Azure Front Door and multi-region deployments

In the previous post, we looked at publishing and securing an API with Azure Front Door and Azure Web Application Firewall. The API ran on Kubernetes, exposed by Kong and Kong Ingress Controller. Kong was configured to require an API key to call the /users API, allowing us to identify the consumer of the API. The traffic flow was as follows:

Consumer -- HTTPS --> Azure Front Door with WAF policy -- HTTPS --> Kong (exposed with Azure Load Balancer) -- HTTP --> API Kubernetes service --> API pods 

Although Kubernetes makes the API(s) highly available, you might want to take extra precautions such as deploying the API in multiple regions. In this post, we will take a look at doing so. That means we will deploy the API in both West and North Europe, in two distinct Kubernetes clusters:

  • we-clu: Kubernetes cluster in West Europe
  • ne-clu: Kubernetes cluster in North Europe

The flow is very similar of course:

Consumer -- HTTPS --> Azure Front Door with WAF policy -- HTTPS --> Kong (exposed with Azure Load Balancer; region to connect to depends on Front Door configuration and health probes) -- HTTP --> API Kubernetes service --> API pods  

Let’s take a look at the configuration! By the way, the supporting files to deploy the Kubernetes objects are here: https://github.com/gbaeke/api-kong/tree/master. To deploy Kong, check out this post.

Kubernetes

We deploy a Kubernetes cluster in each region, install Helm, deploy Kong, deploy our API and configure ingresses and related Kong custom resource definitions (CRDs). The result is an external IP address in each region that leads to the Kong proxy. Search for “kong” on this blog to find posts with more details about this deployment.

Note that the API deployment specifies an environment variable that will contain the string WE or NE. This environment variable will be displayed in the output when we call the API. Here is the API deployment for West Europe:

apiVersion: v1
kind: Service
metadata:
  name: func
spec:
  ports:
  - port: 80
    protocol: TCP
    targetPort: 80
  selector:
    app: func
  type: ClusterIP
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: func
spec:
  replicas: 2
  selector:
    matchLabels:
      app: func
  template:
    metadata:
      labels:
        app: func
    spec:
      containers:
      - name: func
        image: gbaeke/ingfunc
        env:
        - name: REGION
          value: "WE"
        ports:
        - containerPort: 80 

When we call the API and Azure Front Door uses the backend in West Europe, the result will be:

WE included in the response from the West Europe cluster

Origin APIs

The origin APIs need to be exposed on the public Internet using a DNS name. Azure Front Door requires a public backend to connect to. Naturally, the backend can be configured to only accept incoming requests from Front Door. In our case, the APIs are available on the public IP of the Kong proxy. The following names were used:

  • api-o-we.baeke.info: Kong proxy in West Europe
  • api-o-ne.baeke.info; Kong proxy in North Europe

Both endpoints are configured to accept TLS connections only, and use a Let’s Encrypt wildcard certificate for *.baeke.info.

Front Door Configuration

The Front Door designer looks the same as in the previous post:

Front Door Designer

However, the backend pool api-o now has two backends:

Two backend hosts, both enabled with same priority and weight

To determine the health of the backend, Front Door needs to be configured with a health probe that returns status code 200. If we were to specify the probe below, the health probe would fail:

Errrrrrr, this won’t work

The health probe would hit Kong’s proxy and return a 404 (Not Found). We did not create a route for /, only for /users. With Azure Front Door, when all health probes fail, all backends are considered healthy. Yes, you read that right.

Although we could create a route called /health that returns a 200, we will use the following probe just to make it work:

Fixing the health probe (quick and dirty fix); just can the /users API

If you are exposing multiple APIs on each cluster, the health probe above would not make sense. Also note that the purpose of the health probe is to determine if the cluster is up or not. It will not fix one API behaving badly or being removed accidentally!

You can check the health probes from the portal:

Yep! Health probes in West and North Europe are at 100%

Connection test

When I connect from my home laptop in Belgium, I get the following response:

Connection to West Europe cluster

When I connect from my second home in Dublin 🤷‍♀️ I get:

Connection from a VM in North Europe (I was kidding about the second home)

If you enable logging to Log Analytics, you can check this in the FrontDoorAccessLog:

Connection from home via Brussels (West Europe would show DB in Tenant_x)

When I remove the /users API in West Europe (kubectl delete deploy func), my home laptop will connect to North Europe as expected:

I didn’t fake this! It’s 100% real, my laptop now connects to the North Europe cluster via Front Door as expected

Note that the calls will not fail from the moment you delete the /users API (the health probe here). That depends on the following setting (backends in Front Door designer):

When should the backend be determined healthy or unhealthy; decrease sample size and or samples to make it go faster

The backend health percentage graph indicates the probe failure as well:

We’ve lost West Europe folks!

Conclusion

When you are going for a multi-region deployment of services, Azure Front Door is one of the options. Of course, there is much more to a multi-region deployment than the “front-end stuff” described in this post. What do you do with databases for instance? Can you use active-active write regions (e.g. Cosmos DB) or does your database only support active/passive with read replicas?

As in other load balancing and fail over solutions, proper health probes are crucial in the design. Think about what a good health probe can be and what it means when it is not available. One option is to just write a health probe exposed via an endpoint such as /health that merely returns a 200 status code. But your health probe could also be designed to connect to backend systems such as databases or queues to determine the health of the system.

Hopefully, this post gives you some ideas to start! Follow me on Twitter for updates.