Spinnaker is a multi-cloud, multi-region automated deployment tool. Open sourced by Netflix and heavily contributed to by Google, it supports all major cloud providers including Kubernetes.
Last month, Kayenta was open sourced, a canary analysis engine. Canary analysis is a technique to reduce the risk from deploying a new version of software into production. A new version of the software, referred to as the canary, is deployed to a small subset of users alongside the stable running version. Traffic is split between these two versions such that a portion of incoming requests is diverted to the canary. This approach can quickly uncover any problems with the new version without impacting the majority of users.
The quality of the canary version is assessed by comparing key metrics that describe the behavior of the old and new versions. If there is a significant degradation in these metrics, the canary is aborted and all of the traffic is routed to the stable version in an effort to minimize the impact of unexpected behavior.
Ordina helps companies through digital transformation using three main focus areas:
Embracing a DevOps culture and corresponding practices allows teams to focus on delivering value for the business, by changing the communication structures of the organization. Through automation, teams are empowered and capable of delivering applications much faster to production.
Having a modular decoupled architecture, our second focus area, fits well with this model. Making these changes to our architecture in combination with a culture of automation, results in a lot more moving parts in our application landscape.
Naturally, the next step is tackling the underlying infrastructure accomodate this new architecture and way of working. Cloud automation is therefore our final focus area in digital transformations.
Releasing more often doesn’t only allow new features reaching the user faster, it also fastens the feedback loops, improves reliability and availability, developer productivity and efficiency. Spinnaker plays a crucial part in all of this, as it allows more frequent and faster deployments, without sacrificing safety.
Automated canary analysis, demonstrated in this codelab, is a powerful tool in that sense.
- Introducing our Rick & Morty demo
- Running the demo scenario
The purpose of this codelab is to simplify getting up-and-running with automated canary analysis using Spinnaker on Kubernetes.
We’re using Google Cloud Platform for this demonstration. Monitoring and logging will be handled by Stackdriver, which is integrated completely with GCP.
The canary functionality we’re going to use in this setup requires the use of a specific cluster version with full rights:
- You must be an Owner of the project containing your cluster.
- You must use Kubernetes v1.10.2-gke.0 or later.
Introducing our Rick & Morty demo
Rick & Morty is a television show following the misadventures of cynical mad scientist Rick Sanchez and his good-hearted but fretful grandson Morty Smith, who split their time between domestic life and interdimensional adventures.
Our demo application is a Java Spring Boot application, running on an Apache Tomcat server, packaged inside a docker container.
The docker container runs on Kubernetes managed by Google Cloud Platform (GKE).
The application exposes an endpoint on
http://localhost:8080 and can be run locally by executing
./mvnw spring-boot:run, assuming you have a JRE or JDK (v8+) installed.
The endpoint returns an HTML with a background of Pickle Rick. In season three episode three, Rick turns himself into a pickle in an attempt at escaping family therapy.
Pickle Rick will act as our initial green deployment, running stabily on production.
We will try to replace it with a blue deployment.
Mr. Meeseeks, featured in season one episode five, will be the protagonist of that deployment. Meeseeks are creatures who are created to serve a singular purpose for which they will go to any length to fulfill. After they serve their purpose, they expire and vanish into the air. Their motivation to help others comes from the fact that existence is painful to a Meeseeks, and the only way to be removed from existence is to complete the task they were called to perform.
Meeseeks can however summon other Meeseeks to help, which could spiral out of control if the task at hand is unsolvable.
Therefore, Meeseeks are quite dangerous and a good candidate for our misbehaving blue deployment.
Aside from the HTTP endpoint, our demo application also prints out a number of character names from the series.
Blue Green Differences
Aside from the leading character in our two versions, there are two specific differences in the code between both versions.
The following commit shows moving from the green version to the blue version: 24cc45cf
First of all, the background image changes, which gives a clear visual indication of which version is currently deployed.
Since using Meeseeks could get out of hand quickly, keeping track of how many times the Meeseeks HTTP endpoint has been hit makes a lot of sense. Hence, the blue version prints an extra
Meeseeks in the logs, for every request to the endpoint.
Using this setup, we should be able to consider logs as a source of information for judging the canary healthiness.
Note that the Github repository can constantly switch between Pickle Rick and Meeseeks. Before starting a build and making deployments, make sure your fork is aligned with the green version. If this isn’t the case, switching to green is demonstrated in the following commit: 784e616a
Setup Continuous Integration
Making changes to our application will be the trigger for our pipelines. Therefore, we should have a simple continuous integration flow set up. We could use Jenkins or any other build server that uses webhooks, but since our entire demo is being deployed on GCP, we can use the build server from GCP instead.
First of all, fork the demo application repository.
In the GCP console, open build triggers underneath the Container Registry (GCR) tab.
Select Github as repository hosting, and select the forked repository to create a trigger for. Configure the trigger to activate on any branch, using a
cloudbuild.yaml file located in the root of the repository.
This will run a maven build and docker build, and push the created docker image into the GCR.
Throughout this guide we refer to the official documentation for individual parts of the installation already covered by the Spinnaker team. However, as reference we also compiled an exhaustive list of commands to execute based on the commands found in those articles. This means you could skip the official documentation and simply execute those commands. However, we still recommend going through the docs to get more context. The list of commands to execute can be found at the end of this chapter.
Follow the guide on Spinnaker’s website: https://www.spinnaker.io/setup/quickstart/halyard-gke
Create a cluster as mentioned here: https://cloud.google.com/monitoring/kubernetes-engine/installing
gcloud components update gcloud auth login gcloud config set project <PROJECT_NAME>
Find out the latest supported cluster version with the following command:
gcloud container get-server-config --zone=$ZONE
Create a cluster for your specific zone (e.g. europe-west1-d) and preferred cluster version (v1.10.2-gke.0 or later):
CLUSTER_VERSION=1.10.2-gke.1 GCP_PROJECT=$(gcloud info --format='value(config.project)') ZONE=europe-west1-d CLOUDSDK_CONTAINER_USE_V1_API=false CLOUDSDK_API_CLIENT_OVERRIDES_CONTAINER=v1beta1 gcloud beta container clusters create spinnaker \ --zone=$ZONE \ --project=$GCP_PROJECT \ --cluster-version=$CLUSTER_VERSION \ --enable-stackdriver-kubernetes \ --enable-legacy-authorization
--enable-legacy-authorization are passed.
Navigate to the Google Cloud Console and enable the following APIs:
This section complements official documentation with some recommendations and extras.
Postpone running the
hal deploy apply command until the end of this chapter.
During the Halyard on GKE guide on Spinnaker’s website, remember to use the right zone when creating the Halyard VM.
gcloud compute instances create $HALYARD_HOST \ --project=$GCP_PROJECT \ --zone=$ZONE \ --scopes=cloud-platform \ --service-account=$HALYARD_SA_EMAIL \ --image-project=ubuntu-os-cloud \ --image-family=ubuntu-1404-lts \ --machine-type=n1-standard-4
When SSH’ing into the Halyard VM, also remember to use the right zone.
gcloud compute ssh $HALYARD_HOST \ --project=$GCP_PROJECT \ --zone=$ZONE \ --ssh-flag="-L 9000:localhost:9000" \ --ssh-flag="-L 8084:localhost:8084"
Before you perform
hal deploy apply, add the Docker registry corresponding to your region. In case your project is located in Europe, add the eu.gcr.io registry as illustrated below.
hal config provider docker-registry account add gcr-eu \ --address eu.gcr.io \ --password-file ~/.gcp/gcp.json \ --username _json_key hal config provider kubernetes account edit my-k8s-account --docker-registries my-gcr-account gcr-eu
Enable Stackdriver access for Spinnaker in GCP’s IAM settings.
Add the following roles to the member with name
- Logging Admin
- Monitoring Admin
Automated Canary Analysis
Before you perform
hal deploy apply, enable automated canary analysis.
Follow the guide further down, but first of all, set some variables while still SSH’d in the Halyard VM.
One of these variables is the Spinnaker bucket automatically created when installing Halyard.
Look for the right bucket identifier in the GCP GKE buckets dashboard.
PROJECT_ID=$(gcloud info --format='value(config.project)') JSON_PATH=~/.gcp/gcp.json MY_SPINNAKER_BUCKET=spin-48b89b5e-dd67-446a-ad9f-66e8783e9822
Follow the official canary quickstart documentation.
Configure the default metrics store.
hal config canary edit --default-metrics-store stackdriver
And finally execute the rollout.
hal deploy apply
Sometimes an issue might occur with credentials on the Halyard VM:
! ERROR Unable to communicate with your Kubernetes cluster: Failure executing: GET at: https://220.127.116.11/api/v1/namespaces. Message: Forbidden! User gke_spinnaker-demo-184310_europe-west1-d_spinnaker-alpha doesn't have permission. namespaces is forbidden: User "client" cannot list namespaces at the cluster scope: Unknown user "client".. ? Unable to authenticate with your Kubernetes cluster. Try using kubectl to verify your credentials.
In this case, enable legacy authentication in the GKE UI for the cluster.
You can monitor deployment locally on your own PC by running
kubectl get pods -w --all-namespaces.
For this to work, kubectl needs permissions to talk to the cluster.
You can use gcloud to populate your kubeconfig file with credentials to access the cluster.
This can help you to look into specific logs of each Spinnaker pod or follow up on deployments handled by Spinnaker.
You can find out which commands are sent to GCP by enabling audit logging. Turn on audit logging: https://cloud.google.com/monitoring/audit-logging & https://cloud.google.com/logging/docs/audit/configure-data-access#example
Comprehensive list of commands
These are all the commands we have executed in order to get everything set up.
Fill in the
<PLACEHOLDER> placeholders according to your preferences.
ZONE=<ZONE_NAME> CLUSTER_VERSION=<CLUSTER_VERSION> GCP_PROJECT=<PROJECT_NAME> gcloud components update gcloud auth login gcloud config set project $PROJECT_NAME gcloud container get-server-config --zone=$ZONE CLOUDSDK_CONTAINER_USE_V1_API=false CLOUDSDK_API_CLIENT_OVERRIDES_CONTAINER=v1beta1 gcloud beta container clusters create spinnaker \ --zone=$ZONE \ --project=$GCP_PROJECT \ --cluster-version=$CLUSTER_VERSION \ --enable-stackdriver-kubernetes \ --enable-legacy-authorization HALYARD_SA=halyard-service-account gcloud iam service-accounts create $HALYARD_SA \ --project=$GCP_PROJECT \ --display-name $HALYARD_SA HALYARD_SA_EMAIL=$(gcloud iam service-accounts list \ --project=$GCP_PROJECT \ --filter="displayName:$HALYARD_SA" \ --format='value(email)') gcloud projects add-iam-policy-binding $GCP_PROJECT \ --role roles/iam.serviceAccountKeyAdmin \ --member serviceAccount:$HALYARD_SA_EMAIL gcloud projects add-iam-policy-binding $GCP_PROJECT \ --role roles/container.admin \ --member serviceAccount:$HALYARD_SA_EMAIL GCS_SA=gcs-service-account gcloud iam service-accounts create $GCS_SA \ --project=$GCP_PROJECT \ --display-name $GCS_SA GCS_SA_EMAIL=$(gcloud iam service-accounts list \ --project=$GCP_PROJECT \ --filter="displayName:$GCS_SA" \ --format='value(email)') gcloud projects add-iam-policy-binding $GCP_PROJECT \ --role roles/storage.admin \ --member serviceAccount:$GCS_SA_EMAIL gcloud projects add-iam-policy-binding $GCP_PROJECT \ --member serviceAccount:$GCS_SA_EMAIL \ --role roles/browser HALYARD_HOST=$(echo $USER-halyard-`date +%m%d` | tr '_.' '-') gcloud compute instances create $HALYARD_HOST \ --project=$GCP_PROJECT \ --zone=$ZONE \ --scopes=cloud-platform \ --service-account=$HALYARD_SA_EMAIL \ --image-project=ubuntu-os-cloud \ --image-family=ubuntu-1404-lts \ --machine-type=n1-standard-4 gcloud compute ssh $HALYARD_HOST \ --project=$GCP_PROJECT \ --zone=$ZONE \ --ssh-flag="-L 9000:localhost:9000" \ --ssh-flag="-L 8084:localhost:8084"
Inside the Halyard VM:
KUBECTL_LATEST=$(curl -s https://storage.googleapis.com/kubernetes-release/release/stable.txt) curl -LO https://storage.googleapis.com/kubernetes-release/release/$KUBECTL_LATEST/bin/linux/amd64/kubectl chmod +x kubectl sudo mv kubectl /usr/local/bin/kubectl curl -O https://raw.githubusercontent.com/spinnaker/halyard/master/install/debian/InstallHalyard.sh sudo bash InstallHalyard.sh . ~/.bashrc GKE_CLUSTER_NAME=spinnaker GKE_CLUSTER_ZONE=europe-west1-d PROJECT_ID=$(gcloud info --format='value(config.project)') gcloud config set container/use_client_certificate true gcloud container clusters get-credentials $GKE_CLUSTER_NAME \ --zone=$GKE_CLUSTER_ZONE GCS_SA=gcs-service-account GCS_SA_DEST=~/.gcp/gcp.json mkdir -p $(dirname $GCS_SA_DEST) GCS_SA_EMAIL=$(gcloud iam service-accounts list \ --filter="displayName:$GCS_SA" \ --format='value(email)') gcloud iam service-accounts keys create $GCS_SA_DEST \ --iam-account $GCS_SA_EMAIL hal config version edit --version $(hal version latest -q) hal config storage gcs edit \ --project $PROJECT_ID \ --json-path $GCS_SA_DEST hal config storage edit --type gcs hal config provider docker-registry enable hal config provider docker-registry account add my-gcr-account \ --address gcr.io \ --password-file $GCS_SA_DEST \ --username _json_key hal config provider kubernetes enable hal config provider kubernetes account add my-k8s-account \ --docker-registries my-gcr-account \ --context $(kubectl config current-context) # Only required in case you want to use eu.gcr.io hal config provider docker-registry account add gcr-eu \ --address eu.gcr.io \ --password-file $GCS_SA_DEST \ --username _json_key hal config provider kubernetes account edit my-k8s-account --docker-registries my-gcr-account gcr-eu hal config deploy edit --account-name my-k8s-account hal config deploy edit --type distributed MY_SPINNAKER_BUCKET=<SPINNAKER_BUCKET_ID> hal config canary enable hal config canary google enable hal config canary google account add my-google-account \ --project $PROJECT_ID \ --json-path $GCS_SA_DEST \ --bucket $MY_SPINNAKER_BUCKET hal config canary google edit --gcs-enabled true \ --stackdriver-enabled true hal config canary edit --default-metrics-store stackdriver hal deploy apply hal deploy connect
This guide uses the Kubernetes V1 provider, but you can use V2 just as well.
Follow the official documentation to enable the V2 provider.
Visit localhost:9000 to open the Spinnaker UI.
In the applications page, create a new application:
Under Infrastructure, the Clusters view should normally be opened automatically. Click the Config link on the top right and enable Canary for this project.
This should enable Canary Analysis for the project.
The result should be that the Spinnaker menu for this project should be changed.
Pipelines are now nested underneath Delivery, which also now boasts Canary Configs and Canary Reports.
In case this is not visualised directly, you can refresh the cache by clicking on the Spinnaker logo on the top left of the page, and clicking the Refresh all caches link in the Actions drop down.
Under Infrastructure, switch to the Load Balancers view and create a load balancer.
Fill in the stack, port, target port and type
Under Infrastructure, switch to the Clusters view and create a Server Group.
Once the server group is created, it will show up like this:
By clicking on the little load balancer icon on the right-hand side, we can now visit the Ingress IP through the load balancer view on the side of the page.
Back in the server group section, clicking on the little green chicklet will display container information on the side of the page, including logs of the application.
Let’s do this for PROD as well.
Follow exactly the same steps as for DEV, except use
prod as Stack instead of
Once the PROD load balancer and server group are deployed, we’d like to make sure we never have downtime on PROD.
We can set up a Traffic Guard, responsible for making sure our production cluster always has active instances.
Go to the Config link on the top right of the page, and add a Traffic Guard.
Now that we’ve deployed a single version of our application to DEV and PROD, it’s time to create a pipeline.
This will enable us to continuously deploy new versions of our application without having to manually create new server groups every time.
Head over to the pipelines view and create a new pipeline called Deploy to DEV.
Under the first “Configuration” stage, configure an automated trigger.
Now add a stage to deploy our application.
We now have to add a server group as deployment configuration. We can reuse the existing deployment as a template.
Change the strategy to
It’s important to change the image being deployed, otherwise, we’d always deploy the image of our existing server group.
Go down to the Container section and select the Image from Trigger.
This will automatically change the container image at the top of the dialog box under Basic Settings.
Keep all other settings as they are.
Save the server group configuration, and save the pipeline.
When we now select the pipelines view, we can see the newly created Deploy to DEV pipeline.
We can test this by either starting a manual execution, or committing a change to our application GIT repository.
Create new pipeline Deploy to PROD.
Add a new Find Image from Cluster stage. This stage will allow us to look for the image we deployed to DEV, and pass that information on to upcoming stages.
Add a new Deploy stage to deploy the new DEV version into production.
Under deploy configuration, add a server group based on the one in DEV.
Make sure to set the right load balancer, i.e.
Scrolling down to the Container section, select the image found in DEV by the Find Image stage.
Since this is a new version we’d like to push to production, it would be a good idea to build in some safety measures to protect us from unexpected failure.
Using a canary release strategy allows us to limit the blast radius of potential issues that might arrise.
In the Basic Settings section, set the stack as
prod and the detail as
canary to indicate that this deployment is our canary deployment. Also use the
None strategy, since we just want to deploy this canary server group next to the one already active in production.
Now let’s test this out.
Change the application to respond with PickleRicks if that’s not already the case. Otherwise, make an insignificant change to the application and push the changes to GIT (master branch).
This should trigger a build, which should push a docker image to the GCR.
That on its turn should trigger the deployment to DEV, which - if successful - should trigger a deployment to PROD.
Once that’s done, your cluster view should look like this:
V001 on DEV, it has replaced the existing manual server group deployment using the highlander strategy.
Currently our canary is registered under the same load balancer as our production cluster. This means traffic is split between the canary and production.
We could test the canary manually by going to the ingress endpoint of the load balancer as we did on DEV. This could be sufficient for your needs, but Spinnaker offers automated canary analysis (aka. ACA), capable of automatically investigating traffic sent to the canary.
The ACA engine Kayenta will compare certain metrics between the currently running production version, and the newly deployed canary version.
Since comparing a fresh deployment with an old, potentially degraded deployment, could produce unwanted results, it’s advised to deploy both a canary and a current production instance labelled baseline, next to each other.
In the Deploy to PROD pipeline configure screen, add a stage in parallel with Find Image from DEV by clicking on the outer-left Configuration stage, and adding a new stage from that point on.
From that point forward, add another Deploy stage, with the prod server group as template.
At the bottom of the Deployment Cluster Configuration, switch the Container Image to the Find Image result for prod.
baseline as detail, and keep the strategy as
Save the pipeline.
We now have our basic setup of both a baseline and canary server group to perform canary analysis.
Our specific demo scenario uses Meeseeks from Rick and Morty as the new version to deploy.
As people who watched the series probably will know, Meeseeks can quickly become a threat to our way of living if we let nature run its course. Therefore, when switching to Meeseeks, we also write Meeseeks in the logs to keep track of them.
GCP uses Stackdriver for logging and monitoring, so if we’d like to use the logs as a source of information for canary analysis, we should make a Stackdriver metric using the Meeseeks logs.
In the GCP left-hand menu, under the Stackdriver section, you can find Logging and drill down to Logs-based metrics.
Add a new metric using the following filter, replacing the location and project_id by the zone name and project id from earlier in this guide:
(resource.labels.cluster_name="spinnaker" AND resource.labels.location="europe-west1-d" AND resource.labels.namespace_name="default" AND resource.labels.project_id="spinnaker-demo-184310" AND textPayload:"Meeseeks”)
Back in Spinnaker, head over to the Canary Configs view under Delivery.
Create a new Canary Config called Demo-config, and add a filter template.
The template will filter based on the replication controller of the server group:
Now we can add actual metrics to analyse.
Create a new Metrics Group called Meeseeks, with one metric underneath.
Since we’d also like to know whether our CPU or memory consumption has increased, let’s add some system metrics as well.
We can investigate which filters we can construct by using the GCP REST API.
Add a new group called Boring System Metrics, and add the following two metrics.
The only thing left to do for this Canary Config, is to configure thresholds for the Metric Groups.
The marginal is treated as a lower bound. If an interval analysis fails to reach the marginal limit, the entire canary release will be halted and no further intervals will be analysed. The pass limit is the upper bound, qualifying the analysis as a success. Anything in between will be recorded and next intervals will be analysed.
Save the Canary Config, and go back to the Deploy to PROD pipeline configuration.
Join both canary and baseline deployments into the Canary Analysis stage, by using the Depends On configuration.
Configure the canary analysis stage as follows.
Rollout or Rollback
After the Canary Analysis has run, the new version can safely replace the existing production server group.
Add a stage called Deploy to PROD, copying the production server group as template, and use the red/black (aka. blue/green) deployment strategy to avoid any downtime.
At the bottom of the Deployment Cluster Configuration, switch the Container Image to the Find Image result for DEV.
Regardless whether this pipeline actually succeeds or not, we need to make sure to clean up afterwards. Add a new pipeline called Tear Down Canary, with the following trigger.
Add two Destroy Server Group stages in parallel.
Configure the first one to destroy our baseline server group.
And finally also destroy the canary server group.
Running the demo scenario
As explained in the introduction of the demo application, we have two versions of our application. As long as we keep deploying green versions with minor changes to other parts of the application (not impacting Meeseeks logs), the whole pipeline should pass, including the canary analysis.
For a canary test to be successful, we need data.
The more data our test can gather, the more informed the decision will be.
In our demo scenario, we can continuously refresh the page to generate more load and more Meeseeks in the logs, but we can also use a script for that.
In the root of the demo repository, a script called
randomload.sh can be used to generate calls to the PROD ingress endpoint at a random interval.
The script uses HTTPie to make calls, but you can also replace it with
Also, make sure you change the IP address in your forked repository’s file.
A successful canary release would look like this.
Meeseeks logs should occur at a similar rate in the canary and the baseline server group.
CPU and RAM metrics are also part of the comparison. In the example below, the canary CPU metrics deviated too much from the baseline, resulting in a failure for that metric group. However, the weight of those metrics were not high enough to fail the verdict, but it did cause the outcome to be labeled marginal.
When switching to the blue Meeseeks version, the initial DEV deploy would succeed, but our canary analysis should fail after one or two intervals.
A failed canary release would look like this.
The canary server group generated a noticeable higher amount of Meeseeks than our baseline server group, resulting in a failed analysis.
Even though canary CPU and RAM metrics were quite in sync with the baseline, our Meeseeks metrics were enough to fail the entire pipeline.
Pickle Rick and Mr. Meeseeks have shown us the power of automated canary analysis using Spinnaker. There are still a few considerations we have to take into account, such as the importance of choosing the right metrics and filters and iterating on those after each canary release. Yet, having a tool like this at our disposal allows us to release more often to production, without compromising safety or quality. By reducing manual and ad hoc analysis only the most stable releases are deployed to production in a highly automated way.