Databricks Asset Bundles (DAB)

Databricks Asset Bundles are the platform's first-party answer to "how do I treat my workspace like infrastructure?" A bundle is a directory containing a databricks.yml manifest and the source files it references — notebooks, Python wheels, DLT pipeline definitions, ML experiments, dashboards. The Databricks CLI reads the manifest, resolves variables and target overrides, and deploys the whole package to a workspace as a coherent unit. The same bundle definition produces dev, staging, and prod environments with only target-level differences.

Bundles replace the older "deploy notebooks with a Bash script and configure jobs in the UI" workflow with declarative IaC that is checked into source control, code-reviewed, and applied through CI/CD.

1. What a Bundle Is

A bundle is a directory tree like this:

my_bundle/
├── databricks.yml              # the manifest
├── README.md
├── src/
│   ├── ingest_orders.py
│   ├── transform_silver.py
│   └── dlt_pipeline.py
├── notebooks/
│   └── exploration.ipynb
├── resources/
│   ├── orders_job.yml          # split out for readability
│   └── orders_pipeline.yml
└── tests/
    └── test_transform_silver.py

Anything you can configure in the workspace UI — jobs, DLT pipelines, ML model registrations, model serving endpoints, AI/BI dashboards, SQL warehouses, secret scopes — can be declared in YAML and shipped via a bundle. Workspaces become deterministic and disposable.

2. The databricks.yml Structure

The manifest has four top-level blocks: bundle, workspace, resources, and targets. A minimal manifest looks like this:

bundle:
  name: orders_pipeline

workspace:
  host: https://adb-1234567890123456.7.azuredatabricks.net

variables:
  catalog:
    description: Unity Catalog target catalog
    default: dev_orders
  notifications_email:
    default: data-eng@example.com

resources:
  jobs:
    orders_etl:
      name: orders_etl_${bundle.target}
      tasks:
        - task_key: ingest
          notebook_task:
            notebook_path: ./src/ingest_orders.py
          job_cluster_key: shared_cluster
        - task_key: transform
          depends_on:
            - task_key: ingest
          notebook_task:
            notebook_path: ./src/transform_silver.py
          job_cluster_key: shared_cluster
      job_clusters:
        - job_cluster_key: shared_cluster
          new_cluster:
            spark_version: 15.4.x-scala2.12
            node_type_id: Standard_D4ds_v5
            num_workers: 2
      email_notifications:
        on_failure: [${var.notifications_email}]

targets:
  dev:
    mode: development
    default: true
    workspace:
      root_path: /Users/${workspace.current_user.userName}/.bundle/${bundle.name}/dev
    variables:
      catalog: dev_orders

  prod:
    mode: production
    workspace:
      root_path: /Shared/.bundle/${bundle.name}/prod
    run_as:
      service_principal_name: 0123abcd-...-....
    variables:
      catalog: prod_orders
      notifications_email: oncall@example.com

Key concepts:

• bundle — identity (name, optional git stamp). The CLI uses this to namespace deployed assets.
• workspace — default workspace host and the root path under which the CLI uploads source files.
• variables — typed inputs you can override per target or on the CLI with --var key=value. Substitute as ${var.name}.
• resources — the actual stuff being deployed: jobs, pipelines, models, model_serving_endpoints, experiments, schemas, volumes, dashboards, etc.
• targets — environments. Each target can override workspace, variables, run-as identity, and any resource field.

3. Targets: dev / staging / prod

The mode field on a target is the most consequential setting. It changes how the CLI rewrites resources before deploy:

• mode: development — prefixes resource names with [dev your.user], pauses schedules, sets run_as to the deploying user. Designed for safe iteration in a personal workspace path.
• mode: production — deploys with the configured names, schedules are active, requires a service principal as run_as, and refuses to deploy from a personal user folder. The CLI also warns if you deploy from a dirty git tree.

This makes it easy to databricks bundle deploy from a laptop into a personal sandbox while the same bundle goes through CI for production.

4. Resource Patterns

Multi-task Job

Already shown above. The pattern that matters: declare job_clusters once and reference them via job_cluster_key from each task so all tasks share warm compute.

DLT Pipeline

resources:
  pipelines:
    orders_dlt:
      name: orders_dlt_${bundle.target}
      catalog: ${var.catalog}
      target: silver
      libraries:
        - notebook:
            path: ./src/dlt_pipeline.py
      configuration:
        bronze.path: /Volumes/${var.catalog}/raw/orders
      clusters:
        - label: default
          autoscale:
            min_workers: 1
            max_workers: 4
      development: false
      photon: true
      serverless: true

Model Serving Endpoint

resources:
  model_serving_endpoints:
    fraud_scorer:
      name: fraud-scorer-${bundle.target}
      config:
        served_entities:
          - entity_name: ${var.catalog}.ml.fraud_model
            entity_version: "7"
            workload_size: Small
            scale_to_zero_enabled: true
        traffic_config:
          routes:
            - served_model_name: fraud_model-7
              traffic_percentage: 100
        auto_capture_config:
          catalog_name: ${var.catalog}
          schema_name: inference_logs
          table_name_prefix: fraud_scorer

MLflow Experiment

resources:
  experiments:
    fraud_training:
      name: /Shared/experiments/fraud_training_${bundle.target}
      tags:
        - key: owner
          value: ml-platform

5. Deploy and Run from the CLI

The CLI ships as databricks. Bundle commands are under databricks bundle.

# Validate the manifest, resolve variables, show what will be deployed
databricks bundle validate --target dev

# Upload sources and apply resources to the dev target
databricks bundle deploy --target dev

# Trigger a job by its resource key (not its display name)
databricks bundle run orders_etl --target dev

# Tail logs from the most recent run of a job
databricks bundle run orders_etl --target dev --no-wait
databricks jobs list-runs --job-id $(databricks bundle summary --target dev -o json | jq -r '.resources.jobs.orders_etl.id')

# Promote to production
databricks bundle deploy --target prod

# Tear down everything the bundle created in a target
databricks bundle destroy --target dev

6. CI/CD with GitHub Actions

The standard pattern is: validate on every PR, deploy to staging on merge to main, deploy to prod on a tagged release. Authenticate with an OAuth machine-to-machine (M2M) service principal token stored as a GitHub secret.

name: deploy-bundle

on:
  pull_request:
    branches: [main]
  push:
    branches: [main]
    tags: ['v*']

jobs:
  validate:
    if: github.event_name == 'pull_request'
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: databricks/setup-cli@main
      - run: databricks bundle validate --target staging
        env:
          DATABRICKS_HOST: ${{ secrets.DATABRICKS_HOST }}
          DATABRICKS_CLIENT_ID: ${{ secrets.DATABRICKS_CLIENT_ID }}
          DATABRICKS_CLIENT_SECRET: ${{ secrets.DATABRICKS_CLIENT_SECRET }}

  deploy-staging:
    if: github.event_name == 'push' && github.ref == 'refs/heads/main'
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: databricks/setup-cli@main
      - run: databricks bundle deploy --target staging
        env:
          DATABRICKS_HOST: ${{ secrets.DATABRICKS_HOST }}
          DATABRICKS_CLIENT_ID: ${{ secrets.DATABRICKS_CLIENT_ID }}
          DATABRICKS_CLIENT_SECRET: ${{ secrets.DATABRICKS_CLIENT_SECRET }}

  deploy-prod:
    if: startsWith(github.ref, 'refs/tags/v')
    runs-on: ubuntu-latest
    environment: production    # require manual approval
    steps:
      - uses: actions/checkout@v4
      - uses: databricks/setup-cli@main
      - run: databricks bundle deploy --target prod
        env:
          DATABRICKS_HOST: ${{ secrets.DATABRICKS_PROD_HOST }}
          DATABRICKS_CLIENT_ID: ${{ secrets.DATABRICKS_PROD_CLIENT_ID }}
          DATABRICKS_CLIENT_SECRET: ${{ secrets.DATABRICKS_PROD_CLIENT_SECRET }}

The environment: production line ties the prod job to a GitHub Environment, which can require reviewer approval before the step runs — a cheap and effective production guardrail.

7. Bundle Templates

For new projects, bootstrap with databricks bundle init, which scaffolds a bundle from a template. Built-in templates exist for default Python jobs, DLT pipelines, MLOps stacks, and dbt projects, and you can author custom templates as Go templates over a YAML config.

# Use a built-in template
databricks bundle init default-python

# Use a custom template from a git repo
databricks bundle init https://github.com/example/my-dab-template

8. Bundles vs the Terraform Databricks Provider

The Terraform Databricks provider has been the IaC tool for Databricks for years. It is broader in scope — it manages workspace creation, networking, accounts, and other control-plane resources. Asset Bundles cover only what lives inside a workspace, but they cover it more idiomatically.

The pragmatic split: Terraform owns the platform (workspace creation, metastore, account-level groups). Bundles own the application (this team's jobs, pipelines, ML models, dashboards). The two compose well — Terraform stands up the workspace, then each application team ships into it with their own bundle and CI/CD.

Common Interview Questions: