Lakehouse reference architectures (download)

This article covers architectural guidance for the lakehouse in terms of data source, ingestion, transformation, querying and processing, serving, analysis, and storage.

Each reference architecture has a downloadable PDF in 11 x 17 (A3) format.

While the lakehouse on Databricks is an open platform that integrates with a large ecosystem of partner tools, the reference architectures focus only on AWS services and the Databricks lakehouse. The cloud provider services shown are selected to illustrate the concepts and are not exhaustive.

Download: Reference architecture for the Databricks lakehouse on AWS

The AWS reference architecture shows the following AWS-specific services for ingesting, storage, serving, and analysis:

Amazon Redshift as a source for Lakehouse Federation
Amazon AppFlow and AWS Glue for batch ingest
AWS IoT Core, Amazon Kinesis, and AWS DMS for streaming ingest
Amazon S3 as the object storage
Amazon RDS and Amazon DynamoDB as operational databases
Amazon QuickSight as BI tool
Amazon Bedrock as a unified API to foundation models from leading AI startups and Amazon

Organization of the reference architectures

The reference architecture is structured along the swim lanes Source, Ingest, Transform, Query/Process, Serve, Analysis, and Storage:

Source

The architecture distinguishes between semi-structured and unstructured data (sensors and IoT, media, files/logs), and structured data (RDBMS, business applications). SQL sources (RDBMS) can also be integrated into the lakehouse and Unity Catalog without ETL through lakehouse federation. In addition, data might be loaded from other cloud providers.
Ingest

Data can be ingested into the lakehouse via batch or streaming:
- Databricks LakeFlow Connect offers built-in connectors for ingestion from enterprise applications and databases. The resulting ingestion pipeline is governed by Unity Catalog and is powered by serverless compute and Delta Live Tables.
- Files delivered to cloud storage can be loaded directly using the Databricks Auto Loader.
- For batch ingestion of data from enterprise applications into Delta Lake, the Databricks lakehouse relies on partner ingest tools with specific adapters for these systems of record.
- Streaming events can be ingested directly from event streaming systems such as Kafka using Databricks Structured Streaming. Streaming sources can be sensors, IoT, or change data capture processes.
Storage

Data is typically stored in the cloud storage system where the ETL pipelines use the medallion architecture to store data in a curated way as Delta files/tables.
Transform and Query / process

The Databricks lakehouse uses its engines Apache Spark and Photon for all transformations and queries.

DLT (Delta Live Tables) is a declarative framework for simplifying and optimizing reliable, maintainable, and testable data processing pipelines.

Powered by Apache Spark and Photon, the Databricks Data Intelligence Platform supports both types of workloads: SQL queries via SQL warehouses, and SQL, Python and Scala workloads via workspace clusters.

For data science (ML Modeling and Gen AI), the Databricks AI and Machine Learning platform provides specialized ML runtimes for AutoML and for coding ML jobs. All data science and MLOps workflows are best supported by MLflow.
Serving

For DWH and BI use cases, the Databricks lakehouse provides Databricks SQL, the data warehouse powered by SQL warehouses, and serverless SQL warehouses.

For machine learning, model serving is a scalable, real-time, enterprise-grade model serving capability hosted in the Databricks control plane. Mosaic AI Gateway is Databricks solution for governing and monitoring access to supported generative AI models and their associated model serving endpoints.

Operational databases: External systems, such as operational databases, can be used to store and deliver final data products to user applications.

Collaboration: Business partners get secure access to the data they need through Delta Sharing. Based on Delta Sharing, the Databricks Marketplace is an open forum for exchanging data products.
Analysis

The final business applications are in this swim lane. Examples include custom clients such as AI applications connected to Mosaic AI Model Serving for real-time inference or applications that access data pushed from the lakehouse to an operational database.

For BI use cases, analysts typically use BI tools to access the data warehouse. SQL developers can additionally use the Databricks SQL Editor (not shown in the diagram) for queries and dashboarding.

The Data Intelligence Platform also offers dashboards to build data visualizations and share insights.
Integrate
- The Databricks platform integrates with standard identity providers for user management and single sign on (SSO).
- External AI services like OpenAI, LangChain or HuggingFace can be used directly from within the Databricks Intelligence Platform.
- External orchestrators can either use the comprehensive REST API or dedicated connectors to external orchestration tools like Apache Airflow.
- Unity Catalog is used for all data & AI governance in the Databricks Intelligence Platform and can integrate other databases into its governance through Lakehouse Federation.
  
  Additionally, Unity Catalog can be integrated into other enterprise catalogs. Contact the enterprise catalog vendor for details.

Common capabilities for all workloads

In addition, the Databricks lakehouse comes with management capabilities that support all workloads:

Data and AI governance

The central data and AI governance system in the Databricks Data Intelligence Platform is Unity Catalog. Unity Catalog provides a single place to manage data access policies that apply across all workspaces and supports all assets created or used in the lakehouse, such as tables, volumes, features (feature store), and models (model registry). Unity Catalog can also be used to capture runtime data lineage across queries run on Databricks.

Databricks lakehouse monitoring allows you to monitor the quality of the data in all of the tables in your account. It can also track the performance of machine learning models and model-serving endpoints.

For observability, system tables is a Databricks-hosted analytical store of your account’s operational data. System tables can be used for historical observability across your account.
Data intelligence engine

The Databricks Data Intelligence Platform allows your entire organization to use data and AI. It is powered by DatabricksIQ and combines generative AI with the unification benefits of a lakehouse to understand the unique semantics of your data.

The Databricks Assistant is available in Databricks notebooks, SQL editor, and file editor as a context-aware AI assistant for developers.
Automation & Orchestration

Databricks Jobs orchestrate data processing, machine learning, and analytics pipelines on the Databricks Data Intelligence Platform. Delta Live Tables allow you to build reliable and maintainable ETL pipelines with declarative syntax. The platform also supports CI/CD and MLOps

High-level use cases for the Data Intelligence Platform on AWS

Databricks LakeFlow Connect offers built-in connectors for ingestion from enterprise applications and databases. The resulting ingestion pipeline is governed by Unity Catalog and is powered by serverless compute and Delta Live Tables. LakeFlow Connect leverages efficient incremental reads and writes to make data ingestion faster, scalable, and more cost-efficient, while your data remains fresh for downstream consumption.

Use case: Ingestion with Lakeflow Connect:

Ingestion with Lakeflow Connect on Databricks on AWS.

Download: Batch ETL reference architecture for Databricks on AWS

Use case: Batch ETL

Batch ETL reference architecture on Databricks on AWS.

Download: Batch ETL reference architecture for Databricks on AWS

Ingest tools use source-specific adapters to read data from the source and then either store it in the cloud storage from where Auto Loader can read it, or call Databricks directly (for example, with partner ingest tools integrated into the Databricks lakehouse). To load the data, the Databricks ETL and processing engine - via DLT - runs the queries. Single or multitask workflows can be orchestrated by Databricks Jobs and governed by Unity Catalog (access control, audit, lineage, and so on). If low-latency operational systems require access to specific golden tables, they can be exported to an operational database such as an RDBMS or key-value store at the end of the ETL pipeline.

Use case: Streaming and change data capture (CDC)

Spark structured streaming architecture on Databricks on AWS.

Download: Spark structured streaming architecture for Databricks on AWS

The Databricks ETL engine Spark Structured Streaming to read from event queues such as Apache Kafka or AWS Kinesis. The downstream steps follow the approach of the Batch use case above.

Real-time change data capture (CDC) typically uses an event queue to store the extracted events. From there, the use case follows the streaming use case.

If CDC is done in batch where the extracted records are stored in cloud storage first, then Databricks Autoloader can read them and the use case follows Batch ETL.

Use case: Machine learning and AI

Download: Machine learning and AI reference architecture for Databricks on AWS

For machine learning, the Databricks Data Intelligence Platform provides Mosaic AI, which comes with state-of-the-art machine and deep learning libraries. It provides capabilities such as Feature Store and model registry (both integrated into Unity Catalog), low-code features with AutoML, and MLflow integration into the data science lifecycle.

All data science-related assets (tables, features, and models) are governed by Unity Catalog and data scientists can use Databricks Jobs to orchestrate their jobs.

For deploying models in a scalable and enterprise-grade way, use the MLOps capabilities to publish the models in model serving.

Use case: Retrieval Augmented Generation (Gen AI)

Download: Gen AI RAG reference architecture for Databricks on AWS

For generative AI use cases, Mosaic AI comes with state-of-the-art libraries and specific Gen AI capabilities from prompt engineering to fine-tuning of existing models and pre-training from scratch. The above architecture shows an example of how vector search can be integrated to create a RAG (retrieval augmented generation) AI application.

For deploying models in a scalable and enterprise-grade way, use the MLOps capabilities to publish the models in model serving.

Use case: BI and SQL analytics

Download: BI and SQL analytics reference architecture for Databricks on AWS

For BI use cases, business analysts can use dashboards, the Databricks SQL editor or specific BI tools such as Tableau or Amazon QuickSight. In all cases, the engine is Databricks SQL (serverless or non-serverless) and data discovery, exploration, lineage, and access control are provided by Unity Catalog.

Use case: Lakehouse federation

Download: Lakehouse federation reference architecture for Databricks on AWS

Lakehouse federation allows external data SQL databases (such as MySQL, Postgres, or Redshift) to be integrated with Databricks.

All workloads (AI, DWH, and BI) can benefit from this without the need to ETL the data into object storage first. The external source catalog is mapped into the Unity catalog and fine-grained access control can be applied to access via the Databricks platform.