Lakebase

Learning Objectives

• Understand what Lakebase is and why Databricks built an operational database on top of its analytics lakehouse
• Identify the features that make Lakebase suited to AI agents — serverless Postgres, instant branching, and scale-to-zero
• Evaluate when a lakehouse-native operational database is the right choice versus a standalone Postgres service

What Is Lakebase?

Lakebase is a fully managed, serverless Postgres database from Databricks, built for the applications and AI agents that run on top of modern data platforms. It reached general availability in February 2026 and represents a deliberate expansion for Databricks: the company built its business on analytical workloads — large-scale queries over a data lakehouse — and Lakebase adds the operational side, the fast, transactional reads and writes that live applications need.

Lakebase is built on Neon, the serverless-Postgres company Databricks acquired in 2025. The strategic bet behind that deal is striking: before the acquisition, roughly 80 percent of databases created on Neon were spun up by AI agents rather than human developers. Agents that write code, build apps, and manage state need somewhere to put that state — and they need to create and tear down databases far faster than a human ever would.

Access

Lakebase runs inside the Databricks platform rather than as a standalone product. Access details:

There is no standalone signup separate from Databricks — Lakebase is provisioned from within a Databricks workspace, and it shares that workspace's security and governance.

Core Capabilities

Serverless Postgres with Scale-to-Zero

Lakebase separates compute from storage, so database compute can scale up under load and shut off entirely when idle. Combined with sub-second start times, this "scale-to-zero" behavior means you pay for compute only when the database is actually working — a natural fit for bursty agent workloads that sit idle between tasks.

Instant Database Branching

Lakebase can create isolated, zero-copy clones of a production database in seconds. Because the branch shares underlying storage until it is written to, spinning up a full copy for testing, a new agent, or a preview environment is cheap and near-instant — the same branching workflow developers expect from version control, applied to live data.

AI-Agent State and Memory

Lakebase is designed as the transactional engine for AI applications, giving agents persistent memory that stays consistent with the lakehouse. It works as a supported checkpointer backend for agent frameworks such as LangGraph, so an agent's working state is durably stored and recoverable rather than living only in memory.

Native Lakehouse Sync

Sync tables keep data moving between Lakebase and the Databricks lakehouse in both directions, so operational data an app writes can flow back into analytics without a separate pipeline, and analytical results can be served back to the app. Because it supports pgvector, Lakebase can also store embeddings for retrieval-augmented generation (RAG) directly alongside application data.

Recovery and Reliability

Automated backups and point-in-time recovery let teams restore a database to a specific millisecond within a configurable retention window — protecting against application bugs or accidental deletions without manual snapshot management.

Pricing

Lakebase billing is usage-based, measured in Databricks Units (DBUs) based on the capacity-unit hours a workload consumes, with storage billed separately. Because it is part of the Databricks platform, cost sits inside your existing Databricks account rather than a separate database bill.

Strengths

• Lakehouse-native: Operational data and analytics share one governed platform — no separate database to secure, sync, and manage
• Built for agents: Fast database creation, branching, and scale-to-zero match how AI agents actually spin up and discard state
• Standard Postgres: It is real Postgres 17 with pgvector and the usual extensions, so existing tools, drivers, and skills transfer directly
• Cost efficiency at idle: Scale-to-zero means bursty or intermittent workloads do not pay for constantly-running compute
• Unified governance: Unity Catalog governs Lakebase data alongside the rest of the lakehouse

Limitations & Considerations

• Databricks-bound: Lakebase's advantages assume you are already on Databricks — it is not a general-purpose database you would adopt in isolation
• Cloud availability still expanding: Generally available on AWS, with Azure in beta and Google Cloud still to come at launch
• Newer operational track record: Databricks' analytics heritage is deep, but a managed operational database is a newer surface than long-established Postgres services
• Usage-based cost modeling: DBU-based billing takes some modeling to predict for steady high-throughput workloads compared with a flat provisioned instance elsewhere

Best Use Cases

When to choose alternatives:

• Standalone serverless Postgres outside Databricks → Neon or Supabase offer the same engine without the platform tie
• Pure vector search at scale → A dedicated vector database like Pinecone or Weaviate
• Established managed Postgres on a specific cloud → Cloud-native services (Amazon Aurora, Azure Database for PostgreSQL) are mature and well understood

Key Takeaways

• Lakebase is Databricks' fully managed, serverless Postgres database, generally available since February 2026, bringing an operational database into the analytics lakehouse
• It is built on Neon, the serverless-Postgres company Databricks acquired — a bet on a future where AI agents, not humans, create most databases
• Scale-to-zero, instant zero-copy branching, pgvector, and LangGraph support make it well suited to AI agents and the apps around them
• Its value is strongest for teams already on Databricks who want operational data and analytics governed on one platform, rather than as a standalone Postgres service