Using k6 to do load testing of the canary before rolling out traffic
Here's a demo (sped up):
Here's what the Canary
webhook can look like. This is pre-rollout
webhook, so it happens before any traffic is placed on the canary. If the webhook passes the thresholds, the rest of the Flagger analysis and promotion process occurs
See the k6 docs for a full description of options available in scripts
apiVersion: flagger.app/v1beta1
kind: Canary
...
spec:
analysis:
...
webhooks:
- name: k6-load-test
timeout: 5m
type: pre-rollout
url: http://<k6_loadtester_service_name>.<k6_loadtester_namespace>:<k6_loadtester_service_port>/launch-test
metadata:
script: |
import http from 'k6/http';
import { sleep } from 'k6';
export const options = {
vus: 2,
duration: '30s',
thresholds: {
http_req_duration: ['p(95)<50']
},
ext: {
loadimpact: {
name: '<cluster>/<your_service>',
projectID: <project id>, # Optional, will use the default project otherwise
},
},
};
export default function () {
http.get('http://<your_service>-canary.<namespace>:<service_port>/');
sleep(0.10);
}
upload_to_cloud: "true"
slack_channels: "channel1,channel2"
notification_context: "My Cluster: `dev-us-east-1`" # Additional context to be added to the end of messages
min_failure_delay: "2m" # Fail all successive runs after a failure (keyed to the namespace + name + phase) within the given duration (defaults to 2m). This prevents reruns. Set this to a duration slightly above the testing interval
wait_for_results: "true" # Wait until the K6 analysis is completed before returning. This is required to fail/succeed on thresholds (defaults to true)
env_vars: "{\"KEY\": \"value\"}" # Injects additional environment variables at runtime
kubernetes_secrets: "{\"TEST_VAR\": \"other-namespace/secret-name/secret-key\"}" # Injects additional environment variables from secrets, at runtime
Use the k6 environment variables feature to inject configurations and secrets to your script. To do so, mount your configs as environment variables onto the load tester and reference them with ${__ENV.<VAR_NAME>}
You can also refer to other secrets by using the kubernetes_secrets
setting in metadata. This is useful if your secrets are not located in the same namespace as the load tester or if you wish to limit the amount of secret to mount to the load tester. Note that you will need to assign a Kubernetes service account that can read the secrets in question to the load tester deployment
In order to send results to K6 cloud, the following conditions must be met:
- The script itself must support it. As shown above, in the
ext.loadimpact
, your script must define a test name and, optionally, a project ID - You must pass the
upload_to_cloud: "true"
attribute in your Canary's metadata - A
K6_CLOUD_TOKEN
environment variable must be set on the load tester's deployment
Once all of this is setup, results will be streamed to the cloud
Deploy this as a Service + Deployment beside Flagger:
- Set the
K6_CLOUD_TOKEN
environment variable if any of your tests will be uploaded to k6 cloud - Set the
SLACK_TOKEN
environment variable to allow slack updates
See the example directory for a full example on how the loadtester can be deployed along with a Canary referencing it
helm repo add flagger-k6-webhook https://grafana.github.io/flagger-k6-webhook
helm upgrade -i k6-loadtester flagger-k6-webhook
--namespace=flagger
--set webhook.vars.K6_CLOUD_TOKEN=token
--set webhook.vars.SLACK_TOKEN=slack_token
Whenever a k6 test is started, that process has to be managed by the application in order to avoid zombie processes.
This requires additional resources and the number of parallel test runs is therefore limited.
By default, that limit is set to 1,000 parallel k6 processes which can be configured using the MAX_CONCURRENT_TESTS
environment variable or the --max-concurrent-tests
flag.
If a new test request is received while the limit is reached, the request will be rejected with a HTTP 429 status.
The response also includes a Retry-After
header that should be respected by the client.