No comments yet.
It uses k3s as Kubernetes flavor and Hetzner Cloud as provider. For this test I used extremely high concurrency so the tool hung twice in the middle of the process because I was hitting the Hetzner API too hard, so I had to interrupt it and continue.
Excluding the time it paused/hung due to the API, I calculated around 11 minutes total for the cluster creation. This includes:
- creating all the resources (cloud instances, firewall, load balancer for the Kubernetes API)
- deploying k3s to the control plane (3 masters) and the 297 worker nodes
- installing the Hetzner Cloud Controller Manager, to link K8s nodes to the cloud instances as seen by the Hetzner control panel and be able to provision load balancers out of the box
- installing the Hetzner CSI driver, to be able to provision block storage volumes out of the box
- installing the Cluster Autoscaler, to allow configuring autoscaling node pools
- installing the Rancher System Upgrade Controller, to handle k3s upgrades very easily
I believe this is a world record, or at least I have never heard of a tool managed or not to create clusters that comes even close to mine in terms of speed. Am I wrong? I would be curious to hear if there is a tool even faster than mine.
Of course not many people need to create clusters with 300 nodes from the get go, but it was a fun experiment. In the version I will release in the coming few weeks (v2.0) I will likely limit concurrency to more sane levels to reduce the risk of hitting the Hetzner API too hard, although most people will create small clusters to begin with. During my testing, I was able to create a cluster without hanging etc with max 100 nodes. Beyond that it almost always hangs and there are retries to create/power on cloud instances. So yeah I am going to limit the concurrency just in case to perhaps 10 or 20 servers per time for real life use.
What do you think of this? Let me know! I am curious to hear feedback or any comments on the subject :)
