Setup an easy way for local development that uses a multi node kubernets cluster

[yuvipanda]

Right now, development happens mostly with minikube. This has a couple of differences from production:

1. Build host & run host are always the same, so it's much quicker than production
2. We don't use an image registry, so any bugs around that will be uncaught.

Locally developing binderhub only requires the following:

1. Access to a .kube/config file that knows how to connect to a kubernetes cluster
2. An appropriately configured JupyterHub

The easiest way to accomplish these is to have a 'dev' kubernetes cluster that's small, often destroyed, and has the following structure:

1. We give access to it liberally
2. Each person gets their own namespace to play with
3. We also have a central docker image repository with namespaces for each person

This also makes testing quicker, since the network over at GKE is probably faster than local computer networks (at least for those of us still in the land of the free). It shouldn't be too expensive either, and we can probably do this soon.

[yuvipanda]

Another option is to investigate https://www.telepresence.io/

[richarddli]

(I work with the Telepresence team, and am not familiar with Jupyter and how it works). Hopefully the below info is helpful:

• Telepresence code a service locally, while proxying that service to a remote K8S cluster. So if for dev you need to run multiple different services in a K8S cluster but want to use, say, the Flask auto reload functionality, Telepresence works well.
• Telepresence doesn't alleviate the need to have a deployment system that builds your Docker image, pushes it to a repo, does a kubectl apply, etc. Given that you're in minikube, I'm guessing that you already have such a system. If not, you might want to look at https://forge.sh.

If there are specific enhancements or support you need if you want to try this, please feel free to hop on our Gitter (https://gitter.im/datawire/telepresence).

[manics]

Is this a suitable alternative #1318 ?
... now that I've read the issue properly it's obviously not 😄 Do you want to keep this open anyway?