Databricks security architecture on AWS

Summary

• Databricks on AWS uses a two-plane architecture: a Databricks-managed control plane for workspace management, orchestration, and governance, and a compute plane that runs either as classic compute in your own AWS VPC or as serverless compute in the Databricks account.
• It is secure by default: secure cluster connectivity (SCC) means customer VPCs have no open ports and cluster nodes have no public IP addresses, with encryption in transit and authenticated workspace access enabled automatically.
• Private networking uses a customer-managed VPC plus AWS PrivateLink for private control-plane connectivity, so REST API and SCC relay traffic can stay off the public internet.
• Data is encrypted at rest with support for customer-managed AWS KMS keys, and Unity Catalog governs data and AI access with fine-grained RBAC and ABAC controls, lineage, and auditing across workspaces.
• Network controls scale through three deployment architectures (Managed security, Hardened connectivity, and Isolated environment), and Databricks publishes a Security Reference Architecture with Terraform templates.

Databricks security architecture on AWS

The Databricks security architecture on AWS is built on a two-plane model with security enforced by default and layered network controls you can tighten as your requirements grow. The control plane is the Databricks-managed layer that handles workspace management, cluster orchestration, job scheduling, and governance. The compute plane is where your data is processed, and it runs in one of two ways: classic compute in your own AWS VPC, or serverless compute hosted in the Databricks account. This separation, combined with encryption, private connectivity, and Unity Catalog governance, lets you keep your data in your trust boundary while Databricks operates the platform.

Why the Databricks Platform is secure on AWS

• Secure cluster connectivity (SCC) by default. All new workspaces enable SCC: customer VPCs have no open ports and classic compute nodes have no public IP addresses. Clusters connect to the control plane through a secure tunnel over HTTPS (port 443) via the SCC relay, which separates web and REST API traffic from cluster management. Serverless compute resources also have no public IP addresses.
• Customer-managed VPC. Deploy Databricks into your own VPC for control over IP ranges, routing, and security groups, and to integrate with your existing network architecture. A customer-managed VPC is required to enable classic PrivateLink.
• AWS PrivateLink. Add classic compute plane PrivateLink so REST API and SCC relay traffic between your clusters and the control plane travels over a private network instead of the public internet. Front-end (workspace UI and API) private access is also supported for fully private connectivity.
• Encryption in transit and at rest. Connections between the control plane and compute plane use transport-level encryption, and control-plane and SCC relay connections use TLS with certificate pinning. For data at rest, Databricks supports customer-managed AWS KMS keys for workspace storage, managed services, and Unity Catalog external locations. See Security and compliance.
• Governance with Unity Catalog. Unity Catalog provides unified governance for data and AI with fine-grained RBAC and ABAC access control, centralized policy management, auditing, and lineage. These controls apply consistently across all network architectures and workspaces.
• Layered network architectures. Choose the level of isolation you need across three deployment architectures: Managed security (baseline, with SCC and a customer-managed VPC), Hardened connectivity (adds context-based ingress, VPC endpoints, and serverless egress controls), and Isolated environment (adds inbound PrivateLink and a required firewall for full private connectivity). Classic compute egress can optionally be routed through an external firewall such as AWS Network Firewall for inspection and policy enforcement.
• Independently tested. Databricks tests the platform with annual third-party penetration tests and runs a public bug bounty program, under a shared responsibility model for secure deployment.

Getting started

• Read the Network reference architecture overview to understand the control plane, compute plane, and the three deployment architectures.
• Start with Managed security, then add classic PrivateLink and a customer-managed VPC as your isolation requirements grow.
• Deploy with the Databricks Security Reference Architecture (Terraform) and review Security and compliance for encryption and controls.

FAQs

What are the control plane and compute plane on AWS?

The control plane is the Databricks-managed layer for workspace management, orchestration, and governance. The compute plane processes your data and runs either as classic compute in your own AWS VPC or as serverless compute in the Databricks account.

Does Databricks require public IP addresses on AWS?

No. With secure cluster connectivity, enabled by default, customer VPCs have no open ports and classic compute nodes have no public IP addresses; serverless compute resources also have no public IP addresses.

How does Databricks keep AWS network traffic private?

Deploy into a customer-managed VPC and enable AWS PrivateLink so REST API and SCC relay traffic between your clusters and the control plane stays on a private network, with front-end PrivateLink available for private workspace UI and API access.

How is data encrypted on Databricks for AWS?

Traffic between the control and compute planes is encrypted in transit with TLS, and data at rest can be protected with customer-managed AWS KMS keys for workspace storage, managed services, and Unity Catalog external locations.