Skip to main content
Azure Databricks networking guide

Azure Databricks Private Link vs NCC choose the pattern by traffic flow

Private Link and Network Connectivity Configurations are not competing controls. The right design depends on who is connecting, which compute plane is involved and where the traffic needs to go.

Read the networking guide
Inbound Private Link Classic compute Serverless NCC
Databricks network flow map
Direction matters
Origin
Control
Destination

User / BI / API

inbound access

Private Link

Workspace

UI and APIs

Classic compute

cluster control traffic

Back-end PL

Control plane

REST / SCC relay

Serverless compute

outbound data access

NCC + PE

Azure resource

Storage / SQL / Key Vault

Classic compute

customer VNet path

VNet + PE

Azure / on-prem

data sources and services

Start with the traffic path. Then choose the network control that protects that path.

Start with the flow

Private Link and NCC solve different parts of the network

Private Link creates private network paths. NCC is the Databricks account-level construct used to manage serverless connectivity. The distinction becomes clear once the origin and destination are explicit.

User or application → workspace
How do I keep workspace access private?
Inbound Private Link
Classic compute → Databricks control plane
How do I keep cluster control traffic private?
Classic compute plane Private Link
Serverless compute → Azure resource
How do I privatize serverless data access?
NCC + private endpoint rule
Classic compute → Azure resource or on-prem
How do I control the customer VNet data path?
VNet injection + PE / routing

Inbound access corridor

Private access from the corporate network to Databricks.

Client

Corporate user

VPN / ExpressRoute

Transit

Private endpoint

databricks_ui_api

Destination

Workspace

UI · REST API

Browser authentication

Private browser sign-in needs the required authentication endpoint and DNS path.

Public access is separate

Creating a private endpoint does not by itself make the workspace private-only.

Inbound Private Link protects the path into Databricks. It does not define how compute reaches your data sources.

Inbound Private Link

Use Private Link when the workspace itself must be reached privately

Inbound Private Link secures user, API and application access to Azure Databricks. The private endpoint normally lives in a transit network that users can reach from the corporate network.

Protect the workspace entry point

Web UI and REST API traffic can stay on private networking.

Plan authentication and DNS together

The endpoint exists only if clients resolve and route to it correctly.

Decide whether public access remains available

Private connectivity and public-access policy are separate design choices.

Classic compute

Classic compute has two different private paths

VNet injection places classic compute in your network. From there, one path goes to the Databricks control plane and another goes to storage, databases or on-premises systems.

Control-plane path

Classic compute plane Private Link keeps REST and secure-cluster-connectivity relay traffic on private networking.

Data path

Private endpoints, routes, peering and VPN or ExpressRoute control access from the customer VNet to data sources.

Do not combine the two mentally

A private control-plane path does not automatically create a private path to ADLS, SQL or an on-premises ERP.

Classic compute path split

One cluster, two network responsibilities.

Customer VNet

Classic compute

Path A

To Databricks control plane

Back-end Private Link

REST API and SCC relay traffic.

Path B

To customer resources

Private endpoint Routes VPN / ER

ADLS, SQL, services and on-premises sources.

VNet injection gives classic compute customer-managed networking. It does not apply to serverless compute.

Serverless networking

NCC is the serverless network boundary, not a replacement for Private Link

A Network Connectivity Configuration is an account-level regional object. Workspaces attach to it so serverless compute can use the network controls defined for that NCC.

Attached workspaces

Manufacturing DEV
Manufacturing PRD

Regional control

NCC

workspace attachments
private endpoint rules

Private destinations

Azure Storage
Azure SQL / Key Vault

10 NCCs

current maximum per region per account

100 private endpoints

per region, distributed across NCCs

50 workspaces

current maximum attachments per NCC

Scope

Regional and account-level

The NCC can be attached to several workspaces in the same region.

Isolation

Separate where the boundary differs

Dev, test, prod or business units often deserve separate NCCs when their network policy differs.

Direction

Outbound only

NCC does not secure inbound access from users to the Databricks workspace.

Serverless egress

NCC is only one part of serverless egress control

The network pattern depends on whether the destination needs a private endpoint, a controlled Azure service path or a restricted external destination.

Private path

Azure PaaS

Use NCC private endpoint rules for supported Azure resources when traffic must stay on private connectivity.

NCC + PE

Private VNet path

Internal service

Serverless compute can privately reach supported resources in a customer VNet through the Databricks private connectivity pattern and a Standard Load Balancer.

NCC + LB

Azure firewall path

Storage

Service-endpoint access and Azure Network Security Perimeter rules can control storage access without a private endpoint.

NSP / service tag

External egress

FQDN / IP

Serverless network policies restrict outbound destinations and apply even when the path uses a private endpoint.

Network policy

Private connectivity and egress authorization are separate layers. A private endpoint does not automatically allow the destination.

Decision guide

Choose the networking pattern in three questions

Do not start with a product name. Start with direction, compute plane and destination.

Question 1

Is the traffic going into Databricks or out from compute?

Into Databricks

User, BI client or API

Use inbound Private Link when the workspace entry path must be private.

Out from compute

Question 2: classic or serverless?

Classic

Use VNet controls for customer resources and back-end Private Link for control-plane isolation.

Serverless

Question 3: does the destination require a private path?

Serverless + private destination

Use NCC private endpoint rules

Serverless + controlled non-private path

Use the appropriate firewall, NSP and network-policy controls

Hybrid manufacturing network

Two compute planes, two private connectivity patterns.

Plant

ERP / MES

on-premises

Classic path

VNet compute

VPN / ExpressRoute

Landing

ADLS

private endpoint

Analytics

Serverless

SQL / jobs

Serverless path

NCC

private endpoint rule

Serving

ADLS / Azure SQL

private data access

Food manufacturing example

A hybrid factory can need both classic VNet networking and NCC

A common manufacturing design combines private plant connectivity with serverless analytics. The mistake is assuming one network pattern covers both compute planes.

Operational ingestion

Classic compute can use the customer VNet to reach ERP, MES or file services over VPN or ExpressRoute.

Serverless analytics

Serverless SQL and jobs use NCC-managed connectivity to private Azure destinations.

One platform, separate paths

The data platform can be unified while the network controls remain compute-plane specific.

Validation

Validate the private path layer by layer

A private endpoint can exist while the workload still fails. Test name resolution, routing, endpoint state and identity separately.

01

Resolve

Does the hostname resolve to the expected private destination?

02

Route

Can the actual workload reach that network path?

03

Establish

Is the private endpoint approved and established on the target resource?

04

Authorize

Does the Databricks identity have permission on the data or service?

Test from the same compute plane that will run in production. A successful laptop test or classic-cluster test does not prove serverless connectivity.

Automation

Automate NCC after the network boundary is clear

This guide focuses on architecture. Once the serverless path is decided, the NCC, private endpoint rules and workspace binding can be automated with the Databricks Terraform provider.

Read the NCC Terraform guide

Terraform implementation

Three Databricks account resources

databricks_mws_network_connectivity_config
databricks_mws_ncc_private_endpoint_rule
databricks_mws_ncc_binding

Keep the implementation guide separate so architecture decisions and Terraform mechanics do not compete on the same page.

Common mistakes

Avoid networking designs that mix compute planes and traffic directions

Most failures come from applying the right control to the wrong path, or validating the design from a different workload than the one that will run in production.

Avoid

Treating NCC as a replacement for Private Link

Better default

Classify the traffic direction first

Avoid

Reusing classic VNet assumptions for serverless

Better default

Design serverless egress as its own network plane

Avoid

Creating a private endpoint and assuming public access is disabled

Better default

Configure private connectivity and public-access policy separately

Avoid

Testing from a laptop or classic cluster only

Better default

Validate from the actual production compute plane

Avoid

Sharing one NCC across unrelated security boundaries

Better default

Separate NCCs when environments or business units require isolation

Avoid

Troubleshooting RBAC before DNS and routing

Better default

Validate resolve, route, endpoint state and identity in order

How Food For Analytics implements it

We design the network around the workload path, not around one connectivity feature

In Titan, classic and serverless compute can coexist. We define the network boundary for each workload, automate the required controls and validate the path from the compute that will actually use it.

01 · Map

Classify every traffic path

Inbound, classic control-plane, classic data-plane and serverless egress are treated separately.

02 · Implement

Apply the right control

Private Link
NCC
VNet routing
DNS

03 · Validate

Prove the end-to-end path

DNS, routing, private endpoint state and data permissions are tested from the real workload.

FAQ

Frequently asked questions

Practical answers about Azure Databricks Private Link, NCC and serverless networking.

Is Databricks NCC the same as Azure Private Link?

No. Azure Private Link is the private connectivity technology. A Databricks Network Connectivity Configuration is an account-level regional construct used to manage serverless network connectivity, including private endpoint rules.

When do I need an NCC?

Use an NCC when Azure Databricks serverless compute needs controlled connectivity to customer resources. Private endpoint rules inside the NCC provide private connectivity to supported Azure resources, while other serverless controls can govern service-endpoint, firewall and external egress paths.

Does NCC replace VNet injection?

No. VNet injection applies to classic compute and places classic compute resources in a customer-managed VNet. Serverless compute runs in a Databricks-managed network and uses serverless networking controls such as NCC.

Does configuring Private Link automatically disable public access?

No. Private connectivity and public network access are separate controls. If the security requirement is private-only access, configure the relevant public-access or ingress restrictions as part of the design.

What is the difference between classic compute plane Private Link and a private endpoint to ADLS?

Classic compute plane Private Link secures traffic from classic compute to Azure Databricks control-plane services. A private endpoint to ADLS secures the data path from the compute plane to the storage account. They protect different flows.

Can serverless compute privately reach resources inside a VNet?

Yes, for supported patterns. Azure Databricks can establish serverless private connectivity to resources exposed behind a supported Standard Load Balancer in the customer VNet.

Can one NCC be shared across workspaces?

Yes. An NCC can be attached to multiple workspaces in the same region, subject to current service limits. Separate NCCs are usually clearer when dev, test, prod or business units require distinct isolation boundaries.

How many private endpoints can an NCC design use?

Current Azure Databricks limits allow up to 10 NCCs per region per account, 100 private endpoints per region distributed across those NCCs, and up to 50 workspace attachments per NCC. Verify current limits before production rollout.

Practical next step

Validate your Databricks network path before production

We can review the traffic flows, compute planes, private endpoints, DNS, routing and serverless egress controls before they become part of the production platform.

Read the NCC Terraform guide