Deploy containers to Kubernetes

[feynmanliang]

• Adds K8s as a hard dependency, using minikube for local development
• Removes docker-compose dependency (all Clipper resources are provided on the k8s cluster)
• Removes fabric dependency (remote administration is done through K8s api now The R container that was added (R model deployment #195) reintroduced this
• Adds ClipperK8s class to handle Kubernetes interface

To test:

• Configure local k8s cluster using minikube

minikube start --insecure-registry localhost:5000
eval $(minikube docker-env)

• Run the following code:

from subprocess import Popen, PIPE
import numpy as np
import clipper_admin

# Sets host to be minikube's virtualbox network interface
clipper_host_public_ip = Popen(['minikube', 'ip'], stdout=PIPE).communicate()[0].strip()

clipper = clipper_admin.Clipper(host=clipper_host_public_ip)
clipper.start() # starts clipper query managers in k8s

Clipper and its components should be deployed to minkube

• Run kubectl get svc to get the k8s service port mappings. Substitute in the corresponding values into clipper_manager.py (this could be improved by making it a constructor argument)

• To deploy to clipper on k8s:

clipper.register_application("hello_world", "feature-sum-model", "doubles", "-1.0", 100000)
def feature_sum_function(xs):
      return [str(np.sum(x)) for x in xs]

# serializes predict function, builds docker images, creates k8s deployment which registers with query manager
clipper.deploy_predict_function("feature-sum-model", 5, feature_sum_function, "doubles")

• Test the endpoint by POSTing queries to $(minikube ip):1337/hello_world/predict

• Cluster resources can be deprovisioned with clipper.stop_all()

---

This change is 

[AmplabJenkins]

Can one of the admins verify this patch?

[feynmanliang]

@dcrankshaw ready for review

[dcrankshaw]

I'm currently reviewing this. I'm actually refactoring it a bit and adding some additional functionality as well, which I will submit as a pull request to this branch. So let's treat this as basically a feature branch for now.

From what I've looked at so far, this looks great. The functionality I'm adding is mostly refactoring the python package and adding plain Docker support so we don't break existing workflows when this gets merged.

[dcrankshaw]

Replaced by #251