Databricks

databricks Best Practices

This guide outlines the essential best practices for developing on Databricks. Adhere to these rules to ensure your code is maintainable, performant, and secure.

1. Code Organization and Structure

Leverage Git folders and Databricks Asset Bundles for all projects. Treat notebooks as version-controlled code, not isolated scripts. Extract reusable logic into Python modules.

• Version Control: Always use Git folders for notebooks and source code. ❌ BAD: Storing notebooks directly in Workspace without Git integration. ✅ GOOD:

  # In a Git-synced notebook (e.g., /Repos/user/my-repo/notebooks/my_pipeline.py)
  # This notebook is version-controlled and can import local modules.
  from ..src.utils import process_data
  
  df = spark.read.table("raw_data")
  processed_df = process_data(df)
  processed_df.write.mode("overwrite").saveAsTable("processed_data")
  

• Module Extraction: For any logic beyond simple notebook orchestration, extract it into Python modules (.py files) within your Git repository. ❌ BAD:

  # In a notebook cell
  def complex_transformation(df):
      # 50+ lines of transformation logic
      return df
  

  ✅ GOOD:

  # /Repos/user/my-repo/src/transformations.py
  def complex_transformation(df):
      # Modular, testable logic
      return df
  
  # In a notebook
  from src.transformations import complex_transformation
  df = complex_transformation(spark.read.table("staging"))
  

• Project Structure with Bundles: Use Databricks Asset Bundles to define and deploy your entire project (jobs, pipelines, models, notebooks, infrastructure) as a single, versioned unit. ✅ GOOD:

  # databricks-bundle.yml
  bundle:
    name: my-data-pipeline
  resources:
    jobs:
      my_etl_job:
        name: My ETL Job
        tasks:
          - task_key: process_data
            notebook_task:
              notebook_path: ./notebooks/main_pipeline.py
            new_cluster:
              spark_version: "14.3.x-scala2.12"
              node_type_id: "Standard_DS3_v2"
              num_workers: 3
  

• Pin Dependencies: Always pin Python package versions in requirements.txt for reproducibility, especially for serverless compute. ❌ BAD:

  pandas
  numpy
  

  ✅ GOOD:

  pandas==2.2.3
  numpy==1.26.4
  

2. Common Patterns and Anti-patterns

Leverage Databricks-native services for all heavy lifting. Avoid using Databricks Apps compute for data processing.

• Offload Heavy Processing: Databricks Apps compute is for UI rendering. Use Databricks SQL for ad-hoc queries, Lakeflow Jobs for batch pipelines, and Model Serving for AI inference. ❌ BAD:

  # In a Databricks App (e.g., Flask app)
  df = spark.sql("SELECT * FROM large_table").toPandas() # Pulls large data to app compute
  processed_data = df.groupby('col').sum() # Heavy processing on app compute
  

  ✅ GOOD:

  # In a Databricks App
  # Use Databricks SQL for queries
  from databricks.sdk.service.sql import StatementExecutionAPI
  # ... authenticate ...
  statement_execution = StatementExecutionAPI(api_client)
  result = statement_execution.execute_statement(
      warehouse_id="your_sql_warehouse_id",
      statement="SELECT SUM(col) FROM large_table GROUP BY col",
      # ... handle async results ...
  )
  # For batch processing, trigger a Lakeflow Job
  # For AI inference, call a Model Serving endpoint
  

• Serverless Compute Compatibility: Ensure data is in Unity Catalog, use Databricks Runtime 14.3+, and avoid JARs. ❌ BAD:

  # Attempting to use a custom JAR for a data source on serverless
  spark.sparkContext.addJar("s3://my-bucket/custom-connector.jar")
  df = spark.read.format("com.example.CustomSource").load(...)
  

  ✅ GOOD:

  # Use native ingestion methods for serverless
  # For cloud storage:
  df = spark.readStream.format("cloudFiles") \
      .option("cloudFiles.format", "json") \
      .load("s3://my-bucket/raw_data")
  
  # For external databases (query federation):
  df = spark.read.table("my_catalog.external_schema.external_table")
  

3. Performance Considerations

Optimize resource usage and data access patterns.

• Databricks Apps Startup: Keep initialization lightweight. Load heavy resources only when needed. ❌ BAD:

  # app.py
  import large_model_library
  model = large_model_library.load_model("path/to/model") # Blocks startup
  

  ✅ GOOD:

  # app.py
  model = None
  def get_model():
      nonlocal model
      if model is None:
          import large_model_library
          model = large_model_library.load_model("path/to/model") # Lazy load
      return model
  

• In-Memory Caching (Apps): Cache frequently used data or API responses. ✅ GOOD:

  from functools import lru_cache
  
  @lru_cache(maxsize=128)
  def get_cached_data(key):
      # Expensive operation, e.g., query Databricks SQL
      return fetch_data_from_warehouse(key)
  

4. Common Pitfalls and Gotchas

Avoid common missteps that lead to instability or security vulnerabilities.

• Graceful Shutdown (Databricks Apps): Implement SIGTERM handling to shut down within 15 seconds. ❌ BAD: No SIGTERM handler, app gets SIGKILL. ✅ GOOD:

  import signal
  import sys
  import time
  
  def signal_handler(signum, frame):
      print("SIGTERM received, initiating graceful shutdown...")
      # Perform cleanup, close connections, etc.
      sys.exit(0)
  
  signal.signal(signal.SIGTERM, signal_handler)
  # ... your app logic ...
  

• Logging (Databricks Apps): Log to stdout and stderr only. ❌ BAD: with open("app.log", "a") as f: f.write(...) ✅ GOOD: print("Log message"); import logging; logging.info("Log message")

• Secrets Management: Never expose raw secrets. Use valueFrom in app config. ❌ BAD: env: MY_API_KEY: "super-secret-value" ✅ GOOD:

  env:
    MY_API_KEY:
      valueFrom:
        secret:
          name: "my-scope/my-api-key"
  

• Privileged Operations (Databricks Apps): Apps run as non-privileged users. ❌ BAD: Attempting apt-get install or root access. ✅ GOOD: Use Python/Node.js package managers (PyPI, npm).

5. Error Handling

Implement robust error handling and observability.

• Global Exception Handling (Databricks Apps): Prevent crashes, return proper HTTP errors. ❌ BAD: Uncaught exceptions exposing stack traces. ✅ GOOD:

  from flask import Flask, jsonify
  app = Flask(__name__)
  
  @app.errorhandler(Exception)
  def handle_exception(e):
      # Log the full exception for internal debugging
      app.logger.error(f"Unhandled exception: {e}", exc_info=True)
      return jsonify(error="An unexpected error occurred"), 500
  

• Observability: Enable monitoring for Lakeflow Jobs and Pipelines. ✅ GOOD: Configure job logging and alerts within Databricks.

6. Request/Response Patterns

Use Databricks SDKs/APIs and secure data exchange.

• Interoperability: Use the unified Databricks REST API or higher-level SDKs (Python, Java, Go, R) for cross-system integration. ❌ BAD: Custom HTTP clients for Databricks API calls. ✅ GOOD:

  from databricks.sdk import WorkspaceClient
  w = WorkspaceClient()
  jobs = w.jobs.list()
  for job in jobs:
      print(job.settings.name)
  

• Parameterized SQL: Prevent SQL injection. ❌ BAD: cursor.execute(f"SELECT * FROM users WHERE name = '{user_input}'") ✅ GOOD: cursor.execute("SELECT * FROM users WHERE name = %s", (user_input,))

• Asynchronous Requests (Apps for Long-Running Operations): Avoid synchronous waits. ✅ GOOD:

  # Initial request to start a job
  job_run = w.jobs.run_now(job_id="my_long_job")
  run_id = job_run.run_id
  
  # Periodically poll for status
  while w.jobs.get_run(run_id).state.life_cycle_state not in ["TERMINATED", "SKIPPED"]:
      time.sleep(10)
  

7. Testing Approaches

Embrace automated testing for reliability.

• Unit Testing: Unit test extracted Python modules locally or in CI/CD. ✅ GOOD:

  # test_transformations.py
  import pytest
  from src.transformations import complex_transformation
  from pyspark.sql import SparkSession
  
  @pytest.fixture(scope="session")
  def spark():
      return SparkSession.builder.appName("pytest-spark").getOrCreate()
  
  def test_complex_transformation(spark):
      data = [("A", 1), ("B", 2)]
      input_df = spark.createDataFrame(data, ["col1", "col2"])
      output_df = complex_transformation(input_df)
      assert output_df.count() == 2
      # ... more assertions ...
  

• CI/CD Pipeline: Implement a full CI/CD pipeline using Databricks Asset Bundles, GitHub Actions, Azure DevOps, etc. ✅ GOOD:

  # .github/workflows/databricks_ci.yml
  name: Databricks CI/CD
  on: [push]
  jobs:
    deploy:
      runs-on: ubuntu-latest
      steps:
        - uses: actions/checkout@v4
        - uses: databricks/setup-cli@v0
        - run: databricks bundle deploy -t production # Deploy using the bundle
          env:
            DATABRICKS_HOST: ${{ secrets.DATABRICKS_HOST }}
            DATABRICKS_TOKEN: ${{ secrets.DATABRICKS_TOKEN }}