Kubernetes Gated Deployments facilitates A/B tests on application deployments to perform automated regression testing and canary analysis.
Kubernetes Gated Deployments extends the Kubernetes API to add a new type of
object called GatedDeployments
, that allows engineers to specify which
deployments and decision plugins to include in the A/B test. It uses a
controller in the Kubernetes control plane that is responsible for implementing
this behavior by retrieving and analyzing metrics from the backend specified,
decisioning the A/B test, and either rolling back or continuing the deployment.
- A
GatedDeployment
object is added to the cluster (see the section for Installing the controller) - The controller fetches
GatedDeployment
objects using the Kubernetes API - When the
treatment
deployment specified in theGatedDeployment
object is deployed and eligible for an A/B test, i.e., it has more than zero replicas and a different pod spec than thecontrol
deployment, the controller will start the experiment - The controller will poll the decision plugins and determine if the
treatment
deployment is causing harm to the metrics measured - The controller will either roll back the
treatment
deployment (by setting the number of replicas to zero), or promote thetreatment
deployment by setting thecontrol
deployment's image to that of thetreatment
deployment, followed by scaling thetreatment
deployment down to zero replicas
To create the GatedDeployment
controller on an existing Kubernetes cluster,
run the following:
kubectl apply -f gated-deployments.yml
This creates all the necessary resources and deploys the controller in the
kubernetes-gated-deployments
namespace.
Alternatively, Helm can be used to install and manage the resources and controller. To install, run the following:
helm install helm/kubernetes-gated-deployments --name kubernetes-gated-deployments
See the Developing section for running locally during development.
Create two identical deployments with different names (e.g.,
example-rest-service-control
and example-rest-service-treatment
). Initially,
the number of replicas for the treatment deployment must be set to 0 and control
deployment will be the only one taking production traffic.
Example deployment manifests are available here.
NOTE: The names of the deployments cannot be a prefix of the other. The gated
deployment controller uses the deployment name as the host prefix (since pod
names are of the form <deployment name>-<xxxxx>
and if one deployment name is
the prefix of the other, it will include data from all the pods. For example,
using example-rest-service
for control and example-rest-service-treatment
for treatment will result in control including the data for treatment as well.
Create an
example-rest-service-gated-deployment.yml
file like below:
apiVersion: 'kubernetes-client.io/v1'
kind: GatedDeployment
metadata:
name: example-rest-service
deploymentDescriptor:
control:
name: example-rest-service-control
treatment:
name: example-rest-service-treatment
decisionPlugins:
- name: newRelicPerformance
accountId: 807783
secretName: newrelic-secrets
secretKey: example-rest-service
appName: example-rest-service
minSamples: 50
maxTime: 600
testPath: /shopper/products
Save the file and run:
kubectl apply -f example-rest-service-gated-deployment.yml
In the example above, the GatedDeployment
object specifies that we want to
gate our deployments on the performance of the /shopper/products
path, between
the example-rest-service-control
deployment and
example-rest-service-treatment
deployment (the latter of which we deploy new
changes to). In this case, we specify that we want the controller to use the
newRelicPerformance
decision plugin to analyze performance data, which will be
retrieved from New Relic (which our application is instrumented with).
For this plugin, you will also need to create a secret containing the NewRelic
API key; an example is shown below. In this case, newRelic.secretName
is set
to newrelic-secrets
, and newRelic.secretKey
is set to
example-rest-service
. This means that the controller will look in its deployed
namespace for a secret called newrelic-secrets
, and look in the secret data
for the value corresponding to the key example-rest-service
.
apiVersion: v1
kind: Secret
metadata:
name: newrelic-secrets
type: Opaque
data:
example-rest-service: aW5zaWdodHNBcGlLZXk=
Within the deploymentDescriptor
section of the GatedDeployment
object, these
are the possible options to customize. All options are required unless
explicitly specified as optional.
Property | Description |
---|---|
control |
Section describing the control deployment. |
control.name |
Name of the control deployment. |
treatment |
Section describing the treatment deployment. This should be the one normally deployed, e.g. as part of your CICD pipeline. |
treatment.name |
Name of the treatment deployment. |
decisionPlugins |
Section containing the list of decision plugin config objects. See Plugin configurations below for details on specific plugins. |
Each type of plugin will require its own configuration. The following parameters are common to all plugins:
Property | Description |
---|---|
name |
The plugin name. This allows the plugin factory to find the correct plugin class. |
maxTime (optional) |
The maximum amount of time the experiment will run, at which point the A/B test will stop and automatically roll out the treatment deployment to the control deployment. When not specified, this defaults to 600 seconds (10 minutes) |
Plugins are designed to return one of three values:
WAIT
: if the analysis cannot make a conclusion about the metric yet, e.g., it requires a minimum amount of time or if the result is not yet statistically significantPASS
: if the treatment version does no harm to the metric analyzedFAIL
: if the treatment does harm to the metric analyzed
Property | Description |
---|---|
name |
Must be newRelicPerformance |
accountId |
Account ID of the New Relic account integrated with your application. |
secretName |
Name of the secret where your New Relic API keys can be found. This should be created in the namespace where kubernetes-gated-deployments is deployed. |
secretKey |
Name of the key in the secret specified in secretName that contains the New Relic Insights API key, used to run NRQL to collect performance data. |
appName |
Name of the New Relic application. |
testPath |
Path that you want to measure performance of for both deployments. |
minSamples |
The minimum number of samples required for each deployment before testing for significance. |
zScoreThreshold (optional) |
The Z Score threshold for Mann-Whitney U test. Defaults to 1.96, which corresponds to a p-value of 0.05 |
harmThreshold (optional) |
Maximum allowable ratio of treatment to control U values from the Mann-Whitney U Test before treatment is marked as causing harm. This defaults to 1.5. |
To contribute a new plugin, create a new plugins class in lib/plugins that is a subclass of Plugin
. At minimum, you should implement the following methods:
build
: this should create the plugin with any necessary setup_poll
: this is called periodically, and it should fetch and analyze metrics to return aDecisionResult
The following methods are implemented by default:
onExperimentStart
: this is called when the experiment starts, and sets the experiment start timeonExperimentStop
: this is called when an experiment ends, and clears the experiment start timeonExperimentPoll
: this is called on every polling interval; it will check if the maximum experiment duration has been reached and returnPASS
if it has, or it will return the result from_poll
.
To roll out a new version, update the treatment deployment with the new image and set the number of replicas to a non zero value (depending on the percentage of traffic you want to send to the new version).
Once the treatment deploy is rolled out, the gated deployment controller will
start a new experiment and start polling for decisions from the decision
plugins. The experiment runs until either all plugins have returned PASS
, or
any single plugin returns FAIL
, at which point the controller will set the
gatedDeployStatus
annotation on the treatment deployment to either noHarm
or
harm
respectively.
An example command to get the value of the annotation
kubectl get deploy -o jsonpath='{.metadata.annotations.gatedDeployStatus}' example-rest-service-treatment
This value can be periodically polled to check if the new version is causing harm or not in the CI/CD pipeline of the application. If the deployment causes no harm, the controller automatically rolls it out the new version to the control deployment. The status of the rollout can be checked using the below command.
kubectl rollout status deploy/example-rest-service-control
See CONTRIBUTING.md for how to contribute to this project.
You can develop locally with Minikube.
On Linux, the kvm2
driver provides better performance than the default
virtualbox
driver, but either will work:
minikube start --vm-driver=kvm2
minikube start
will configure your kubeconfig
for your local Minikube
cluster and set the current context to be for Minikube. With that configuration
you can run the kubernetes-gated-deployment
controller on your host operating
system:
npm start
Kubernetes Gated Deployments is MIT licensed.
- Steven Fu
- Satish Ravi
- Jacob Brooks
- Silas Boyd-Wickizer