Canary Deployments, A/B Testing, and Microservices With Ambassador

Canary deployments are a popular technique for incrementally testing changes on real-world traffic. In a traditional application, canary deployments occur on the granularity of the entire application. This limits the utility of canary deployments, as a single feature cannot be tested against real-world traffic.

With a microservices architecture, this is no longer the case. A single service team is able to test their updates with real-world users.

Unlike a monolith, a microservices team is able to:

• Test multiple versions of their service simultaneously.
• Control which versions of their service are being tested and when.

With these capabilities, a service team is able to not only conduct canary deployments, but run multiple different versions of their service for A/B testing, validating specific fixes, and so forth. The key to enabling these use cases is a Layer 7 reverse proxy.

Reverse Proxies and Layer 7

A reverse proxy capable of managing Layer 7 traffic is necessary to support these capabilities. Incoming traffic is routed through the reverse proxy, which then routes traffic to different versions of a service.

In this example, we'll use the Envoy Proxy. Envoy is a very lightweight, high-performance proxy designed for modern microservices architectures. Configuring Envoy, however, is a complicated exercise given the breadth of its features. Thus we'll use Ambassador, which is an open source API Gateway built around Envoy. Ambassador integrates Kubernetes-native functionality (e.g., annotations) into Envoy.

The picture below illustrates how an end user sends traffic to an L7 reverse proxy, which can then route that traffic to different versions of the service depending on criteria. For example, the proxy could route between a stable 1.1 version and a canary 1.2 version based on a weighted round robin policy, while sending only authenticated developer requests to the development 1.3 version.

Deploying a Canary

Ambassador is configured using Kubernetes annotations. We've already seen this dynamic configuration at work when we deployed a service that dynamically registered a URL with Ambassador. Now, we'll use that same workflow to configure Ambassador to route different percentages of traffic to two different versions of the service.

1. We'll start by deploying the same example service as we've done previously. This deploys a hello-world-stable service. Note: if you'd prefer to use Java, clone the Spring Boot Java template instead of the Python template.

   git clone https://github.com/datawire/hello-world-python
   cd hello-world-python
   forge deploy

2. Now, let's make a change to the app.py in the source tree. (If you're not familiar with Python, be aware that whitespace matters.)

   def root():
     return "Hello World Canary! (up %s)\n" % elapsed()

3. We're going to deploy this change as a canary. We do this by specifying using the canary profile defined in the service.yaml. This profile contains the parameter weight: 10.0, which tells Ambassador that 10% of the overall traffic should be routed to this deployment.

   forge --profile canary deploy

Testing the Canary

1. You'll see a new service, hello-world-canary be deployed. Now, let's test this out. Get the external IP address of Ambassador:

   kubectl get services ambassador
   NAME         CLUSTER-IP      EXTERNAL-IP      PORT(S)        AGE
   ambassador   10.11.250.208   35.190.189.139   80:31622/TCP   4d

2. In order to see the results of this canary, we need to simulate some real traffic. We'll run a loop to issue a stream of GET requests to the service, and see how Ambassador routes between the stableand canary versions.

   while true; do curl http://$AMBASSADOR_URL/hello/; done

Rollback

Sometimes your canary isn't working, and you want to roll back so that all your production traffic goes to the stable version. Rolling back is as simple as deleting the canary deployment.

1. Let's delete the canary deployment and service.

   kubectl delete svc,deploy hello-world-canary

2. If we rerun our loop, we'll see that traffic goes straight to the stable version (you may see the canary service still responding as the pod terminates).

   while true; do curl http://$AMBASSADOR_URL/hello/; done