Deployments using Traffic Splitting

Overview

You can use traffic splitting for most deployment scenarios, including canary, blue-green, A/B testing, and so on. The ability to control traffic flow to different versions of an application makes it easy to roll out a new application version with minimal effort and interruption to production traffic.

Before You Begin

  1. Install kubectl.

  2. (Optional) If you want to view metrics, ensure that you have deployed Prometheus and Grafana. Refer to the Monitoring and Tracing guide for instructions.

  3. Set up a Kubernetes cluster with F5 NGINX Service Mesh deployed with the following configuration:

    • --mtls-mode is set to permissive or off states.
    • (Optional) --prometheus-address is pointed to the Prometheus instance you created above.

  4. Enable automatic sidecar injection for the default namespace.

  5. Download all the example files:

The NGINX Plus Ingress Controller’s custom resource TransportServer has the same Kubernetes short name(ts) as the custom resource TrafficSplit. If you have the NGINX Plus Ingress Controller installed, use the full name trafficsplit(s) instead of ts in the following instructions.
Avoid configuring traffic policies such as TrafficSplits, RateLimits, and CircuitBreakers for headless services. These policies will not work as expected because NGINX Service Mesh has no way to tie each pod IP address to its headless service.

Objectives

Follow the steps in this guide to learn how to use traffic splitting for various deployment strategies.

Deploy the Production Version of the Target App

  1. First, let’s begin by deploying the “production” v1.0 target app, the load balancer Service, and the ingress gateway.

    For simplicity, this guide uses a simple NGINX reverse proxy for the ingress gateway. For production usage and for more advanced ingress control, we recommend using the NGINX Ingress Controller for Kubernetes. Refer to Deploy NGINX Ingress Controller with NGINX Service Mesh to learn more.

    Command:

    kubectl apply -f target-svc.yaml -f target-v1.0.yaml -f gateway.yaml

    Expectation: All Pods and Services deploy successfully.

    Use kubectl to make sure the Pods and Services deploy successfully.

    Example:

    bash
    $ kubectl get pods
NAME                           READY   STATUS    RESTARTS   AGE
gateway-58c6c76dd-4mmht        2/2     Running   0          2m
target-v1-0-6f69fc48f6-mzcf2   2/2     Running   0          2m
    bash
    $ kubectl get svc
NAME          TYPE           CLUSTER-IP      EXTERNAL-IP     PORT(S)        AGE
gateway-svc   LoadBalancer   10.0.0.2        1.2.3.4         80:30975/TCP   2m
target-svc    ClusterIP      10.0.0.3        <none>          80/TCP         2m
target-v1-0   ClusterIP      10.0.0.4        <none>          80/TCP         2m

    To better understand what is going on here, let’s take a quick look at what we deployed here:

    • gateway: simple NGINX reverse proxy that forwards traffic to the target app. Besides providing a single point of ingress to the cluster, using the gateway lets us use the nginx-meshctl top command to check traffic metrics between it and the backend Services it is sending traffic to.
    • target-svc: the root Service that connects to all the different versions of the target app.
    • target: for our example we will be deploying 3 different versions of the target app. The target app is a basic NGINX server that returns the target version. Each one has its own Service tagged with its version number. These are the Services that the root target-svc sends requests to.

  2. Once the Pods and Services are ready, generate traffic to target-svc. Use a different bash window for this step so you can watch the traffic change as you are doing the deployments.

    Commands:

    • Get the external IP for gateway-svc:

      kubectl get svc gateway-svc

    • Save the IP address as an environment variable:

      export GATEWAY_IP=<gateway external IP>

    • Start a loop that sends a request to that IP once per second for 5 minutes. Rerun as needed:

      for i in $(seq 1 300); do curl $GATEWAY_IP; sleep 1; done

    Expectation: Requests will start to come in to target-svc. At this point you should only see target v1.0 responses.

  3. Back in your original bash window, use the mesh CLI to check traffic metrics.

    Command: nginx-meshctl top
    Expectation: The target-v1-0 deployment will show 100% incoming success rate and the gateway deployment will show 100% outgoing success rate. The top command only shows traffic from the last 30s. top provides a quick look at your Services for immediate debugging and to see if there’s any anomalies that need further investigation. For more detailed and accurate traffic monitoring, we recommend using Grafana. Refer to traffic metrics for details.

    Example:

    bash
    $ nginx-meshctl top
Deployment   Incoming Success  Outgoing Success  NumRequests
gateway                        100.00%           10
target-v1-0  100.00%                             10

Deploy a New Version of the Target App using a Canary Deployment

Using traffic splits we can use a variety of deployment strategies. Whether using a blue-green deployment, canary deployment, or a hybrid of different deployment strategies, traffic splits make the process extremely easy.

For this version of the target app, let’s try using a canary deployment strategy.

  1. Apply the traffic split so that once a new version is deployed, it will not receive any traffic until we are ready. Ideally we would apply this at the same time as the first target version, target-svc, and gateway. To make it easier to see what is happening though, we are applying it in this separate step.

    Command:

    kubectl apply -f trafficsplit.yaml

    Expectation: The traffic split is applied successfully. Use kubectl get ts to see the current traffic splits.

    Use kubectl describe ts target-ts to see details about the traffic split we just applied. Currently the traffic split is configured to send 100% of traffic to target v1.0.

    yaml
    apiVersion: split.smi-spec.io/v1alpha3
kind: TrafficSplit
metadata:
  name: target-ts
spec:
  service: target-svc
  backends:
  - service: target-v1-0
    weight: 100

  2. Now let’s deploy target v2.0. To show a scenario where an upgrade is failing, this version of target is configured to return a 500 error status code instead of a successful 200.

    Command:

    kubectl apply -f target-v2.0-failing.yaml

    Expectation: Target v2.0 will deploy to the cluster successfully. You should see the new target-v2-0 Pod and Service in the kubectl get pods/kubectl get svc output. Since we deployed the traffic split, if you look at your other bash window where the traffic is being generated you should still only see responses from target v1.0. If you check nginx-meshctl top you should see the same deployments as before. This is because no traffic has been sent to or received from target v2.0.

  3. For this deployment we’ll send 10% of traffic to target v2.0 while 90% is still going to target v1.0. Open trafficsplit.yaml in the editor of your choice and add a new backend for target-v2-0 with a weight of 10. Change the weight of target-v1-0 to 90.

    yaml
    apiVersion: split.smi-spec.io/v1alpha3
kind: TrafficSplit
metadata:
  name: target-ts
spec:
  service: target-svc
  backends:
  - service: target-v1-0
    weight: 90
  - service: target-v2-0
    weight: 10

  4. After updating trafficsplit.yaml, save and apply it.

    Command:

    kubectl apply -f trafficsplit.yaml

    Expectation: After applying the updated traffic split, you should start seeing responses from target v2.0 in the other bash where traffic is being generated. Because of the weight we set in the previous step, about 1 out of 10 requests will be sent to v2.0. Something to keep in mind is that these are weighted, so it will not be exactly 1 in 10, but it will be close.

  5. Check the traffic metrics now that v2.0 is available.

    Command:

    nginx-meshctl top

    Expectation:

    • target-v1-0 deployment will still show 100% incoming success rate
    • target-v2-0 deployment will show 0% incoming success rate
    • gateway deployment will show the appropriate percentage of successful outgoing requests

    Example:

    bash
    $ nginx-meshctl top
Deployment   Incoming Success  Outgoing Success  NumRequests
gateway                        90.00%            10
target-v1-0  100.00%                             9
target-v2-0  0.00%                               1

  6. It looks like v2.0 doesn’t work! We can see that because the incoming success rate to target-v2 is 0%. Thankfully, using traffic splitting, it is easy to redirect all traffic back to v1.0 without doing a complicated rollback. To update the traffic split, simply update trafficsplit.yaml to send 100% of traffic to v1.0 and 0% of traffic to v2.0 and re-apply it.

    You can either explicitly set the weight of target-v2-0 to 0 or remove the target-v2-0 backend completely. The result will be the same.

    At this point you can delete v2.0 from the cluster.

    Command:

    kubectl delete -f target-v2.0-failing.yaml

Deploy a New Version of the Target App using a Blue-Green Deployment

For this version of the target app, let’s use a blue-green deployment.

  1. Deploy v2.1 of target, which fixes the issue causing the failing requests that we saw in v2.0.

    Command:

    kubectl apply -f target-v2.1-successful.yaml

    Expectation: Target v2.1 will deploy successfully. You should see the new target-v2-1 Pod and Service in the kubectl get pods/kubectl get svc output. Just as with target-v2-0 though, we have the traffic split configured to send all traffic to target-v1-0 until we are ready to do the actual deployment and make target-v2-1 available for traffic.

  2. Since we are doing a blue-green deployment, we will configure the traffic split to send all traffic to target v2.1. Open trafficsplit.yaml in the editor of your choice and add a new backend for target-v2-1 with a weight of 100. Change the weight of target-v1-0 to 0. You could also delete the target-v1-0 backend completely, but with this type of deployment it’s easier to set the weight to 0 in case you need to roll back quickly.

    yaml
    apiVersion: split.smi-spec.io/v1alpha3
kind: TrafficSplit
metadata:
  name: target-ts
spec:
  service: target-svc
  backends:
  - service: target-v1-0
    weight: 0
  - service: target-v2-1
    weight: 100

  3. After updating trafficsplit.yaml, save and apply it.

    Command:

    kubectl apply -f trafficsplit.yaml

    Expectation: After applying the updated traffic split, you should start seeing responses from target v2.1 in the other bash where traffic is being generated. Because of the weight we set in the previous step, all traffic should be going to v2.1.

  4. Check the traffic metrics now that v2.1 is available.

    Command:

    nginx-meshctl top

    Expectation:

    • target-v1-0 deployment will not show up, although keep in mind that it will take a bit for the previous requests to move out of the 30s metric window. If you see target-v1-0, try again in 30s or so.
    • target-v2-1 deployment will show 100% incoming success rate
    • gateway deployment will show 100% outgoing success rate

    Example:

    bash
    $ nginx-meshctl top
Deployment   Incoming Success  Outgoing Success  NumRequests
gateway                        100.00%           10
target-v2-1  100.00%                             10

  5. Since target v2.1 is working as expected, we can delete v1.0 from the cluster. If v2.1 had started failing, we could have quickly rolled back to v1.0 just as we did earlier.

    Command:

    kubectl delete -f target-v1.0.yaml

A/B Testing with Traffic Splits

If you want to implement A/B testing, you can create an HTTPRouteGroup resource and associate the HTTPRouteGroup with the traffic split.

Consider the following configuration:

yaml
apiVersion: specs.smi-spec.io/v1alpha3
kind: HTTPRouteGroup
metadata:
  name: target-hrg
  namespace: default
spec:
  matches:
  - name: firefox-users
    headers:
      user-agent: ".*Firefox.*"

The HTTPRouteGroup is used to describe HTTP traffic. The spec.matches field defines a list of routes that an application can serve. Routes are made up of the following match conditions: pathRegex, headers, and HTTP methods. In the target-hrg example above, we have defined one route, firefox-users, using the header filter user-agent: ".*Firefox.*". Incoming HTTP traffic that has the user-agent header set to a value that matches the regex ".*Firefox.*" satisfies the firefox-users match condition.

A route with multiple match conditions (pathRegex, headers, and/or HTTP methods) within a single match represent an AND condition. This means that all match conditions must be satisfied for the traffic to match the route.

To associate the target-hrg HTTPRouteGroup with the traffic split we need to add the matches field to our traffic split spec:

yaml
apiVersion: split.smi-spec.io/v1alpha3
kind: TrafficSplit
metadata:
  name: target-ts
spec:
  service: target-svc
  backends:
  - service: target-v2-1
    weight: 0
  - service: target-v3-0
    weight: 100
  matches:
  - kind: HTTPRouteGroup
    name: target-hrg

Traffic split matches allow you to associate one or more HTTPRouteGroups with a traffic split. The matches field in the traffic split spec maps the HTTPRouteGroup’s matches directives to the traffic split. This means that the traffic split only applies if the request’s attributes satisfy the match conditions outlined in the match directives.

If there are multiple HTTPRouteGroups listed in the traffic split matches field and/or multiple matches defined in the HTTPRouteGroup, the traffic split will be applied if the request satisfies any of the specified matches.

In this example, all traffic sent to the root Service target-svc that contains the string Firefox in the user-agent header will be routed to the target-v2-1 backend. All other traffic will be sent to the root Service target-svc.

  1. To demonstrate how to A/B test with traffic splits, let’s create a new version of the target application:

    Command:

    kubectl apply -f target-v3.0.yaml

    Expectation: The target-v3-0 Pod and Service deploy successfully. At this point there should be three Pods and four Services running.

    Example:

    bash
    $ kubectl get pods
NAME                            READY   STATUS    RESTARTS   AGE
gateway-58c6c76dd-4mmht         2/2     Running   0          2m
target-v2-1-6f69fc48f6-mzcf2    2/2     Running   0          2m
target-v3-0-5f6fc9cf99-tps6k    2/2     Running   0          2m
    bash
    $ kubectl get svc
NAME          TYPE           CLUSTER-IP      EXTERNAL-IP     PORT(S)        AGE
gateway-svc   LoadBalancer   10.0.0.2        1.2.3.4         80:30975/TCP   2m
target-svc    ClusterIP      10.0.0.3        <none>          80/TCP         2m
target-v2-1   ClusterIP      10.0.0.4        <none>          80/TCP         2m
target-v3-0   ClusterIP      10.0.0.5        <none>          80/TCP         2m

    In the terminal window where you are generating traffic to the gateway Service you should still see all responses coming from the target-v2-1 backend. You may need to restart the loop that generates traffic:

    for i in $(seq 1 300); do curl $GATEWAY_IP; sleep 1; done

  2. Create the target-hrg HTTPRouteGroup and update the traffic split:

    Command:

    kubectl apply -f trafficsplit-matches.yaml

    Expectation: The target-hrg HTTPRouteGroup is created and the target-ts traffic split is updated.

    Use kubectl get and kubectl describe for httproutegroups and trafficsplits to make sure the resources were created or updated.

    In the terminal window where you are generating traffic to the gateway Service you should now see responses from both the target-v2-1 and target-v3 backends. To test the A/B traffic shaping, open another terminal window and generate traffic to the gateway Service with the header user-agent: Firefox:

    for i in $(seq 1 100); do curl $GATEWAY_IP -H "user-agent:Firefox"; sleep 1; done

    Since the user-agent header is set to “Firefox”, you should see responses from the target-v3-0 backend only.

Traffic Splitting based on path and HTTP methods

In addition to supporting traffic splitting based on header filters, NGINX Service Mesh also supports traffic splitting based on path and HTTP methods. To demonstrate this let’s update the target-hrg and add a new match.

  1. Edit the HTTPRouteGroup target-hrg:

    Command:

    kubectl edit httproutegroup target-hrg

    Add the get-api-requests route to the list of matches:

    yaml
    apiVersion: specs.smi-spec.io/v1alpha3
kind: HTTPRouteGroup
metadata:
  name: target-hrg
  namespace: default
spec:
  matches:
  - name: firefox-users
    headers:
      user-agent: ".*Firefox.*"
  - name: get-api-requests
    pathRegex: "/api"
    methods:
    - GET

    Save and close the editor.

    The get-api-requests route will match all GET requests to the /api endpoint. By adding this route to the target-hrg matches, the target-ts traffic split will now have both the firefox-users and get-api-requests matches applied to it. Since multiple matches are applied with an OR operator, if an incoming HTTP request to target-svc matches either firefox-users or get-api-requests, the traffic split will be applied, and the request will be routed to the target-v3-0 backend Service. All other incoming HTTP requests will be routed to the root target-svc, which will forward the request to one of the target services based on the load-balancing algorithm of the mesh.

    Expectations:

    • In the terminal window where requests to the gateway Service have the user-agent:Firefox header set, you should still see responses from the target-v3-0 backend only.

    • To test the get-api-requests route, start a new for loop that sends GET requests to the /api endpoint:

      for i in $(seq 1 100); do curl $GATEWAY_IP/api; sleep 1; done

      You should only see responses from the target-v3-0 backend. If you remove the /api path from the request you should see responses from both the target-v2-1 and target-v3-0 backends.

Cleanup

Delete all the resources from your cluster:

Command:

kubectl delete -f gateway.yaml -f target-svc.yaml -f target-v2.1-successful.yaml -f target-v3.0.yaml -f trafficsplit-matches.yaml

Use Case

An example use case for traffic splitting can be seen in this blog. The blog uses a self-referential TrafficSplit configuration in which the root service is also listed as a backend service. Even though the Service Mesh Interface specification mentions that self-referential configurations are invalid, NGINX Service Mesh supports this type of configuration due to its value in use cases like the example defined in the blog.

Summary

These are just a couple examples of how you can use traffic splits for a deployment. Whether you want to do a gradual roll out of 5% increments or send 5% to two staging backends while 90% goes to production or any other combination of splits, traffic splits offer a convenient way to handle almost any deployment strategy you need.

Resources

View source
Edit this page
Create a new issue