databricks-spark-declarative-pipelines

name

databricks-spark-declarative-pipelines

description

Creates, configures, and updates Databricks Lakeflow Spark Declarative Pipelines (SDP/LDP) using serverless compute. Handles streaming tables, materialized views, CDC, SCD Type 2, and Auto Loader ingestion patterns. Use when building data pipelines, working with Delta Live Tables, ingesting streaming data, implementing change data capture, or when the user mentions SDP, LDP, DLT, Lakeflow pipelines, streaming tables, or bronze/silver/gold medallion architectures.

Lakeflow Spark Declarative Pipelines (SDP)

IMPORTANT: If this is a new pipeline (one does not already exist), see Quick Start. Be sure to use whatever language user has specified only (Python or SQL). Be sure to use Databricks Asset Bundles for new projects.

---

Critical Rules (always follow)

• MUST confirm language as Python or SQL. Stick with that language unless told otherwise.
• MUST if not modifying an existing pipeline, use Quick Start below.
• MUST create serverless pipelines by default. ** Only use classic clusters if user explicitly requires R language, Spark RDD APIs, or JAR libraries.

Required Steps

Copy this checklist and verify each item:

- [ ] Language selected: Python or SQL
- [ ] Compute type decided: serverless or classic compute
- [ ] Decide on multiple catalogs or schemas vs. all in one default schema
- [ ] Consider what should be parameterized at the pipeline level to make deployment easy.
- [ ] Consider [Multi-Schema Patterns](#multi-schema-patterns) below, ask if unclear on best choices.
- [ ] Consider [Modern Defaults](#modern-defaults) below, ask if unclear on best choices.


## Quick Start: Initialize New Pipeline Project

**RECOMMENDED**: Use `databricks pipelines init` to create production-ready Asset Bundle projects with multi-environment support.

### When to Use Bundle Initialization

Use bundle initialization for **New pipeline projects** for a professional structure from the start

Use manual workflow for:
- Quick prototyping without multi-environment needs
- Existing manual projects you want to continue
- Learning/experimentation

### Step 1: Initialize Project

I will automatically run this command when you request a new pipeline:

```bash
databricks pipelines init

Interactive Prompts:

• Project name: e.g., customer_orders_pipeline
• Initial catalog: Unity Catalog name (e.g., main, prod_catalog)
• Personal schema per user?: yes for dev (each user gets their own schema), no for prod
• Language: SQL or Python (auto-detected from your request - see language detection below)

Generated Structure:

my_pipeline/
├── databricks.yml              # Multi-environment config (dev/prod)
├── resources/
│   └── *_etl.pipeline.yml      # Pipeline resource definition
└── src/
    └── *_etl/
        ├── explorations/       # Exploratory code in .ipynb
        └── transformations/    # Your .sql or .py files here

Step 2: Customize Transformations

Replace the example code created by the init process with custom transformation files in src/transformations/ based on provided requirements, using best practice guidance from this skill.

For Python pipelines using cloudFiles: Ask the user where to store Auto Loader schema metadata. Recommend:

/Volumes/{catalog}/{schema}/{pipeline_name}_metadata/schemas

Step 3: Deploy and Run

# Deploy to workspace (dev by default)
databricks bundle deploy

# Run pipeline
databricks bundle run my_pipeline_etl

# Deploy to production
databricks bundle deploy --target prod

Quick Reference

Concept	Details
Names	SDP = Spark Declarative Pipelines = LDP = Lakeflow Declarative Pipelines = Lakeflow Pipelines (all interchangeable)
Python Import	from pyspark import pipelines as dp
Primary Decorators	@dp.table(), @dp.materialized_view(), @dp.temporary_view()
Temporary Views	@dp.temporary_view() creates in-pipeline temporary views (no catalog/schema, no cluster_by). Useful for intermediate logic before AUTO CDC or when a view needs multiple references without persistence.
Replaces	Delta Live Tables (DLT) with import dlt
Based On	Apache Spark 4.1+ (Databricks' modern data pipeline framework)
Docs	https://docs.databricks.com/aws/en/ldp/developer/python-dev

---

Detailed guides

Ingestion patterns: Use 1-ingestion-patterns.md when planning how to get new data into your Lakeflow pipeline —- covers file formats, batch/streaming options, and tips for incremental and full loads. (Keywords: Auto Loader, Kafka, Event Hub, Kinesis, file formats)

Streaming pipeline patterns: See 2-streaming-patterns.md for designing pipelines with streaming data sources, change data detection, triggers, and windowing. (Keywords: deduplication, windowing, stateful operations, joins)

SCD query patterns: See 3-scd-query-patterns.md for querying Slowly Changing Dimensions Type 2 history tables, including current state queries, point-in-time analysis, temporal joins, and change tracking. (Keywords: SCD Type 2 history tables, temporal joins, querying historical data)

Performance tuning: Use 4-performance-tuning.md for optimizing pipelines with Liquid Clustering, state management, and best practices for high-performance streaming workloads. (Keywords: Liquid Clustering, optimization, state management)

Python API reference: See 5-python-api.md for the modern pyspark.pipelines (dp) API reference and migration from legacy dlt API patterns. (Keywords: dp API, dlt API comparison)

DLT migration: Use 6-dlt-migration.md when migrating existing Delta Live Tables (DLT) pipelines to Spark Declarative Pipelines (SDP). (Keywords: migrating DLT pipelines to SDP)

Advanced configuration: See 7-advanced-configuration.md for advanced pipeline settings including development mode, continuous execution, notifications, Python dependencies, and custom cluster configurations. (Keywords: extra_settings parameter reference, examples)

Project initialization: Use 8-project-initialization.md for setting up new pipeline projects with databricks pipelines init, Asset Bundles, multi-environment deployments, and language detection logic. (Keywords: databricks pipelines init, Asset Bundles, language detection, migration guides)

AUTO CDC patterns: Use 9-auto_cdc.md for implementing Change Data Capture with AUTO CDC, including Slow Changing Dimensions (SCD Type 1 and Type 2) for tracking changes and deduplication. (Keywords: AUTO CDC, Slow Changing Dimension, SCD, SCD Type 1, SCD Type 2, change data capture, deduplication)

---

Workflow

1. Determine the task type:

   Setting up new project? → Read 8-project-initialization.md first Creating new pipeline? → Read 1-ingestion-patterns.md Creating stream table? → Read 2-streaming-patterns.md Querying SCD history tables? → Read 3-scd-query-patterns.md Implementing AUTO CDC or SCD? → Read 9-auto_cdc.md Performance issues? → Read 4-performance-tuning.md Using Python API? → Read 5-python-api.md Migrating from DLT? → Read 6-dlt-migration.md Advanced configuration? → Read 7-advanced-configuration.md Validating? → Read validation-checklist.md

2. Follow the instructions in the relevant guide

3. Repeat for next task type

---

Official Documentation

• Lakeflow Spark Declarative Pipelines Overview - Main documentation hub
• SQL Language Reference - SQL syntax for streaming tables and materialized views
• Python Language Reference - pyspark.pipelines API
• Loading Data - Auto Loader, Kafka, Kinesis ingestion
• Change Data Capture (CDC) - AUTO CDC, SCD Type 1/2

Medallion Architecture Pattern

Bronze Layer (Raw)

• Raw data ingested from sources in original format
• Minimal transformations (append-only, add metadata like _ingested_at, _source_file)
• Single source of truth preserving data lineage

Silver Layer (Validated)

• Cleaned and validated data.
• Might deduplicate here with auto_cdc, but often wait until the final step for auto_cdc if possible.
• Business logic applied (type casting, quality checks, filtering invalid records)
• Enterprise view of key business entities
• Enables self-service analytics and ML

Gold Layer (Business-Ready)

• Aggregated, denormalized, project-specific tables
• Optimized for consumption (reporting, dashboards, BI tools)
• Fewer joins, read-optimized data models
• Kimball star schema tables - dim_<entity_name>, fact_<entity_name>
• Deduplication often happens here via Slow Changing Dimensions (SCD), using auto_cdc. Sometimes that will happen upstream in silver instead, such as when joining multiple tables or business users plan to query the table from silver.

Typical Flow (Can vary)
Bronze: read_files() or spark.readStream.format("cloudFiles") → streaming table
Silver: read bronze → filter/clean/validate → streaming table Gold: read silver → aggregate/denormalize → auto_cdc or materialized view

Sources:

• https://www.databricks.com/glossary/medallion-architecture
• https://docs.databricks.com/aws/en/lakehouse/medallion
• https://www.databricks.com/blog/2022/06/24/data-warehousing-modeling-techniques-and-their-implementation-on-the-databricks-lakehouse-platform.html

For medallion architecture (bronze/silver/gold), two approaches work:

• Flat with naming (template default): bronze_*.sql, silver_*.sql, gold_*.sql
• Subdirectories: bronze/orders.sql, silver/cleaned.sql, gold/summary.sql

Both work with the transformations/** glob pattern. Choose based on preference.

See 8-project-initialization.md for complete details on bundle initialization, migration, and troubleshooting.

---

General SDP development guidance

Step 1: Write Pipeline Files Locally

Create .sql or .py files in a local folder:

my_pipeline/
├── bronze/
│   ├── ingest_orders.sql       # SQL (default for most cases)
│   └── ingest_events.py        # Python (for complex logic)
├── silver/
│   └── clean_orders.sql
└── gold/
    └── daily_summary.sql

SQL Example (bronze/ingest_orders.sql):

CREATE OR REFRESH STREAMING TABLE bronze_orders
CLUSTER BY (order_date)
AS
SELECT
  *,
  current_timestamp() AS _ingested_at,
  _metadata.file_path AS _source_file
FROM read_files(
  '/Volumes/catalog/schema/raw/orders/',
  format => 'json',
  schemaHints => 'order_id STRING, customer_id STRING, amount DECIMAL(10,2), order_date DATE'
);

Python Example (bronze/ingest_events.py):

from pyspark import pipelines as dp
from pyspark.sql.functions import col, current_timestamp

# Get schema location from pipeline configuration
schema_location_base = spark.conf.get("schema_location_base")

@dp.table(name="bronze_events", cluster_by=["event_date"])
def bronze_events():
    return (
        spark.readStream.format("cloudFiles")
        .option("cloudFiles.format", "json")
        .option("cloudFiles.schemaLocation", f"{schema_location_base}/bronze_events")
        .load("/Volumes/catalog/schema/raw/events/")
        .withColumn("_ingested_at", current_timestamp())
        .withColumn("_source_file", col("_metadata.file_path"))
    )

IMPORTANT for Python Pipelines: When using spark.readStream.format("cloudFiles") for cloud storage ingestion, with schema inference (no schema specified), you must specify a schema location.

Always ask the user where to store Auto Loader schema metadata. Recommend:

/Volumes/{catalog}/{schema}/{pipeline_name}_metadata/schemas

Example: /Volumes/my_catalog/pipeline_metadata/orders_pipeline_metadata/schemas

Never use the source data volume - this causes permission conflicts. The schema location should be configured in the pipeline settings and accessed via spark.conf.get("schema_location_base").

Language Selection:

CRITICAL RULE: If the user explicitly mentions "Python" in their request (e.g., "Python Spark Declarative Pipeline", "Python SDP", "use Python"), ALWAYS use Python without asking. The same applies to SQL - if they say "SQL pipeline", use SQL.

• Explicit language request: User says "Python" → Use Python. User says "SQL" → Use SQL. Do not ask for clarification.
• Auto-detection (only when no explicit language mentioned):
  • SQL indicators: "sql files", "simple transformations", "aggregations", "materialized view", "CREATE OR REFRESH"
  • Python indicators: ".py files", "UDF", "complex logic", "ML inference", "external API", "@dp.table", "pandas", "decorator"
• Prompt for clarification only when language intent is truly ambiguous (no explicit mention, mixed signals)
• Default to SQL only when ambiguous AND no Python indicators present

See 8-project-initialization.md for detailed language detection logic.

Option 1: Pipelines with DABs:

Use asset bundles and pipeline CLI. See Quick Start and 8-project-initialization.md for complete details.

Option 2: Manual Workflow (Advanced)

For rapid prototyping, experimentation, or when you prefer direct control without Asset Bundles, use the manual workflow with MCP tools.

Use MCP tools to create, run, and iterate on serverless SDP pipelines. The primary tool is create_or_update_pipeline which handles the entire lifecycle.

IMPORTANT: Always create serverless pipelines (default). Only use classic clusters if user explicitly ask for classic, pro, advances compute or requires R language, Spark RDD APIs, or JAR libraries.

See 10-mcp-approach.md for detailed guide.

Best Practices (2026)

Project Structure

• Default to databricks pipelines init for new projects (creates Asset Bundle)