Lakeflow Spark Declarative Pipelines (SDP)

Critical Rules (always follow)

Syntax: CREATE OR REFRESH (not CREATE OR REPLACE)

• MUST use CREATE OR REFRESH for SDP objects:
  • CREATE OR REFRESH STREAMING TABLE - for streaming tables
  • CREATE OR REFRESH MATERIALIZED VIEW - for materialized views
• NEVER use CREATE OR REPLACE - that is standard SQL syntax, not SDP syntax

Simplicity First

• MUST create the minimal number of tables to solve the task
• Simplicity first: prefer single pipeline even for multi-schema setups - use fully qualified names (catalog.schema.table)
• When asked to "create a silver table" or "create a gold table", create ONE table - not a multi-layer pipeline
• Don't add intermediate tables, staging tables, or helper views unless explicitly requested
• A silver transformation = 1 streaming table reading from bronze
• A gold aggregation = 1 materialized view reading from silver
• Create bronze→silver→gold chains when the user asks for a "pipeline" or "medallion architecture" or full/detailed ingestion. Otherwise keep it simple - don't over engineer.

Language Selection

• MUST know the language (Python or SQL). For simple task / pipeline / table creation, pick SQL. For complex pipeline with parametrized information, or if the user mentions python-related items pick python. If you have a doubt, ask the user. Stick with that language unless told otherwise.

Other Rules

• MUST create serverless pipelines by default. Only use classic clusters if user explicitly requires R language, Spark RDD APIs, or JAR libraries.
• MUST choose the right workflow based on context (see below).
• When the user provides table schema and asks for code, respond directly with the code. Don't ask clarifying questions if the request is clear.

Tools

• List files in volume: databricks fs ls dbfs:/Volumes/{catalog}/{schema}/{volume}/{path} --profile {PROFILE}
• Query data: databricks experimental aitools tools query --profile {PROFILE} --warehouse abc123 "SELECT 1 FROM catalog.schema.table"
• Discover schema: databricks experimental aitools tools discover-schema --profile {PROFILE} catalog.schema.table1 catalog.schema.table2
• Pipelines CLI: databricks pipelines init|deploy|run|logs|stop or use databricks pipelines --help for more options

Choose Your Workflow

First, determine which workflow to use:

Option A: Standalone New Pipeline Project (use databricks pipelines init)

Use this when the user wants to create a new, standalone SDP project that will have its own DAB:

• User asks: "Create a new pipeline", "Build me an SDP", "Set up a new data pipeline"
• No existing databricks.yml in the workspace
• The pipeline IS the project (not part of a larger demo/app)

Use databricks pipeline CLI commands:

databricks pipelines init --output-dir . --config-file init-config.json

Example init-config.json:

{
  "project_name": "customer_pipeline",
  "initial_catalog": "prod_catalog",
  "use_personal_schema": "no",
  "initial_language": "sql"
}

→ See 1-project-initialization.md →

Option B: Pipeline within Existing Bundle (edit the bundle)

Use this when the pipeline is part of an existing DAB project:

• There's already a databricks.yml file in the project
• User is adding a pipeline to an existing app/demo

→ See 1-project-initialization.md for adding pipelines to existing bundles

Option C: Rapid Iteration with MCP Tools (no bundle management)

Use this when you need to quickly create, test, and iterate on a pipeline without managing bundle files:

• User wants to "just run a pipeline and see if it works"
• Part of a larger demo where bundle is managed separately, or the DAB bundle will be created at the end as you want to quickly test the project first
• Prototyping or experimenting with pipeline logic
• User explicitly asks to use MCP tools

→ See 2-mcp-approach.md for MCP-based workflow

Required Checklist

Before writing pipeline code, make sure you have:

- [ ] Language selected: Python or SQL
- [ ] Read the syntax basics: **SQL**: Always Read [sql/1-syntax-basics.md](references/sql/1-syntax-basics.md), **Python**: Always Read [python/1-syntax-basics.md](references/python/1-syntax-basics.md)
- [ ] Workflow chosen: Standalone DAB / Existing DAB / MCP iteration
- [ ] Compute type: serverless (default) or classic
- [ ] Schema strategy: single schema with prefixes vs. multi-schema
- [ ] Consider [Multi-Schema Patterns](#multi-schema-patterns) and [Modern Defaults](#modern-defaults)

Then read additional guides based on what the pipeline needs, when you need it:

Quick Reference

Legacy APIs (Do NOT Use)

Streaming Table vs Materialized View

Use streaming tables for windowed aggregations to enable incremental processing. Use materialized views for simple aggregations that recompute fully on each refresh.

Task-Based Routing

After choosing your workflow (see Choose Your Workflow), determine the specific task:

Choose documentation by language:

SQL Documentation

Python Documentation

General Documentation

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

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/existing.

See 1-project-initialization.md for complete details on bundle initialization, migration, and troubleshooting.

General SDP development guidance

SQL Example:

CREATE OR REFRESH STREAMING TABLE bronze_orders
CLUSTER BY (order_date)
AS SELECT *, current_timestamp() AS _ingested_at
FROM STREAM read_files('/Volumes/catalog/schema/raw/orders/', format => 'json');

Python Example:

from pyspark import pipelines as dp

@dp.table(name="bronze_events", cluster_by=["event_date"])
def bronze_events():
    return spark.readStream.format("cloudFiles").option("cloudFiles.format", "json").load("/Volumes/...")

For detailed syntax, see sql/1-syntax-basics.md or python/1-syntax-basics.md.

Best Practices (2026)

Project Structure

• Standalone pipeline projects: Use databricks pipelines init for Asset Bundle with multi-environment support
• Pipeline in existing bundle: Add to resources/*.pipeline.yml
• Rapid iteration/prototyping: Use MCP tools, formalize in bundle later
• See 1-project-initialization.md for project setup details

Minimal pipeline config pointers

• Define parameters in your pipeline’s configuration and access them in code with spark.conf.get("key").
• In Databricks Asset Bundles, set these under resources.pipelines.<pipeline>.configuration; validate with databricks bundle validate.

Modern Defaults

• Always use raw .sql/.py files for the transformations files - NO notebooks in your pipeline. Pipeline code must be plain files.
• Databricks notebook source for explorations - Use # Databricks notebook source format with # COMMAND ---------- separators for ad-hoc queries. See examples/exploration_notebook.py.
• Serverless compute - Do not use classic clusters unless explicitly required (R, RDD APIs, JAR libraries)
• Unity Catalog (required for serverless)
• CLUSTER BY (Liquid Clustering), not PARTITION BY with ZORDER - see sql/5-performance.md or python/5-performance.md
• read_files() for SQL cloud storage ingestion - always consume a folder, not a single file - see sql/2-ingestion.md

Multi-Schema Patterns

Preferred: One pipeline writing to multiple schemas using fully qualified table names (catalog.schema.table). This keeps dependencies clear and is simpler to manage than multiple pipelines.

• Python: @dp.table(name="catalog.bronze_schema.orders")
• SQL: CREATE OR REFRESH STREAMING TABLE catalog.silver_schema.orders_clean AS ...

For detailed examples, see 3-advanced-configuration.md.

Fallback: If all tables must be in the same schema, use name prefixes (bronze_*, silver_*, gold_*).

Post-Run Validation (Required)

After running a pipeline (via DAB or MCP), you MUST validate both the execution status AND the actual data.

Step 1: Check Pipeline Execution Status

From MCP (run_pipeline or create_or_update_pipeline):

• Check result["success"] and result["state"]
• If failed, check result["message"] and result["errors"] for details

From DAB (databricks bundle run):

• Check the command output for success/failure
• Use get_pipeline(pipeline_id=...) to get detailed status and recent events

Step 2: Validate Output Data

Even if the pipeline reports SUCCESS, you MUST verify the data is correct:

# MCP Tool: get_table_stats_and_schema - validates schema, row counts, and stats
get_table_stats_and_schema(
    catalog="my_catalog",
    schema="my_schema",
    table_names=["bronze_*", "silver_*", "gold_*"]  # Use glob patterns
)

Check for:

• Empty tables (row_count = 0) - indicates ingestion or filtering issues
• Unexpected row counts - joins may have exploded or filtered too much
• Missing columns - schema mismatch or transformation errors
• NULL values in key columns - data quality issues

Step 3: Debug Data Issues

If validation reveals problems, trace upstream to find the root cause:

1. Start from the problematic table - identify what's wrong (empty, wrong counts, bad data)

2. Check its source table - use get_table_stats_and_schema on the upstream table

3. Trace back to bronze - continue until you find where the issue originates

4. Common causes:

   • Bronze empty → source files missing or path incorrect
   • Silver empty → filter too aggressive or join condition wrong
   • Gold wrong counts → aggregation logic error or duplicate keys
   • Data mismatch → type casting issues or NULL handling

5. Fix the SQL/Python code, re-upload, and re-run the pipeline

Do NOT use execute_sql with COUNT queries for validation - get_table_stats_and_schema is faster and returns more information in a single call.

Common Issues

For detailed errors, the result["message"] from create_or_update_pipeline includes suggested next steps. Use get_pipeline(pipeline_id=...) which includes recent events and error details.

Advanced Pipeline Configuration

For advanced configuration options (development mode, continuous pipelines, custom clusters, notifications, Python dependencies, etc.), see 3-advanced-configuration.md.

Platform Constraints

Serverless Pipeline Requirements (Default)

Serverless Limitations (When Classic Clusters Required)

General Constraints

Default to serverless unless user explicitly requires R, RDD APIs, or JAR libraries.

Related Skills

• databricks-jobs - for orchestrating and scheduling pipeline runs
• databricks-bundles - for multi-environment deployment of pipeline projects
• databricks-synthetic-data-gen - for generating test data to feed into pipelines
• databricks-unity-catalog - for catalog/schema/volume management and governance