estafette / estafette-gke-preemptible-killer

Kubernetes controller to spread preemption for preemtible VMs in GKE to avoid mass deletion after 24 hours

Home Page:

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool


This small Kubernetes application loop through a given preemptibles node pool and kill a node before the regular 24h life time of a preemptible VM.



When creating a cluster, all the node are created at the same time and should be deleted after 24h of activity. To prevent large disruption, the estafette-gke-preemptible-killer can be used to kill instances during a random period of time between 12 and 24h. It makes use of the node annotation to store the time to kill value.

How does that work

At a given interval, the application get the list of preemptible nodes and check weither the node should be deleted or not. If the annotation doesn't exist, a time to kill value is added to the node annotation with a random range between 12h and 24h based on the node creation time stamp. When the time to kill time is passed, the Kubernetes node is marked as unschedulable, drained and the instance deleted on GCloud.

Known limitations

  • Selecting node pool is not supported yet, the code is processing ALL preemptible nodes attached to the cluster, and there is no way to limit it even via taints nor annotations
  • This tool increases the chances to have many small disruptions instead of one major disruption.
  • This tool does not guarantee that major disruption is avoided - GCP can trigger large disruption because the way preemptible instances are managed. Ensure your have PDB and enough of replicas, so for better safety just use non-preemptible nodes in different zones. You may also be interested in estafette-gke-node-pool-shifter


You can either use environment variables or flags to configure the following settings:

Environment variable Flag Default Description
BLACKLIST_HOURS --blacklist-hours (-b) List of UTC time intervals in the form of 09:00 - 12:00, 13:00 - 18:00 in which deletion is NOT allowed
DRAIN_TIMEOUT --drain-timeout 300 Max time in second to wait before deleting a node
FILTERS --filters (-f) Label filters in the form of key1: value1[, value2[, ...]][; key2: value3[, value4[, ...]], ...]
INTERVAL --interval (-i) 600 Time in second to wait between each node check
KUBECONFIG --kubeconfig Provide the path to the kube config path, usually located in ~/.kube/config. This argument is only needed if you're running the killer outside of your k8s cluster
METRICS_LISTEN_ADDRESS --metrics-listen-address :9001 The address to listen on for Prometheus metrics requests
METRICS_PATH --metrics-path /metrics The path to listen for Prometheus metrics requests
WHITELIST_HOURS --whitelist-hours (-w) List of UTC time intervals in the form of 09:00 - 12:00, 13:00 - 18:00 in which deletion is allowed and preferred

Create a Google Service Account

In order to have the estafette-gke-preemptible-killer instance delete nodes, create a service account and give the compute.instances.delete permissions.

You can either create the service account and associate the role using the GCloud web console or the cli:

$ export project_id=<PROJECT>
$ gcloud iam --project=$project_id service-accounts create preemptible-killer \
    --display-name preemptible-killer
$ gcloud iam --project=$project_id roles create preemptible_killer \
    --project $project_id \
    --title preemptible-killer \
    --description "Delete compute instances" \
    --permissions compute.instances.delete
$ export service_account_email=$(gcloud iam --project=$project_id service-accounts list --filter preemptible-killer --format 'value([email])')
$ gcloud projects add-iam-policy-binding $project_id \
    --member=serviceAccount:${service_account_email} \
$ gcloud iam --project=$project_id service-accounts keys create \
    --iam-account $service_account_email \


Prepare using Helm:

brew install kubernetes-helm
kubectl -n kube-system create serviceaccount tiller
kubectl create clusterrolebinding tiller --clusterrole=cluster-admin --serviceaccount=kube-system:tiller
helm init --service-account tiller --wait

Then install or upgrade with Helm:

helm repo add estafette
helm upgrade --install estafette-gke-preemptible-killer --namespace estafette estafette/estafette-gke-preemptible-killer

Deploy with Kustomize

Create a kustomization.yaml file:

kind: Kustomization
namespace: default
  app: preemptible-killer
- name: estafette/estafette-gke-preemptible-killer
  newTag: 1.1.21
- name: preemptible-killer-secrets
  - google-service-account.json=google_service_account.json
  type: "Opaque"

Apply manifests:

kubectl apply -k .


To start development run

git clone
cd estafette-ci-api

Before committing your changes run

go test ./...
go mod tidy


In order to test your local changes against an external Kubernetes cluster use the following commands:

# proxy master
kubectl proxy

# in another shell
go build && ./estafette-gke-preemptible-killer -i 10

Note: KUBECONFIG=~/.kube/config as environment variable can also be used if you don't want to use the kubectl proxy command.

For an all-in-one script that launches a kind cluster with 3 nodes, runs estafette-gke-preemptible-killer and then reports on the kill time, run:

go build && ./scripts/all-in-one-test -i 10

where -i 10 are the arguments to be passed to estafette-gke-preemptible-killer, replace with your own test arguments. For safety, it does not remove the kind cluster it leaves behind.


Kubernetes controller to spread preemption for preemtible VMs in GKE to avoid mass deletion after 24 hours

License:MIT License


Language:Go 91.0%Language:Mustache 5.2%Language:Shell 2.9%Language:Dockerfile 0.9%