What's the smallest instance size I can use?

[cisnky]

With 0.8.3 I could use t2.micro for the controller and the worker.

On 0.9.1 after setting up the cluster, I can't connect if using t2.micro.

[cisnky]

Worked my way up till it worked with t2.medium.

[pieterlange]

t2.micro (with 1GB memory) is pushing it for the controller server. On my cluster(s) the kube components peak above 1GB. I would also advise against using the T2 instance type for the controller if you're running anything serious.

Also note that kube-aws comes with a bunch of components by default: heapster and dashboard use some memory and CPU as well.

[mandrean]

From my company's real-world experiences, I'd say bare minimum for production is 1x m3.medium for both controller and etcd. But preferably at least 3x m3.medium for HA.

Skip the t2s for controller/etcd in production. You're screwed when you run out of CPU credits. Only use them for worker nodes.

[gamykla]

I got stuff working with on nanos with 0.8.2 .. I really wish smaller instances worked!

[mumoshu]

Thanks for bringing this up and for the info everyone!
I'm closing this as it is now documented in #162.

[scribnar]

@mandrean What you mean you are screwed if CPU credits expire? The baseline performance is still available so the controller will continue to run albeit with slow performance. I understand thats not a recommended scenario however you comment sounds like that is very dangerous scenario. I may be missing something here. Can you please explain "you are screwed"? Thanks.