How I replaced three Azure emulators with one binary, added Key Vault and ACR, and cut my CI setup to a single step

For a while, local Azure development in my projects looked like this: Azurite for Blob and Queue Storage, the Microsoft Service Bus Emulator for messaging, the Cosmos DB Emulator for document storage, and nothing at all for Key Vault, Container Registry, or Entra. The last three were either skipped in local runs, mocked, or simply hit against a real Azure subscription when I needed them.

That worked, to a point. But the setup cost was real. Each emulator has its own Docker image, its own port range, its own quirks, and its own certificate story. Compose files grew. CI pipelines grew to match. And the services that had no emulator at all stayed untested locally, which is exactly the category where surprises tend to show up in production.

I wanted to see whether a single tool could cover the whole stack without making too many compromises.

What the old setup actually looked like

The previous CI compose file had three separate service containers. Azurite covered BlobServiceClient, QueueServiceClient, and TableServiceClient. The Service Bus Emulator added a second container with its own volume and a RabbitMQ dependency. The Cosmos DB Emulator added a third, and that one needed a minute and a half to initialise before tests could start on x64 machines. On Apple Silicon it did not start at all: the Linux container image is x64-only, and under Docker's Rosetta emulation layer the embedded database engine crashes during startup (GitHub issue #54, open since July 2022 with 149 comments). There is a separate vNext Linux emulator that does run on ARM, but it is a different codebase still in preview with its own coverage gaps. In practice, anyone on an M-series Mac either skipped Cosmos DB local testing entirely, ran a Windows VM, or hit real Azure. Key Vault calls were either mocked or skipped regardless of platform.

The GitHub Actions workflow was about 120 lines and included three separate wait loops, one per emulator, each polling a different health endpoint on a different port. That is not the kind of complexity you notice until you have to debug it at 2am because a container did not start cleanly on a fresh runner image.

What "one binary" means in practice

Topaz is a self-contained binary (or a Docker image if you prefer containers) that runs a local emulation layer covering:

Blob Storage, Table Storage, Queue Storage (Azurite replacement)
Service Bus with AMQP, AMQPS, and the UseDevelopmentEmulator SDK flag
Cosmos DB (control plane + data plane, NoSQL API)
Key Vault (secrets, keys, certificates, RBAC-controlled access)
Container Registry (push, pull, login server, tasks)
Azure Resource Manager (ARM, Bicep, Terraform deployments)
Microsoft Entra ID tenant emulation (OIDC, ROPC, az login)
RBAC (role assignments, permission checks, managed identity flows)

The install story on macOS is one command:

brew tap thecloudtheory/topaz && brew install topaz && topaz-host

On Linux there is a dedicated install script:

curl -fsSL https://raw.githubusercontent.com/TheCloudTheory/Topaz/main/install/get-topaz.sh | bash
topaz-host

Or pull the Docker image:

docker run -p 8899:8899 thecloudtheory/topaz-host

Once it is running, GET /health on port 8899 returns 200. That is the signal every CI script needs.

The CI compose file now

Here is the compose file we use in GitHub Actions. The entire emulator layer is one service:

services:
  topaz:
    image: thecloudtheory/topaz-host:latest
    platform: linux/amd64
    ports:
      - "443:443"     # HTTPS — required for MSAL pre-flight during az login
      - "8899:8899"   # ARM / Resource Manager  (health: GET /health)
      - "8898:8898"   # Key Vault data-plane
      - "8891:8891"   # Blob Storage data-plane
      - "8895:8895"   # Cosmos DB
      - "8889:8889"   # Service Bus AMQP plain
      - "5671:5671"   # Service Bus AMQPS / TLS
    volumes:
      - ../../certificate:/certs:ro
    command:
      - --certificate-file
      - /certs/topaz.crt
      - --certificate-key
      - /certs/topaz.key
      - --log-level
      - Warning

That is it. One service, one image, one wait loop on /health:

for i in $(seq 1 15); do
  if curl -sf https://topaz.local.dev:8899/health; then
    echo "Topaz is ready"
    break
  fi
  echo "Waiting... ($i/15)"
  sleep 2
done

The workflow

The full GitHub Actions workflow is about 80 lines including comments. The relevant structure is:

Trust the Topaz self-signed certificate at the OS level and inside the Azure CLI Python bundle
Configure dnsmasq to resolve *.topaz.local.dev to 127.0.0.1 (one script, no per-resource /etc/hosts entries)
docker compose up -d
Poll https://topaz.local.dev:8899/health until it returns 200
Register Topaz as a custom Azure CLI cloud with az cloud register
az login --username topazadmin@topaz.local.dev --password admin
Run tests

Steps 1 through 3 replace the three container startup sequences plus the separate emulator configuration files we had before. Step 6 is new capability entirely. The old setup had no az login support at all because Azurite does not emulate Entra.

SDK integration: no code changes

The pattern is the same regardless of language: swap the credential for AzureLocalCredential and point the client at the local endpoint. No mocks, no conditional branches.

C# (.NET)

var credential = new AzureLocalCredential(Globals.GlobalAdminId);
var armClient  = new ArmClient(credential, subscriptionId);

The connection strings come back through TopazResourceHelpers, which resolves the correct local endpoint by resource name:

var blobUri   = TopazResourceHelpers.GetBlobServiceUri(storageAccountName);
var kvUri     = TopazResourceHelpers.GetKeyVaultEndpoint(vaultName);
var cosmosUri = TopazResourceHelpers.GetCosmosDbAccountEndpoint(cosmosAccountName);
var sbConn    = TopazResourceHelpers.GetServiceBusConnectionStringWithTls(serviceBusNamespace);

Python

The topaz_sdk package provides the same primitives. The azure-mgmt-* clients accept a base_url and credential_scopes override so no monkey-patching is needed:

from topaz_sdk import GLOBAL_ADMIN_ID, DEFAULT_RESOURCE_MANAGER_PORT, AzureLocalCredential

BASE_URL = f"https://topaz.local.dev:{DEFAULT_RESOURCE_MANAGER_PORT}"

def mgmt_kwargs(subscription_id):
    return {
        "credential": AzureLocalCredential(GLOBAL_ADMIN_ID),
        "subscription_id": subscription_id,
        "base_url": BASE_URL,
        "credential_scopes": [f"{BASE_URL}/.default"],
    }

# Key Vault data plane — point at the per-vault subdomain
secret_client = SecretClient(
    vault_url=f"https://{vault_name}.vault.topaz.local.dev:8898",
    credential=AzureLocalCredential(GLOBAL_ADMIN_ID),
    verify_challenge_resource=False,
)
secret_client.set_secret("db-password", "s3cr3t-password")

No environment variable juggling. No conditional if isDevelopment branches scattered across startup.

RBAC works too

This is the part I did not expect to be useful until I actually needed it. The SecretsRbac example in the repo shows the pattern: create a Key Vault, assign the Key Vault Secrets User built-in role to a managed identity principal, and then verify that the identity can read the secret while an unassigned identity gets a 403.

await armClient
    .GetRoleAssignments(kvScope)
    .CreateOrUpdateAsync(
        WaitUntil.Completed,
        Guid.NewGuid().ToString(),
        new RoleAssignmentCreateOrUpdateContent(
            new ResourceIdentifier($"/providers/Microsoft.Authorization/roleDefinitions/{kvSecretsUserRoleId}"),
            assignedPrincipalId)
        {
            PrincipalType = RoleManagementPrincipalType.ServicePrincipal
        });

Azure RBAC is one of those areas where the gap between a mock and real behavior shows up fastest. The mock returns 200. The real service returns 403 unless the exact scope hierarchy, role definition ID, and principal type are right. Topaz enforces the same checks locally, which means tests that pass locally actually mean something.

Coverage gaps worth knowing about

Cosmos DB data plane covers the NoSQL API. Mongo API and Table API users will hit gaps. SQL and App Service are control-plane only for now. Everything else in the list above, including Event Hubs messaging, works end-to-end. The API coverage docs have the full operation-level breakdown if you want to check a specific call before migrating.

The before and after

Before:

3 Docker images in CI compose (Azurite, Service Bus Emulator, Cosmos DB Emulator)
3 wait loops on 3 different ports
No az login, no Key Vault, no ACR in integration tests
Cosmos DB emulator does not start on Apple Silicon at all
Service Bus namespace provisioning on real Azure: 60-120 seconds, occasionally longer

After (measured on a GitHub-hosted ubuntu-latest runner):

1 Docker image, 1 wait loop
Full Entra login, Key Vault reads with RBAC enforcement, Container Registry push/pull
Topaz ready in 7 seconds (5s start + 2s health poll)
Full job including DNS setup, provisioning, and tests: 38 seconds
Works identically on Apple Silicon, x64, and CI runners

Where to go from here

The compose file, the workflow YAML, and the AllInOne example are all in the Topaz GitHub repo under Examples/. The brew install topaz path works for local development; the Docker path works for CI.

Azurite is one starting point, but the more interesting cases are the ones that had no local option at all. If your application touches Key Vault, Container Registry, Entra, RBAC, or ARM deployments, you are probably either skipping those paths in tests, mocking them, or incurring real Azure costs in CI. Topaz is designed for the full stack, not as a drop-in for a single service.

推荐订阅源

DEV Community