Can anyone who has experience with real deployments advise on what pain-points may be encountered by growing something like this further than N nodes (and talk about what N might be?)
If you're interested in books on the topic, I like Kubernetes in Action (https://www.manning.com/books/kubernetes-in-action) or Kubernetes: Up and running (http://shop.oreilly.com/product/0636920043874.do)
One of the challenges with Kubernetes is that it's pretty fast moving, so it's a good idea to work with resources that are up to date. I know a last commit from May doesn't seem very old but that's going to a least be missing 1.7 and could be missing 1.6 or earlier as well.
Project plug. I am building a free service and tool for instantaneous cloud-running ephemeral Kubernetes clusters called Kubernaut. If you just want to play around with Kubernetes and start learning it is a great tool. Also useful for CI use via Travis: https://github.com/datawire/kubernaut
1) k8s makes a lot of sense for stateless applications (such as your website) but not so much for stateful applications that require a client to connect to the same container every time (there are ways to do it, but they are a pain in the ass.)
2) Tooling is getting better with time, but it's still pretty green. Packages for your usual orchestration tools like Puppet and Ansible are volunteer work so they get easily out of sync, or require more work than you'd expect to get going. Using their suggested YAML format leads to another problem: there's no easy way to keep secrets outside of the configuration files, unless you build your own process around it
3) Some pieces, like a replicated DB, might be easier to be kept outside of the k8s cluster. You can technically run them there, but they weren't designed for running in that kind of environment and sometimes it shows
4) The CI/CD story using Jenkins pipelines is far from solved. There are several packages that provide some solutions, but the documentation tends to be horrible and that leads to days of debugging through trial and error
5) Leaky abstractions everywhere. As an example, the Jenkins plugins to build using your k8s cluster suggests using the "credentials" system, but you need to add them manually after you boot the Jenkins service. Then your slaves stop receiving the credentials and you have to reboot Jenkins (I had to reboot it on average every 3 or 4 builds.)
Don't let that discourage you from using k8s as a "run" cluster, if your app is a shared-nothing, stateless app. It's so much easier to setup (specially on Azure and GCP) and it obviates the need for setting up Puppet + Sensu + load balancing just to make sure your service keeps running when a node dies.
Isn’t that what the Secrets resources are for? Just use those, and mount them into your containers in the filesystem or environment variables.
If I control a physical cluster of machines end-to-end, I might consider Kubernetes (because I don't have any better options and doing the work to actually secure it is probably less work than my alternatives), but it's also one of the reasons--though far from the only reason--that I couldn't begin to consider k8s if I'm running in AWS or another cloud environment.
They are in progress towards encrypting at rest the etcd database. You can experimentally turn it on by following this doc:
https://kubernetes.io/docs/tasks/administer-cluster/encrypt-...
https://docs.google.com/document/d/1lFhPLlvkCo3XFC2xFDPSn0jA...
https://github.com/kubernetes/kubernetes/pull/41939
https://github.com/kubernetes/kubernetes/pull/49350
etc.
There is an implementation to use Hashicorp's Vault to encrypt secrets (using this functionality) which looks quite nice actually:
https://docs.google.com/document/d/15-baW4i7qws1yxxIYjHXqKpk...
2) What exactly do you not like tooling wise.
3) Totally agree. I am actually working on replacing etcd with postgresql though.
1) Write a bunch of bash scripts around kubectl and a bunch of YAML files. While painful, this the way I ended up going for (plus blackbox to GPG encrypt/decrypt secrets on the repo)
2) Try to use your usual tools (Ansible/Puppet) as a replacement for kubectl. This is the dream, but the plugins for Ansible and Puppet only support subsets of the latest k8s features, or require annoying stuff like setting your API endpoint on every task.
In other words, getting reproducible deploys between different k8s clusters (say, one for staging and one for production) is not really a thing yet. I guess there's an argument for saying that one should use k8s-native solutions for that (such as namespaces) but what about having clusters in different data centers?
How do you view that in regard to this discussion? https://github.com/kubernetes/kubernetes/issues/1957