Edit

Microsoft Discovery architecture decisions and fixed settings

You fix several Discovery settings when you create a resource, including deployment-time tags such as NetworkIsolation, the workspace and cluster identities, the encryption model, the supercomputer outbound type, and subnet bindings. Changing these settings later means deleting and re-creating the workspace, supercomputer, or bookshelf. The two Workbench feature tags are separate and can be updated on an existing workspace. This article lists those decisions, the options for each, and where to find the configuration details, so you can agree on them with your security, network, and platform owners before your first deployment.

For deployment paths, prerequisites and their owners, and the onboarding tool stages, see Microsoft Discovery deployment planning overview.

Settings fixed at creation

Setting Resource Can you change it after creation? More information
Deployment-time tags, such as NetworkIsolation, discovery.overridemrgregion, and discovery.enableprivatecluster Workspace, supercomputer, node pool, bookshelf Set these deployment settings before creation. The workspace's Workbench feature tags can be edited separately. Quickstart: Deploy infrastructure
Workspace identity Workspace No Managed identities
Cluster identity Supercomputer No Managed identities
Kubelet and workload identities Supercomputer Yes, by using PATCH Managed identities
Customer-managed keys or Microsoft-managed keys Workspace, supercomputer, bookshelf No Configure customer-managed keys
outboundType and managementSubnetId Supercomputer No Configure secure networking for a supercomputer
Subnets Workspace, supercomputer, node pool, bookshelf No Configure network security
VM size and scale set priority Node pool No. Add another node pool instead. Manage supercomputers by using REST APIs
publicNetworkAccess Workspace, bookshelf Yes Disable public network access
Private endpoints and private DNS Workspace, bookshelf Yes Create private endpoints
Role assignments for users and groups All Yes Role assignments

Data-plane network access

The system enables network hardening of managed resources by default for every workspace and bookshelf. You decide whether clients reach the data plane privately.

Option What you get Choose it when
Quickstart deployment Managed resources protected by a Network Security Perimeter (NSP), public data-plane endpoints protected by Microsoft Entra ID Evaluations and environments without private-network requirements
Network-hardened deployment Private endpoints for the workspace and bookshelf data planes, and network-hardened customer storage Regulated data, landing zones that deny public endpoints, or private-only access requirements

Plan the client access path, such as VPN, Azure ExpressRoute, or virtual network peering, before you set publicNetworkAccess to Disabled. With public access disabled, requests from the public internet receive 403 Forbidden. Private endpoints must be in the same region as the Discovery resource. For details, see Network security.

Supercomputer egress and management plane

The outboundType setting and the management subnet decide how supercomputer traffic leaves your virtual network and how the Discovery service reaches the cluster management plane.

Option Outbound traffic Management plane Requirements
LoadBalancer (default) Unrestricted outbound access through a managed public IP Public IP. Register the Microsoft.ContainerService/EnableServiceTagAuthorizedIPPreview feature to restrict it to the Discovery service. Management subnet optional, recommended for defense in depth
UserDefinedRouting Through your Azure Firewall or network virtual appliance, with no managed public IP for egress Public IP restricted to the Discovery service when the feature is registered Management subnet, route table, and firewall rules in place before you create the supercomputer
Private supercomputer Through your Azure Firewall or network virtual appliance No public IP. Discovery reaches the management plane over Azure Private Link. UserDefinedRouting, a management subnet, Private Link service network policies disabled, a default route to Azure Firewall or a network virtual appliance, and the discovery.enableprivatecluster tag

If an Azure Policy assignment denies public IP creation, use UserDefinedRouting. For more information, see Known issues.

For the firewall rules and management subnet requirements, see Configure secure networking for a supercomputer and Management subnet ID requirements.

Identities

Option Trade-off
One shared user-assigned managed identity Simplest to set up. The workload identity, which agents use through tool execution, inherits every role, including Discovery Platform Contributor and AcrPull.
Separate identities for each slot More setup. The workload identity receives only Storage Blob Data Contributor on the storage account.

Use separate identities when your organization requires least-privilege access, when tools process sensitive data, or when you must separate control-plane and data-plane identities. For the recommended role assignments for each identity, see Granular role assignments per identity.

Encryption keys

Microsoft-managed keys require no action. Customer-managed keys need an Azure Key Vault with soft delete and purge protection, an RSA key, and a user-assigned managed identity with the Key Vault Crypto Service Encryption User role, all in the same region as the Discovery resource.

Choose the key model when you create each workspace, supercomputer, and bookshelf. Discovery requires an explicit key version, so plan to update each Discovery resource after you rotate the key. For details, see Configure customer-managed keys and Rotate a customer-managed key.

Regions

Microsoft Discovery supports the following home regions for its control-plane resources:

  • East US
  • Sweden Central
  • UK South

Region support doesn't guarantee model or compute capacity.

Option Resource placement
Single region All resources in one supported home region
Cross-region Discovery resources and their managed resource groups in a supported home region, the resources inside the managed resource groups, and your virtual network, storage, container registry, and identities, in one target region

For a cross-region deployment, apply the discovery.overridemrgregion tag when you create each workspace, supercomputer, and bookshelf, and confirm that the target region supports the required Microsoft Foundry features. For details, including how data moves between regions, see Deploy Microsoft Discovery across Azure regions.

Workspaces, virtual networks, and subnets

  • Each workspace, bookshelf, and supercomputer requires its own non-overlapping subnets. You can't share delegated subnets between Azure Container Apps environments. See Subnet requirements for workspaces and bookshelves.
  • The quickstart uses a separate virtual network for each workspace. See Quickstart: Deploy infrastructure.
  • Leave room in the private endpoint subnets for future endpoints, because you can't change subnets after creating the resources that use them.
  • Projects in a workspace share its supercomputer and chat model deployments. Use project-level access control to separate teams within a workspace, and separate workspaces when network, identity, or model boundaries must differ.

Workbench features

Set NetworkIsolation when you create the workspace. You can set the two Workbench feature tags at creation or update them later without changing network isolation.

Tag Effect
NetworkIsolation Set to false to enable public access to the preview workbench and other managed resource group resources. When you keep network isolation enabled, connect to the workspace's virtual network through VPN or ExpressRoute to use the workbench.
discovery.workbench.enableGhcpAiFeatures Enables GitHub Copilot and AI features. Users need a GitHub account.
discovery.workbench.enableExtensions Enables the Visual Studio Code Extension Marketplace.

The Bicep template in the quickstart sets the GitHub Copilot and extension tags to true. Set them to false if your organization hasn't approved these features. For details, see Use GitHub Copilot in Microsoft Discovery.

Project storage

  • Discovery Studio downloads files directly from the browser to Blob Storage, so plan a network path from your clients to the storage account. See Configure project storage.
  • For private-only storage, use the network-hardened customer storage configuration.
  • Choose a redundancy option that meets your residency requirements. Geo-redundant options copy data to a secondary region.

Compute and quota

  • Plan model, VM family, and Azure Container Apps quota before deployment. See Quota reservations.
  • Use Regular priority for jobs that can't restart, and Spot only for restartable work.
  • Discovery manages AKS upgrades automatically. Make long-running tools restartable. See Supercomputer AKS upgrades.

Bookshelf

  • A bookshelf needs its own search subnet and private endpoint subnet.
  • Choose the indexSize tag before you create the bookshelf. See Resource quota requirements.
  • Bookshelf infrastructure, including Azure AI Search, Azure Container Apps, and Azure SQL Database, runs continuously after deployment. See Quota reservations.

Business continuity

Discovery doesn't replicate customer state between regional deployments. To recover from a regional outage, plan a secondary deployment with replication for your storage accounts and container registry. See Business continuity and disaster recovery.