Scale OpenFaaS functions to zero replicas after a period of inactivity
Premise: functions (Deployments) can be scaled to 0/0 replicas from 1/1 or N/N replicas when they are not receiving traffic. Traffic is observed from Prometheus metrics collected in the OpenFaaS API Gateway.
Scaling to zero requires an "un-idler" or a blocking HTTP proxy which can reverse the process when incoming requests attempt to access a given function. This is done through the OpenFaaS API Gateway through which every incoming call passes - see Add feature: scale from zero to 1 replicas #685.
faas-idler is implemented as a controller which polls Prometheus metrics on a regular basis and tries to reconcile a desired condition - i.e. zero replicas -> scale down API call.
The build requires Docker and builds a local Docker image.
TAG=0.1.9 make build TAG=0.1.9 make push
docker stack deploy func -c docker-compose.yml
kubectl apply -f faas-idler-dep.yml
Now decorate some functions with the label:
com.openfaas.scale.zero: true and watch the idler scale them to zero. You should also change the
-dry-run flag to
false. For example:
faas-cli store deploy figlet --label "com.openfaas.scale.zero=true"
- Environmental variables:
On Kubernetes the
gateway_url needs to contain the suffix of the namespace you picked at deploy time. This is usually
.openfaas and is pre-configured with a default.
Try using the ClusterIP/Cluster Service instead and port 8080.
||The URL for the API gateway i.e. http://gateway:8080 or http://gateway.openfaas:8080 for Kubernetes|
||host for Prometheus|
||port for Prometheus|
- Command-line args
-dry-run - don't send scaling event
How it works:
gateway_function_invocation_total is measured for activity over
1h of inactivity (or no HTTP requests)
You can view the logs to show reconciliation in action.
kubectl logs -n openfaas -f deploy/faas-idler