Databricks Launches Private Network Gateway for Serverless Compute

Why networking held back serverless

Databricks serverless compute is described by the company as fully managed, intelligent and versionless. According to Databricks, the remaining high-friction step for many teams was networking: giving serverless compute a private path to databases, applications and systems inside their own networks.

The company says that in its customer interviews, more than 75% named their current network configuration and the complexity of dealing with network changes as the main reason specific workloads stayed on classic compute instead of moving to serverless.

The difference comes from architecture. Classic compute runs inside a customer's own virtual network, so reaching data is a matter of peering networks with standard cloud primitives. Serverless runs in a Databricks-managed network, which removes operational overhead but means private resources need an explicit path.

Serverless Private Link provides that path and, according to Databricks, remains the right tool for cloud-managed services such as AWS S3 or Azure Data Lake Storage. For resources customers host themselves, such as databases, Kafka clusters or internal APIs, connecting often meant a separate private endpoint, load balancer and proxy VM for each resource. It also required coordination between the Databricks admin, the cloud admin and the network security team. Databricks says this per-resource approach becomes a bottleneck when a team needs access to dozens of databases, APIs and internal services across VNets, peered networks and on-premises environments.

How Private Network Gateway works

With PNG, a customer delegates one subnet in their virtual network to Databricks, and PNG establishes a managed, encrypted gateway from serverless compute into that subnet. The gateway is configured inside a Network Connectivity Configuration (NCC), the account-level object already used for serverless networking. The NCC is then attached to workspaces, and supported serverless workloads can reach the private destinations the customer allows.

Databricks creates and runs the gateway, so there are no gateway VMs for customers to deploy, patch or scale. The company says PNG scales by network, not by endpoint, and that existing routes, DNS and firewall rules determine what is reachable.

Databricks lists several use cases:

• Private database ingestion: use Lakeflow Connect on serverless compute to ingest from supported databases across a VNet/VPC and connected on-premises environments.
• AI agents: an agent built on Databricks can call a private MCP server hosted by the customer.
• Private package repositories: serverless can install Python libraries from a private PyPI, Artifactory or Nexus repository inside the customer network.
• Own firewall for egress: traffic can be sent through an existing firewall or content inspection service, so egress is filtered and logged with the customer's own tools.
• Customer-owned source IP: SaaS apps such as Salesforce can see a source IP that belongs to the customer.

Control stays with the customer

Databricks says the gateway is dedicated to the customer, not a shared appliance carrying other customers' traffic. Isolation is enforced via encapsulation and network security rules, and the gateway enforces workspace-level access controls before any traffic leaves. Beyond the subnet, the customer's route tables, network security groups, DNS resolvers and security appliances decide what serverless can reach.

Customers choose between two modes. The specific_destinations mode sends only listed fully qualified domain names through PNG, while all other traffic keeps its normal path. The all_traffic mode routes outbound traffic through the customer network, except anything matching a more-specific Private Link or service-endpoint path, which keeps its native route. Databricks suggests this mode for inspecting broad egress through a firewall or sending it out through the customer's own IP addresses.

PNG is additive, according to Databricks. It sits alongside existing private connectivity and egress controls without changing them, so it can be enabled or removed without affecting the existing setup.

Fabiano Seki of Banco Bradesco said the gateway means "no more standing up a private link service and load balancer for every connection." Primož Gale of NLB and Michel Schwandner of Julius Blum GmbH also gave positive feedback in the announcement.

Availability and what it means

PNG is in Public Preview on Azure, with documentation available from Databricks. On AWS it is in Private Preview, and customers should contact their Databricks account team to discuss enrollment. Support for Google Cloud is in development.

For teams that have kept workloads on classic compute because of network complexity, the release targets that specific barrier. Since it is still a preview, organizations will need to evaluate it against their own network and security requirements.