This tutorial will walk through how to perform a canary deployment on Kubernetes. When updating an application, a slow rollout will allow for monitoring of traffic to the new deployment and allow a rollback of the deployment with minimal impact to operations.
A Service is an abstraction of the concept of pods. since pods can be created and destroyed, a service provides a single host:port to talk to the pods configured for this service. Services are created with a set of label selectors, and all requests made to the service will be balanced to pods running with the configured label selectors.
There is no restriction that the pods behind a service need to come from the same deployment, and this will allow us to create a canary deployment to test an update before performing a full update.
The applications we'll use to show the canary deployment are located in the app
folder. The application
uses an interface to abstract out the implementation of the AppInfo service.
The application is supposed to provide information about how the application has been deployed on Kubernetes. It looks for environment variables containing metadata about the application and labels on the container and provides them back as a json message.
package models
type AppInfo struct {
PodName string //Extracts the MY_POD_NAME environment variable
AppName string //Extracts the value for the app label
Namespace string //Extracts the MY_POD_NAMESPACE environment variable
Release string //Extracts the value for the release label
Labels map[string]string //Contains other labels for app
}
The app/models
folder contains the data transfer objects for the application.
The app/service
folder contains the interface definition and the implemenattions of the interface that
extract the AppInfo from the runtime. This folder contains three implementations of the interface.
The app/transport
folder contains the go-kit code to host the
interface implementation as a service.
The http.Handler
returned from MakeInfoServiceHandler
is hosted to translate a request to /v1/appinfo
to
the GetAppInfo
function of the interface implementation and serialize the response back.
There are three dockerfiles in the app
folder that correspond to images built with different implementations
of the interface. The following table shows which structs are being used in which Dockerfile, what tag these
images are pushed up with, and a brief description of what should be expected when running the application.
Struct | Dockerfile | Docker Image | Description |
---|---|---|---|
appInfoBaseline | app/Dockerfile.v1 | runyonsolutions/appinfo:1 |
Does not return Namespace value. |
appInfoBroken | app/Dockerfile.v2 | runyonsolutions/appinfo:2 |
Returns an error |
appInfoWithNamespace | app/Dockerfile.v3 | runyonsolutions/appinfo:3 |
populates Namespace value correctly |
In practice, the application being deployed would probably not have three separate implementations of the same interface, but might represent iterations made to one implementation. The use of three different structs simplifies showing the code running in each image.
The application layout does show how easily it would be to change how a service functions and interchange different storage systems for basic CRUD applications, or different implementations of analytics.
As a baseline, we assume there is a Kubernetes deployment named appinfo that was deployed from the
deployment/appinfo.yaml
file. This deployment can be created via
kubectl create -f https://raw.githubusercontent.com/runyontr/k8s-canary/master/deployment/appinfo.yaml
The deployment is very basic except for two additional configuration. First is Exposing Pod Information through Environment Variables. These are created from the yaml in Deployment YAML
env:
- name: MY_POD_NAME
valueFrom:
fieldRef:
fieldPath: metadata.name
- name: MY_POD_NAMESPACE
valueFrom:
fieldRef:
fieldPath: metadata.namespace
Which create an environment variable MY_POD_NAME
that contains the pod name and MY_POD_NAMESPACE
which contains the namespace where the pod is running.
The second customization is Exposing pod labels into a container.
which creates a file at /etc/labels
containing the pods labels.
Now get the name of the pod that was deployed:
kubectl get pods -l app=appinfo
output:
NAME READY STATUS RESTARTS AGE
appinfo-567b989978-vvzlg 1/1 Running 0 19s
Now we can port forward localhost:8080 to the pod's container port 8080 via
POD_NAME=$(kubectl get pods -l app=appinfo -o jsonpath="{.items[0].metadata.name}")
kubectl port-forward $POD_NAME 8080:8080
output
Forwarding from 127.0.0.1:8080 -> 80
Forwarding from [::1]:8080 -> 80
Open a new terminal and we can send an HTTP request:
curl -s localhost:8080/v1/appinfo | jq .
and get a response like
{
"PodName": "appinfo-567b989978-vvzlg",
"AppName": "appinfo",
"Namespace": "",
"Release": "stable",
"Labels": {
"pod-template-hash": "1236545534"
}
}
In a new terminal run this command to be constantly refreshing the AppInfo, which will be useful to show the real time updates in the runtime in the following section.
while true; do clear; curl -s localhost:8080/v1/appinfo | jq . ; sleep 1; done;
To see how the /etc/labels
file is updated dynamically, we can add a new label to the pod and see
the output of our while loop get adjusted in real time
kubectl label pods $POD_NAME newlabel=realtime
Switch back to the first terminal to stop the port forwarding with CTRL+C.
A service provides load balancing to a selection of pods based on a particular label.
The label we're going
to filter by is app=appinfo
. To see the pods that satisfy this, run
kubectl get pods -l app=appinfo
The service defined in deployment/appinfo-service.yaml
has the label selector defined as app=appinfo
.
This will create
a load balancer (service) that routes requests to pods with the label app=appinfo
.
kubectl create -f https://raw.githubusercontent.com/runyontr/k8s-canary/master/deployment/appinfo-service.yaml
and see it
kubectl get svc appinfo
output:
NAME CLUSTER-IP EXTERNAL-IP PORT(S) AGE
appinfo 10.32.0.232 <nodes> 8080:30753/TCP 14s
There will now be an environment variables in all running containers that start after the creation of this service:
APPINFO_SERVICE_PORT=8080
APPINFO_PORT_8080_TCP_PORT=8080
APPINFO_PORT=tcp:https://10.32.0.232:8080
APPINFO_PORT_8080_TCP_ADDR=10.32.0.232
which will allow all pods running in the cluster to have the connection information for this service readily available.
The type of service we created was NodePort.
When this type of service is created, each node will proxy requests to a specific port on the node (the same port on each node) to the configured port on the pod (8080 for our application).
To obtain the port on the node, run
NODE_PORT=$(kubectl get svc appinfo \
--output=jsonpath='{range .spec.ports[0]}{.nodePort}')
When running on a cloud provider, a firewall rule might need to be added to allow TCP traffic into the nodes at the service port. To enable traffic in Google Cloud, the following command will open traffic to the node port.
gcloud compute firewall-rules create appinfo-service \
--allow=tcp:${NODE_PORT} \
--network ${KUBERNETES_NETWORK_NAME}
where ${KUBERNETES_NETWORK_NAME}
is the name of the network Kubernetes is deployed on.
Now if ${EXTERNAL_IP}
is the public address of one of the nodes in the cluster, then
the service is available at https://${EXTERNAL_IP}:${NODE_PORT}
Similar to the while loop monitoring the port forwarded traffic of a particular pod in Talking to the Deployment, the loop
while true; do clear; curl -s ${EXTERNAL_IP}:${NODE_PORT}/v1/appinfo | jq . ; sleep 1; done;
will show the output of the service. In order to monitor the deployments in the following section, this command should be run in a new terminal.
Looking at the output of the loop, we see that the namespace field is not being populated correctly by the
deployment. One proposed (failed) solution is captured in the implementation appInfoBroken
, which
simulates a developers failed attempt at fixing the issue. As new software can sometimes contain bugs, a slow
rollout of the new version of the software will allow for minimal impact if there is an issue.
To simulate a real deployment, we should have the currently deployed application scaled to handle the current traffic. When looking at how large to scale the current deployment (n), its helpful to understand the system's SLOs. The new canary deployment will be getting 1/(n+1) of the traffic going to the service, and reducing the canary's load by increasing the value of n lower the impact to the error budget when things go wrong.
For this tutorial, we scale to 3 replicas.
kubectl scale --replicas=3 deployment appinfo
deployment "appinfo" scaled
kubectl get pods -l app=appinfo
NAME READY STATUS RESTARTS AGE
appinfo-567b989978-6hbm8 1/1 Running 0 15s
appinfo-567b989978-q58nt 1/1 Running 0 15s
appinfo-567b989978-vvzlg 1/1 Running 0 14m
Looking at the output of the while loop look should now show the responses coming from pods with different names. Additionally, if any labels were applied to the first pod in Chaning Labels, the newly created pods will not have those labels. This should be seen as responses will have different sets of labels.
Now we are ready to deploy our broken canary deployment:
kubectl create -f https://raw.githubusercontent.com/runyontr/k8s-canary/master/deployment/canary-broken.yaml
deployment "appinfo-canary-broken" created
Looking at the running pods, we can now see 4 pods with the app=appinfo
label:
kubectl get pods -l app=appinfo
NAME READY STATUS RESTARTS AGE
appinfo-567b989978-6hbm8 1/1 Running 0 10m
appinfo-567b989978-q58nt 1/1 Running 0 10m
appinfo-567b989978-vvzlg 1/1 Running 0 25m
appinfo-canary-broken-c66665c44-7cfk6 1/1 Running 0 19s
Now looking at the loop in Connection should have about 1/4 of the response coming back with a message of
{"error": "something went wrong"}
and the other 3/4 of response should be responding as before.
This is when application monitoring (e.g. Prometheus/Grafana) would be able to split the metrics between canary and stable pods and show any difference in performance. A future iteration of this tutorial will demonstrate the performance differences with a monitoring solution.
For this demo, the while loop will be used to demonstrate the health of the system, and since there are errors being returned in some responses, a rollback of the deployment is required:
kubectl delete -f https://raw.githubusercontent.com/runyontr/k8s-canary/master/deployment/canary-broken.yaml
At this point, the broken pod should be terminating:
kubectl get pods -l app=appinfo
NAME READY STATUS RESTARTS AGE
appinfo-567b989978-6hbm8 1/1 Running 0 12m
appinfo-567b989978-q58nt 1/1 Running 0 12m
appinfo-567b989978-vvzlg 1/1 Running 0 26m
appinfo-canary-broken-c66665c44-7cfk6 1/1 Terminating 0 2m
and responses will return to being valid 100% of the time. In most shops, this would be enough to justify a Post Mortem and improve any testing process prior to being considered for a release.
After figuring out the issue, a new fix has been created and is ready to be rolled out. Following a similar process, a new canary deployment is created:
kubectl create -f https://raw.githubusercontent.com/runyontr/k8s-canary/master/deployment/canary-fixed.yaml
deployment "appinfo-canary-fixed" created
Looking at the output of the while loop now shows about 1/4 of the requests have the correct namespace value (default
), and
are reporting the release=canary
label.
At the point the team is willing to accept the new version formally, the configuration on the stable app would need to be updated to the docker image of the canary deployment:
kubectl set image deployment/appinfo appinfo-containers=runyonsolutions/appinfo:3
Since the image is being update this should fire off a rolling update of the deployment and new pods should be created
kubectl get pods -l app=appinfo
NAME READY STATUS RESTARTS AGE
appinfo-567b989978-6hbm8 1/1 Terminating 0 16m
appinfo-567b989978-q58nt 1/1 Terminating 0 16m
appinfo-567b989978-vvzlg 1/1 Running 0 30m
appinfo-84d5cf794d-9dpfz 1/1 Running 0 5s
appinfo-84d5cf794d-dd67v 1/1 Running 0 10s
appinfo-84d5cf794d-lq45c 0/1 ContainerCreating 0 2s
appinfo-95bccb844-jnfzn 0/1 Terminating 3 1m
appinfo-canary-fixed-744f96dc75-zbxr9 1/1 Running 0 2m
Describing any of the newly created pods should show the update image. The monitoring loop in Connection
should should show labels release=stable
having the namespace value correctly set.
Finally, we need to clean up the canary app.
kubectl delete -f https://raw.githubusercontent.com/runyontr/k8s-canary/master/deployment/canary-fixed.yaml
Now all pods running behind the service are updated.