All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
50 nodes Kubernetes cluster in under 9 minutes
First tests with the foundation of what will be ClusterNinja: from zero to kubeconfig for a cluster with 50 nodes total (3 masters for HA, 47 worker nodes of different types in different pools and in different locations) created in only 8 minutes and 54 seconds TOTAL.
This includes everything, from creating all the resources (servers, load balancer, private network, firewall) to deploying Kubernetes and components required for managing upgrades as well as provisioning persistent volumes and load balancers out of the box.
The upgrade of the same cluster from Kubernetes 1.26 to 1.27 also took only 4 minutes.
Do you know how long these things take on GCP/AWS/Azure?
Thoughts?
Comments
Does grannies knitting wordpress load pretty quick then?
Francisco
no doubt
that's pretty impressive, was it from different providers and networks?
No, this test was with Hetzner for now with servers in 3 different locations. I need to do more work to do a test like this with multi cloud, although the first experiments (at a smaller scale) were successful.
I was wondering if any providers would be interested in partnering when I am ready to launch or soon after. ClusterNinja will be a "bring your own provider" managed service and will support some popular providers natively, but it will also be possible to use nodes from any provider of your choice. I could potentially partner with some LET providers with good reputation and offer them as recommended options for those who want to save more money.
Which providers on LET would be good to talk with about this?
Assuming that it will spawn on multiple provider, the challenge for the distributed storage is that how to encrypt the connection when all traffic comes from the public IP. You can use CNI that encrypt connection by default while taking hit to the storage performance, and maybe further performance hit when using public connection between providers, depend on the peering.
What I can think of is put IO-heavy apps like database to outside of the cluster for the multi-provider deployment
This work because VMs are created pretty fast by the API on Hetzner Cloud. It might be slower, for instance, on Azure, where VMs takes a bit of time to be created.
Great job! Love all your automatization stuff.
In the alpha version of the multi cloud node I have tested with, the nodes talk to each other over a wireguard mesh. For the storage, I have tested successfully with Longhorn, which I am going to recommend to users due to its ease of use and features, and for workloads that require high performance and do not require replication at storage level (such as databases that handle their own replication), I will recommend/install optionally Rancher's local path provisioner.
All this works beautifully from my initial tests
I saw you added the last sentence. Like I said for performance they can just provision volumes on the root disk with the local path provisioner. These disks are usually NVME or at least SSD, so they are OK for most tasks. But later on I could also add features like automatic provisioning of resources outside the cluster, if there is demand. But to be honest things have changed dramatically over the past 2-3 years, and there are some awesome Kubernetes operators that make Kubernetes the best platform even to run databases, which is something people would have thought more than twice before doing it only 2-3 years ago.
Yeah Hetzner is ridiculously fast, I have to agree. So far it's the fastest I have seen at creating servers. One newly created server is up and running in like 20-30 seconds. I wonder if they keep a bunch of servers ready at all times just to be assigned to an account, because otherwise I can't explain that speed.
It's pretty impressive @vitobotta keep it up!
Another great post @vitobotta. Looking forward to ClusterNinja.
In the OS tab they have (or had) a symbol that said ”rapid deploy” or something similar (they all did), I think that = it’s avaliable stand-by
Thanks!
Oh, I have never noticed that symbol
This seems great! We would possibly be interested if you’re still looking for providers.
Which CNI you deployed with?
This is the only way. Same as with fast FaaS.
The point is not starting fast but maintaining it, day-2 operations & being cheap enough. One provider that failed on "cheap" is/was elest.io IMHO.
Sounds good! I'll take a note about you so I'll get in touch a bit later.
How did they fail? What do you mean? Thanks
more than 10. just creating an EKS cluster (not nodes!!) takes 5mins.
Exactly
Too expensive IMHO for BYOC.
They have a completely different pricing than what I have in mind. I won't add a % on top of the instance price. You will be able to chose whichever nodes you want. ClusterNinja will likely be priced based on the size of the cluster. Still need to think about it but it's likely going that way. WDYT?
Upgrading empty cluster it's easy, for comparison shake, when upgrading the control plane in AWS/GCP it's slower but no downtime
Maybe start with your sales pitch, nodes availability (what happen if hosts run out of node) ?
Are you offering managed services?
The cluster was not empty, I installed several things on purpose to fill nodes etc.
Node availability is not my concern since this is a "bring your own provider" kinda of service, so that is responsibility of the infrastructure provider you choose. I don't provide infrastructure, although that may change if I do some nice partnerships with some providers here.
Yes, it will be a managed Kubernetes service with some nice features but I am not ready to share more details yet since I don't know what will be available at launch if we try to launch in 3-4 months.