Portal:Toolforge/Admin/Certificates in k8s
This page is currently a draft.
More information and discussion about changes to this draft on the talk page.
This page contains information on certificates (PKI, X.509, etc) for the Toolforge Kubernetes cluster.
Kubernetes includes an internal CA which is the main one we use for cluster operations.
By default, kubernetes issued certificates are valid for 1 year. After that period, they should be renewed.
The internal kubernetes CA, generated at deployment time by kubadm expires after 10 years. The current CA is good until
Nov 3 14:13:50 2029 GMT
Worth noting that etcd servers don't use the kubernetes CA, but use the puppetmaster CA instead.
Most certs can be checked for expiration with
sudo kubeadm alpha certs check-expiration on a control plane node.
Use cases and operations
Description of the different certificate types we have in the cluster.
external API access
We have certain entities contacting external the kubernetes API. The authorization/authentication access is managed using a kubernetes ServiceAccount and a x509 certificate. The x509 certificate encodes the ServiceAccount name in the Subject field.
Some examples of this:
- tools-prometheus uses this external API access to scrape metrics.
- TODO: any other example?
Certificates for this use case can be generated using a custom script we have: wmcs-k8s-get-cert .
Usually, the generated cert will be copy&pasted into the private puppet repo to be used as a secret in a puppet module or profile.
Renewing the certificate is just generating a new one and replacing the old one.
internal API access
Some stuff running inside the kubernetes cluster also require a certificate to access the API server and use a ServiceAccount. This certificate is usually crafted as a Kubernetes secret for the utility to use it.
Some examples of this:
- our custom webhook: ingress admission controller
- our custom webhook: registry admission controller
Certificates for this use case can be generated using a custom script we have: wmcs-k8s-secret-for-cert.
After running the script, the secret should be ready to use.
Renewing the certificate is just generating a new one (running the script again and making sure the pod uses it).
Kubelet has two certs:
- A client cert to communicate with the API server
- A serving certificate for the Kubelet API
At this time the serving certificate is a self-signed one managed by kubelet, which should not need manual rotation. Proper, CA-signed rotating certs are stabilizing as a feature set in Kubernetes 1.17, and we should probably switch to that for consistency and as a general improvement. The client cert of kubelet is signed by the cluster CA and expires in 1 year.
All such client certs are rotated when upgrading Kubernetes, but they can be manually rotated with kubeadm as well. This should be as easy as running
kubeadm alpha certs renew on a control plane node as root.
It is possible to configure the kubelet to request upgraded certs on its own when they near expiration. So far, we have not set this flag in the config, expecting our upgrade cycle to be 6 months, roughly.
These certs are automatically generated by the maintain-kubeusers mechanism. When a new tool is created in Striker, the LDAP change is picked up by a polling loop in the maintain-kubeusers deployment, and the service will:
- Create the NFS folder for the tool if it isn't already there because of maintain-dbusers
- Create the necessary folders to set up the KUBECONFIG for the user.
- Create a tool namespace along with all necessary privileges, restrictions and quotas
- Generate a private key
- Request and approve the CSR for the cert to authenticate the new tool with the Kubernetes cluster
- Write out the cert to the appropriate files along with the KUBECONFIG
- Create a configmap named
maintain-kubeusersin the tool namespace that gives the expiration date of the cert to use for automatically regenerating the cert before it expires
- Deleting this configmap will cause the cert to be regenerated on the next iteration. This is the safest way to regenerate the certs manually.
Each cert includes a CN, which functions as the user name in Kubernetes, and can include groups as well ("O:" or organization entries). Tool certs currently have the CN of their tool name and one O of "toolforge".
This service runs in Kubernetes in a specialized namespace just for it using a hand-made Docker image, as is documented in the README of the repo. The toolsbeta version runs the
maintain-kubeusers:beta tag instead of the
:latest tag to facilitate staging and testing live without hurting Toolforge proper. Deploying new code only requires deleting the currently-running pod after refreshing the required image tag.
If someone has a need to rotate their tool user certs for some reason, run:
user@bastion $ sudo -i root@bastion # become <tool-that-needs-help> tools.toolname:~$ kubectl delete cm maintain-kubeusers
This will cause maintain-kubeusers to refresh their certs.
If the certs are deleted, you will need to instead run
kubectl delete cm maintain-kubeusers --namespace tool-$toolname as a cluster admin (such as root on a control plane node) since the tool won't be able to authenticate.
In case of a corrupt
.kube/config file, the same trick applies except, that
maintain-kubeusers will not read invalid YAML. Therefore, you will need to delete the tool's
.kube/config and then as a cluster admin, run
kubectl delete cm maintain-kubeusers --namespace tool-$toolname. That will regenerate their credentials.
All etcd servers use puppetmaster issued certificates (puppet node certificates). The etcd service will only allow communication from clients presenting a certificate signed by the same CA. This means kubernetes components that contact etcd should use puppet node certificates.
In the puppet profile controlling this, we have a mechanism to refresh the certificate and restart the etcd daemon if the puppet node certificate changes (it is reissued or whatever).
Some other interesting docs: