What is Lakebase — Databricks Serverless Postgres

Lakebase is the serverless Postgres database that Databricks took to GA (general availability) in 2026. It is a full-featured OLTP service positioned as an alternative to Aurora, Cloud SQL, and Cosmos DB, but its decisive differentiator is native integration with Lakehouse data via Unity Catalog. Target use cases span AI agent backends, production databases for SaaS applications, and analytics on live production data.

This article walks through every Lakebase feature, the pricing model, comparisons with other managed Postgres services, how Lakehouse Sync behaves, production best practices, and the certification exam outlook — all aligned with the latest 2026 release.

What is Lakebase

Lakebase is a serverless Postgres service integrated into the Databricks Data Intelligence Platform. At GA in 2026, "Lakebase Autoscaling" and "Lakebase Provisioned" were unified into a single UI, completing the following feature set.

• Autoscaling: Automatically scales up and down with the workload
• Scale-to-zero: Shrinks to zero during idle periods — zero cost
• Instant Branching: Clone production instantly to create a test environment
• Automated backups + Point-in-Time Recovery (PITR)
• Up to 8TB of storage per instance
• Built on Postgres 17: Fully compatible with the latest open-source Postgres
• pgvector built in: Supports AI search and RAG patterns
• Multiple Azure regions supported: Ready for global deployments

It is much more than a managed Postgres — the defining feature is native Lakehouse integration.

Why Lakebase Matters

Historically, enterprises have run OLTP (transactional) and OLAP (analytical) workloads on separate systems — OLTP on Aurora / Cloud SQL / Cosmos DB, OLAP on Snowflake / BigQuery / Databricks Lakehouse, and so on.

That separation has serious downsides.

• Data must be copied: OLTP → OLAP requires constant ETL via tools like DMS, Glue, or Snowpipe
• Latency is introduced: Minutes to hours of lag make real-time analytics impossible
• Governance is fragmented: Access controls have to be managed in two separate systems
• Operating costs double: You pay for two separate services and need expertise in both
• Poor fit for AI workloads: AI agents need both OLTP state management and a Lakehouse knowledge base

Lakebase resolves all of this. OLTP data is auto-replicated to Delta tables via Lakehouse Sync, so it becomes available for analytics queries the instant it is written. That is the real value of Lakebase.

Pricing — Autoscaling vs. Always-On

Lakebase pricing is per Capacity Unit (CU) hour priced in DBUs, plus separate storage charges. Two billing models are available.

Autoscaling Plan (fully usage-based)

Configure a min/max autoscale range plus a scale-to-zero timeout. Cost is zero while idle, making it ideal for development, testing, and low-traffic workloads.

• Cost: CU hours used × DBU rate + storage
• Scale-to-zero: automatically stops after the configured timeout
• Cold start: a few seconds to tens of seconds
• Typical use cases: dev environments, test environments, internal tools, PoCs

Always-On Plan (25% discount on baseline capacity)

Designed for always-on production. Baseline capacity is discounted 25%, with bursts absorbed by autoscaling. No long-term commitment is required, yet you get a discount equivalent to Aurora Reserved Instances with zero friction.

• Cost: baseline capacity (25% off) + overage (standard rate)
• Always running: no scale-to-zero
• Autoscaling: spikes are absorbed automatically
• Typical use cases: SaaS production databases, internal core systems, AI agent backends

Storage Pricing

Storage is billed separately and covers Postgres data, WAL (Write-Ahead Log), and backups. Rates are in the same range as Aurora, with no explicit reserved purchase concept.

Head-to-Head: Aurora / Cloud SQL / Cosmos DB

vs Amazon Aurora PostgreSQL

• Where Aurora wins: Easy migration path for existing RDS users, extensive region coverage, and the maturity of Aurora Serverless v2
• Where Lakebase wins: Zero-ETL analytics via Lakehouse Sync, governance through Unity Catalog integration, AI agent-specific features, and immediate Postgres 17 support
• How to choose: AWS-centric with separate analytics (Redshift / Snowflake) → Aurora. Databricks Lakehouse as your analytics platform → Lakebase, no contest

vs Google Cloud SQL for PostgreSQL

• Where Cloud SQL wins: BigQuery integration (Federated Query) and tight affinity with the rest of the GCP ecosystem
• Where Lakebase wins: Multi-cloud reach (runs on AWS / Azure / GCP), scale-to-zero, and Lakehouse Sync
• How to choose: GCP-only with mid-scale data → Cloud SQL. Multi-cloud or Databricks-centric organizations → Lakebase

vs Azure Cosmos DB for PostgreSQL (Citus)

• Where Cosmos DB wins: Horizontal scale via distributed Postgres (Citus) and tight integration with other Azure services
• Where Lakebase wins: Lakehouse integration, Postgres 17 support (Cosmos is on 16), and scale-to-zero
• How to choose: Tens of terabytes with horizontal sharding needs → Cosmos DB. Up to ~8TB with a Lakehouse integration priority → Lakebase

vs Neon / Supabase

• Where Neon / Supabase win: Pricing aimed at individual developers and startups, plus strong community and ecosystem
• Where Lakebase wins: Enterprise-grade governance, Unity Catalog, Lakehouse integration, and AI-focused optimizations
• How to choose: Solo developers and small startups → Neon. Enterprises and existing Databricks customers → Lakebase

How Lakehouse Sync Works

Lakehouse Sync, in Public Preview as of 2026, uses CDC (Change Data Capture) to low-latency replicate Lakebase Postgres tables to Unity Catalog-managed Delta tables.

The traditional OLTP → OLAP data pipeline looked something like this:

1. Configure Debezium / DMS to read Aurora's WAL
2. Stream events to Kafka / Kinesis
3. Ingest with Snowpipe / Glue / Spark
4. Merge into Delta or Iceberg tables
5. BI / ML run analytics queries

Each step demands infrastructure, ongoing operations, and additional cost — and data lag runs from minutes to hours.

With Lakehouse Sync, all of this collapses into flipping a single toggle.

• Enable 'Lakehouse Sync' in the Lakebase project settings
• Select the tables you want to sync
• Specify the destination Unity Catalog catalog / schema
• Changes propagate to the Delta tables automatically within seconds to tens of seconds

As a result, order and inventory tables are available for real-time aggregation in ML models and BI dashboards the moment a write completes.

Lakebase as an AI Agent Backend

Managing state for AI agents has historically been hard to design with traditional databases. Here is why.

• Unpredictable scale: Requires instant response to bursty workloads
• State must be persisted: Conversation history, tool execution logs, user context
• Knowledge base integration: Access to Vector Search and RAG documents on the Lakehouse
• Low-latency writes: Directly impacts user-facing response time

Lakebase is optimized for all of the above.

• Scale-to-zero + autoscaling handles bursts while driving idle cost to zero
• Postgres + pgvector keeps state management and semantic search in the same database
• Lakehouse Sync streams conversation history to Delta tables for analytics and downstream iteration
• Low latency: Response times on par with managed Postgres

For storing state for AI agents that call the Anthropic or OpenAI APIs, this is one of the most refined options available today.

Getting Started

You can launch Lakebase from your Azure Databricks or AWS Databricks Workspace in just a few clicks.

1. Open 'Lakebase' from the Workspace sidebar
2. Click 'Create Project' to create a new Lakebase project
3. Choose Autoscaling or Always-On
4. Set the initial spec (Capacity Unit range + storage)
5. Connection details (host / port / password) are issued
6. Connect from psql, pgAdmin, or your application using the standard Postgres protocol

New Lakebase Autoscaling projects have Postgres password authentication disabled by default, and OAuth or Workspace ID-based secure connections are recommended. If you want to use password authentication, you must explicitly enable it in the project settings.

Production Best Practices

Capacity Planning

Set the Autoscaling minimum CU to your traffic floor and the maximum CU to roughly twice your peak. A scale-to-zero timeout of 5 minutes is standard for development and 30 minutes for production.

Backup Strategy

Automated backups support 7-day PITR by default. In production, the standard pattern is to validate pre-release schema changes with Instant Branching before applying them.

Authentication Strategy

Keep password authentication disabled and connect via OAuth + Service Principal. Issue a Service Principal per app and per environment, and apply the principle of least privilege.

Monitoring

Pull connection counts, query performance, and capacity utilization from Databricks System Tables. Integrate with your existing monitoring stack via CloudWatch or Azure Monitor.

Cost Optimization

For dev, use Autoscaling with a short timeout (5 minutes) to drive idle cost to zero. For production, use Always-On to capture the 25% baseline discount and let autoscaling absorb spikes.

Outlook for Certification Exams

As of May 2026, Lakebase is not in the official Databricks Data Engineer Associate / Professional Exam Guide. However, it is highly likely to enter the scope in a late-2026 or 2027 revision, for the following reasons.

• Databricks itself is promoting 'Lakehouse + OLTP' as a flagship platform direction
• GA was announced front and center at FabCon 2026
• Higher-level features like Unity Catalog and Delta Sharing already appear frequently on the existing exams, so Lakehouse Sync joining the scope feels natural

Recommended topics to study ahead of time:

• How Lakebase Autoscaling and Always-On differ and how to choose between them
• Lakehouse Sync's replication model and latency characteristics
• The 'OLTP × OLAP separation' problem that Lakebase solves
• Why Lakebase is advantageous for AI agent use cases

Frequently Asked Questions