Skip to content

Latest commit

 

History

History
93 lines (77 loc) · 4.29 KB

README.md

File metadata and controls

93 lines (77 loc) · 4.29 KB
Raw

kubecfn

Cloudformation based installer for reasonably secure multi-node kubeadm cluster.

Status

This still has some rough edges, see the issue. There are still issues requiring manual intervention but it's designed to fail graceful in these cases. The rolling upgrades of the masters use the WaitOnResourceSignals UpdatePolicy and scripts to ensure it only continues with the rollout of the cluster is heathly.

We try to be reasonably secure, meaning all components are secured via TLS and RBAC is enabled. Yet, due to the user-data size limits we need to fetch the TLS keys from a S3 bucket. The permission for this is granted as an IAM instance profile, that means you need to deploy a metadata proxy to to block access to the metadata service. This isn't ideal but following the current best practices. We recommend kiam.

Operations

You can either edit the Makefile or use environment variable to override specific settings.

  • After deploying the cluster, you need to install kube2iam. Set your AWS account ID in manifests/kube2iam.yaml and apply the manifest.

Known issues / README FIRST / FIXME

  • Run all kubeadm generation steps in the Docker container (if in doubt, run all). This is required since kubeadm alpha phase kubeconfig has hardcoded file locations.

  • When deleting the stack, it will fail to delete the hosted zone because we created DNS records from the lambda. In this case delete the records manually and retry deletion. (#1)

  • Same is true for cloud-provider-integration managed resources like ELBs. These should be deleted in kubernetes first. If that's not possible, the resources need to be deleted manually so cloudformation deletion can finish. (#1)

  • On rolling upgrades etcd-members are suppose to remove themself from the cluster. This isn't working reliably yet. If this happens, the new replacement node can't join the cluster. This will block the rollout. In this case make sure the old node is actually terminated and remove it from the cluster with /etc/etcdctl-wrapper member remove. (#2)

  • Rolling upgrades for workers is disabled right now. When updating, kill the old instances manually. (#3)

  • Sometimes kubeadm fails, probably when it comes up before etcd reached quorum and fails (but can be restarted) #4

  • Sometimes ignition fails to get assets from s3 and reboots as a slow form or 'retry': coreos/bugs#2280

Create cluster

  • modify config/example.json, make sure to set ClusterState=new!
  • docker build -t cfn-make .
  • docker run -e AWS_ACCESS_KEY_ID=xx -e AWS_SECRET_ACCESS_KEY=yy cfn-make \ create-cluster
  • Wait for the cluster to come up and update it again. This will flip the initial-cluster-state flag in etcd, making sure that further updates can be rolled back and forward reliably:
  • docker run -e AWS_ACCESS_KEY_ID=xx -e AWS_SECRET_ACCESS_KEY=yy cfn-make \ update-cluster
  • Install networking plugin:
  • kubectl apply -f manifests/kube-flanne.yaml
  • modify config/example.json, make sure to set ClusterState=existing, otherwise replaced etcd instances won't be able to join the cluster.

"Dry run"

Cloudformation supports Change Sets which can be used to get the changes CloudFormation will do without actually updating the stack.

Create ChangeSet:

docker run -e AWS_ACCESS_KEY_ID=.. -e AWS_SECRET_ACCESS_KEY=.. -v $PWD:/usr/src/ \
  cfn-make cloudformation OP=create-change-set OPTS=--change-set-name=test2

To view the change set run:

aws --region us-east-1 cloudformation describe-change-set \
   --stack-name int2 --change-set-name test2

Create custom cluster

To create a second cluster, you need to override the name of the cloudformation stack. This can be done with the NAME environment variable. Since the stack uses a existing VPC but brings it's own subnets, the network ranges need to be adjusted in the config.