Hi Daniel, On 2 January 2019 at 4:51:17 pm, Daniel Imberman (daniel.imber...@gmail.com) wrote:
I've also been extremely interested in this as well. I've been messing with GCP VM instances in the past few weeks to try to simplify my local build as well. Would definitely be interested in helping with the AIP + implementation. That's good to hear :) One thing I believe we should do is set up the ci base-image with all of the pip dependencies pre-loaded. A lot of time is wasted pip installing dependencies. We can auto-generate new images whenever a PR is submitted to this repository and then specify the tag in the .travis.yml when building. Sounds like a good idea. I've seen this pattern before, where there's a `app` Dockerfile that depends on an `app-dependencies` image. On the k8s side, I think we need to move away from minikube for k8s testing. I discussed in a previous email setting travis to work with GKE. I'd be careful about coupling k8s stuff too tightly with a docker infrastructure. That can get pretty dicey. I think as long as we're using a separate k8s cluster the k8s executor tests only need to gather the IP addresses + have access to the kubeconfig. I think we should mimic whatever the k8s community does regarding testing. And if this means moving away from minikube and towards actual cloud infrastructure, so be it. We just need to document this decision, and how local testing would work for non-committers, which won't necessarily have access to an actual k8s cluster. We might need to manage expectations here. -- Gerardo Curiel // https://gerar.do
