    azure_client_secret=input('AAD Client Secret: '))

Please see more examples in this document.

Google Cloud Platform native authentication

By default, the Databricks SDK for Python first tries GCP credentials authentication (auth_type='google-credentials', argument). If the SDK is unsuccessful, it then tries Google Cloud Platform (GCP) ID authentication (auth_type='google-id', argument).

The Databricks SDK for Python picks up an OAuth token in the scope of the Google Default Application Credentials (DAC) flow. This means that if you have run gcloud auth application-default login on your development machine, or launch the application on the compute, that is allowed to impersonate the Google Cloud service account specified in google_service_account. Authentication should then work out of the box. See Creating and managing service accounts.

To authenticate as a Google Cloud service account, you must provide one of the following:

• host and google_credentials; or their environment variable or .databrickscfg file field equivalents.
• host and google_service_account; or their environment variable or .databrickscfg file field equivalents.

For example, to use Google ID authentication:

from databricks.sdk import WorkspaceClient
w = WorkspaceClient(host=input('Databricks Workspace URL: '),
                    google_service_account=input('Google Service Account: '))

Overriding .databrickscfg

For Databricks native authentication, you can override the default behavior for using .databrickscfg as follows:

For example, to use a profile named MYPROFILE instead of DEFAULT:

from databricks.sdk import WorkspaceClient
w = WorkspaceClient(profile='MYPROFILE')
# Now call the Databricks workspace APIs as desired...

Additional authentication configuration options

For all authentication methods, you can override the default behavior in client arguments as follows:

For example, to turn on debug HTTP headers:

from databricks.sdk import WorkspaceClient
w = WorkspaceClient(debug_headers=True)
# Now call the Databricks workspace APIs as desired...

Long-running operations

When you invoke a long-running operation, the SDK provides a high-level API to trigger these operations and wait for the related entities to reach the correct state or return the error message in case of failure. All long-running operations return generic Wait instance with result() method to get a result of long-running operation, once it's finished. Databricks SDK for Python picks the most reasonable default timeouts for every method, but sometimes you may find yourself in a situation, where you'd want to provide datetime.timedelta() as the value of timeout argument to result() method.

There are a number of long-running operations in Databricks APIs such as managing:

• Clusters,
• Command execution
• Jobs
• Libraries
• Delta Live Tables pipelines
• Databricks SQL warehouses.

For example, in the Clusters API, once you create a cluster, you receive a cluster ID, and the cluster is in the PENDING state Meanwhile Databricks takes care of provisioning virtual machines from the cloud provider in the background. The cluster is only usable in the RUNNING state and so you have to wait for that state to be reached.

Another example is the API for running a job or repairing the run: right after the run starts, the run is in the PENDING state. The job is only considered to be finished when it is in either the TERMINATED or SKIPPED state. Also you would likely need the error message if the long-running operation times out and fails with an error code. Other times you may want to configure a custom timeout other than the default of 20 minutes.

In the following example, w.clusters.create returns ClusterInfo only once the cluster is in the RUNNING state, otherwise it will timeout in 10 minutes:

import datetime
import logging
from databricks.sdk import WorkspaceClient

w = WorkspaceClient()
info = w.clusters.create_and_wait(cluster_name='Created cluster',
                                  spark_version='12.0.x-scala2.12',
                                  node_type_id='m5d.large',
                                  autotermination_minutes=10,
                                  num_workers=1,
                                  timeout=datetime.timedelta(minutes=10))
logging.info(f'Created: {info}')

Please look at the examples/starting_job_and_waiting.py for a more advanced usage:

import datetime
import logging
import time

from databricks.sdk import WorkspaceClient
import databricks.sdk.service.jobs as j

w = WorkspaceClient()

# create a dummy file on DBFS that just sleeps for 10 seconds
py_on_dbfs = f'/home/{w.current_user.me().user_name}/sample.py'
with w.dbfs.open(py_on_dbfs, write=True, overwrite=True) as f:
    f.write(b'import time; time.sleep(10); print("Hello, World!")')

# trigger one-time-run job and get waiter object
waiter = w.jobs.submit(run_name=f'py-sdk-run-{time.time()}', tasks=[
    j.RunSubmitTaskSettings(
        task_key='hello_world',
        new_cluster=j.BaseClusterInfo(
            spark_version=w.clusters.select_spark_version(long_term_support=True),
            node_type_id=w.clusters.select_node_type(local_disk=True),
            num_workers=1
        ),
        spark_python_task=j.SparkPythonTask(
            python_file=f'dbfs:{py_on_dbfs}'
        ),
    )
])

logging.info(f'starting to poll: {waiter.run_id}')

# callback, that receives a polled entity between state updates
def print_status(run: j.Run):
    statuses = [f'{t.task_key}: {t.state.life_cycle_state}' for t in run.tasks]
    logging.info(f'workflow intermediate status: {", ".join(statuses)}')

# If you want to perform polling in a separate thread, process, or service,
# you can use w.jobs.wait_get_run_job_terminated_or_skipped(
#   run_id=waiter.run_id,
#   timeout=datetime.timedelta(minutes=15),
#   callback=print_status) to achieve the same results.
#
# Waiter interface allows for `w.jobs.submit(..).result()` simplicity in
# the scenarios, where you need to block the calling thread for the job to finish.
run = waiter.result(timeout=datetime.timedelta(minutes=15),
                    callback=print_status)

logging.info(f'job finished: {run.run_page_url}')

Paginated responses

On the platform side the Databricks APIs have different wait to deal with pagination:

• Some APIs follow the offset-plus-limit pagination
• Some start their offsets from 0 and some from 1
• Some use the cursor-based iteration
• Others just return all results in a single response

The Databricks SDK for Python hides this complexity under Iterator[T] abstraction, where multi-page results yield items. Python typing helps to auto-complete the individual item fields.

import logging
from databricks.sdk import WorkspaceClient
w = WorkspaceClient()
for repo in w.repos.list():
    logging.info(f'Found repo: {repo.path}')

Please look at the examples/last_job_runs.py for a more advanced usage:

import logging
from collections import defaultdict
from datetime import datetime, timezone
from databricks.sdk import WorkspaceClient

latest_state = {}
all_jobs = {}
durations = defaultdict(list)

w = WorkspaceClient()
for job in w.jobs.list():
    all_jobs[job.job_id] = job
    for run in w.jobs.list_runs(job_id=job.job_id, expand_tasks=False):
        durations[job.job_id].append(run.run_duration)
        if job.job_id not in latest_state:
            latest_state[job.job_id] = run
            continue
        if run.end_time < latest_state[job.job_id].end_time:
            continue
        latest_state[job.job_id] = run

summary = []
for job_id, run in latest_state.items():
    summary.append({
        'job_name': all_jobs[job_id].settings.name,
        'last_status': run.state.result_state,
        'last_finished': datetime.fromtimestamp(run.end_time/1000, timezone.utc),
        'average_duration': sum(durations[job_id]) / len(durations[job_id])
    })

for line in sorted(summary, key=lambda s: s['last_finished'], reverse=True):
    logging.info(f'Latest: {line}')

Single-Sign-On (SSO) with OAuth

Authorization Code flow with PKCE

For a regular web app running on a server, it's recommended to use the Authorization Code Flow to obtain an Access Token and a Refresh Token. This method is considered safe because the Access Token is transmitted directly to the server hosting the app, without passing through the user's web browser and risking exposure.

To enhance the security of the Authorization Code Flow, the PKCE (Proof Key for Code Exchange) mechanism can be employed. With PKCE, the calling application generates a secret called the Code Verifier, which is verified by the authorization server. The app also creates a transform value of the Code Verifier, called the Code Challenge, and sends it over HTTPS to obtain an Authorization Code. By intercepting the Authorization Code, a malicious attacker cannot exchange it for a token without possessing the Code Verifier.

The presented sample is a Python3 script that uses the Flask web framework along with Databricks SDK for Python to demonstrate how to implement the OAuth Authorization Code flow with PKCE security. It can be used to build an app where each user uses their identity to access Databricks resources. The script can be executed with or without client and secret credentials for a custom OAuth app.

Databricks SDK for Python exposes the oauth_client.initiate_consent() helper to acquire user redirect URL and initiate PKCE state verification. Application developers are expected to persist RefreshableCredentials in the webapp session and restore it via RefreshableCredentials.from_dict(oauth_client, session['creds']) helpers.

Works for both AWS and Azure. Not supported for GCP at the moment.

from databricks.sdk.oauth import OAuthClient

oauth_client = OAuthClient(host='<workspace-url>',
                           client_id='<oauth client ID>',
                           redirect_url=f'http://host.domain/callback',
                           scopes=['clusters'])

import secrets
from flask import Flask, render_template_string, request, redirect, url_for, session

APP_NAME = 'flask-demo'
app = Flask(APP_NAME)
app.secret_key = secrets.token_urlsafe(32)


@app.route('/callback')
def callback():
   from databricks.sdk.oauth import Consent
   consent = Consent.from_dict(oauth_client, session['consent'])
   session['creds'] = consent.exchange_callback_parameters(request.args).as_dict()
   return redirect(url_for('index'))


@app.route('/')
def index():
   if 'creds' not in session:
      consent = oauth_client.initiate_consent()
      session['consent'] = consent.as_dict()
      return redirect(consent.auth_url)

   from databricks.sdk import WorkspaceClient
   from databricks.sdk.oauth import SessionCredentials

   credentials_provider = SessionCredentials.from_dict(oauth_client, session['creds'])
   workspace_client = WorkspaceClient(host=oauth_client.host,
                                      product=APP_NAME,
                                      credentials_provider=credentials_provider)

   return render_template_string('...', w=workspace_client)

SSO for local scripts on development machines

For applications, that do run on developer workstations, Databricks SDK for Python provides auth_type='external-browser' utility, that opens up a browser for a user to go through SSO flow. Azure support is still in the early experimental stage.

from databricks.sdk import WorkspaceClient

host = input('Enter Databricks host: ')

w = WorkspaceClient(host=host, auth_type='external-browser')
clusters = w.clusters.list()

for cl in clusters:
    print(f' - {cl.cluster_name} is {cl.state}')

Creating custom OAuth applications

In order to use OAuth with Databricks SDK for Python, you should use account_client.custom_app_integration.create API.

import logging, getpass
from databricks.sdk import AccountClient
account_client = AccountClient(host='https://accounts.cloud.databricks.com',
                               account_id=input('Databricks Account ID: '),
                               username=input('Username: '),
                               password=getpass.getpass('Password: '))

logging.info('Enrolling all published apps...')
account_client.o_auth_enrollment.create(enable_all_published_apps=True)

status = account_client.o_auth_enrollment.get()
logging.info(f'Enrolled all published apps: {status}')

custom_app = account_client.custom_app_integration.create(
    name='awesome-app',
    redirect_urls=[f'https://host.domain/path/to/callback'],
    confidential=True)
logging.info(f'Created new custom app: '
             f'--client_id {custom_app.client_id} '
             f'--client_secret {custom_app.client_secret}')

Error handling

The Databricks SDK for Python provides a robust error-handling mechanism that allows developers to catch and handle API errors. When an error occurs, the SDK will raise an exception that contains information about the error, such as the HTTP status code, error message, and error details. Developers can catch these exceptions and handle them appropriately in their code.

from databricks.sdk import WorkspaceClient
from databricks.sdk.errors import ResourceDoesNotExist

w = WorkspaceClient()
try:
    w.clusters.get(cluster_id='1234-5678-9012')
except ResourceDoesNotExist as e:
    print(f'Cluster not found: {e}')

The SDK handles inconsistencies in error responses amongst the different services, providing a consistent interface for developers to work with. Simply catch the appropriate exception type and handle the error as needed. The errors returned by the Databricks API are defined in databricks/sdk/errors/platform.py.

Logging

The Databricks SDK for Python seamlessly integrates with the standard Logging facility for Python. This allows developers to easily enable and customize logging for their Databricks Python projects. To enable debug logging in your Databricks Python project, you can follow the example below:

import logging, sys
logging.basicConfig(stream=sys.stderr,
                    level=logging.INFO,
                    format='%(asctime)s [%(name)s][%(levelname)s] %(message)s')
logging.getLogger('databricks.sdk').setLevel(logging.DEBUG)

from databricks.sdk import WorkspaceClient
w = WorkspaceClient(debug_truncate_bytes=1024, debug_headers=False)
for cluster in w.clusters.list():
    logging.info(f'Found cluster: {cluster.cluster_name}')

In the above code snippet, the logging module is imported and the basicConfig() method is used to set the logging level to DEBUG. This will enable logging at the debug level and above. Developers can adjust the logging level as needed to control the verbosity of the logging output. The SDK will log all requests and responses to standard error output, using the format > for requests and < for responses. In some cases, requests or responses may be truncated due to size considerations. If this occurs, the log message will include the text ... (XXX additional elements) to indicate that the request or response has been truncated. To increase the truncation limits, developers can set the debug_truncate_bytes configuration property or the DATABRICKS_DEBUG_TRUNCATE_BYTES environment variable. To protect sensitive data, such as authentication tokens, passwords, or any HTTP headers, the SDK will automatically replace these values with **REDACTED** in the log output. Developers can disable this redaction by setting the debug_headers configuration property to True.

2023-03-22 21:19:21,702 [databricks.sdk][DEBUG] GET /api/2.0/clusters/list
< 200 OK
< {
<   "clusters": [
<     {
<       "autotermination_minutes": 60,
<       "cluster_id": "1109-115255-s1w13zjj",
<       "cluster_name": "DEFAULT Test Cluster",
<       ... truncated for brevity
<     },
<     "... (47 additional elements)"
<   ]
< }

Overall, the logging capabilities provided by the Databricks SDK for Python can be a powerful tool for monitoring and troubleshooting your Databricks Python projects. Developers can use the various logging methods and configuration options provided by the SDK to customize the logging output to their specific needs.

Interaction with dbutils

You can use the client-side implementation of dbutils by accessing dbutils property on the WorkspaceClient. Most of the dbutils.fs operations and dbutils.secrets are implemented natively in Python within Databricks SDK. Non-SDK implementations still require a Databricks cluster, that you have to specify through the cluster_id configuration attribute or DATABRICKS_CLUSTER_ID environment variable. Don't worry if cluster is not running: internally, Databricks SDK for Python calls w.clusters.ensure_cluster_is_running().

from databricks.sdk import WorkspaceClient

w = WorkspaceClient()
dbutils = w.dbutils

files_in_root = dbutils.fs.ls('/')
print(f'number of files in root: {len(files_in_root)}')

Alternatively, you can import dbutils from databricks.sdk.runtime module, but you have to make sure that all configuration is already present in the environment variables:

from databricks.sdk.runtime import dbutils

for secret_scope in dbutils.secrets.listScopes():
    for secret_metadata in dbutils.secrets.list(secret_scope.name):
        print(f'found {secret_metadata.key} secret in {secret_scope.name} scope')

Interface stability

Databricks is actively working on stabilizing the Databricks SDK for Python's interfaces. API clients for all services are generated from specification files that are synchronized from the main platform. You are highly encouraged to pin the exact dependency version and read the changelog where Databricks documents the changes. Databricks may have minor documented backward-incompatible changes, such as renaming some type names to bring more consistency.