Skip to content

Example deploying Eclipse Che on a Kubernetes cluster

License

Notifications You must be signed in to change notification settings

wernight/kubernetes-che

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

35 Commits
 
 
 
 
 
 
 
 

Repository files navigation

kubernetes-che

Example deploying Eclipse Che on a Kubernetes cluster. So what you get is a Che server running on your cluster behind SPDY Proxy to handle authencation and encryption.

Note: Currently alpha state; barely tested.

Usage

Deployment

  1. You need a running Kubernetes cluster (for example Google Container Engine).
  2. Edit kubernetes.yml and see the TODO; update with your values.
  3. Run kubectl apply -f kubernetes.yml
  4. Follow client-side set up (see below).

Client-side set up

For Chrome install SwitchyOmega. Then add a PAC Profile with the PAC Script below:

function FindProxyForURL(url, host) {
  if (dnsDomainIs(host, "che")) {
    # TODO: Replace with your host/IP pointing to SPDY Proxy.
    return "HTTPS che.example.com:44300";
  }

  return "DIRECT";
}

Click the lock to set username/password according to values you've set in kubernetes.yml.

Supposing you've set up your DNS to point to the Ingress Controller external IP (or actually to SPDY Proxy) and the SSL certificate of that host is trusted; you should now be able to connect to http://che/.

How it works

The beginning of kubernetes.yml is to get a valid TLS certificate. It sets up an Ingress Controller which, if you don't already have one, can be used for other services running on your cluster. It's not critical for the system to work, but you need a valid certificate when trying proxy through SPDY Proxy else your browser will refuse to use it.

SPDY Proxy is used to add authentication, encryption, and access che running on your cluster. The rules are set up so that it's only used when trying to access Che.

The critical part is that che resolves to the container running the Docker server used by the Che server, from withing your cluster (and thus also from outside your cluster via SPDY Proxy).

It runs the Che server and links it to a running Docker daemon (Docker-in-Docker is running also on Kubernetes).

Isn't Docker-in-Docker discouraged?

The original blog post is partially outdated. The current docker:dind is slow (because it uses --storage-driver=vfs) but should not corrupt data regardless of the underlying file system.

It's required here and desired for the following reasons:

  • CRITICAL: Kubernetes spawns on multiple machines but you need to have Docker and Che on the same host to be able to mount the same directory on both. These volume mounts are required by Che:
    • lib-copy used to bootstrap a workspace
    • storage used to access some settings like SSH keys
    • workspaces used to run docker run -v ... from within a Che workspace
  • Cleaner: Deleting the Pod or if it fails health checks it'll not block or pollute the file system of Kubernetes itself.
  • Provide a simple way to access the Docker host from within Kubernetes cluster (passed via --remote:che to Che).

What could be improved

  • Requires clients to setup and use that SPDY Proxy (at least until eclipse/che#1560 is fixed)
  • Long lived HTTP requests like the terminal, or possibly other parts, seem to disconnect after a while if there was no I/O. Not tested which part is responsible for this issue.

Releases

No releases published

Packages

No packages published