Support using our E2E workflow to build a Docker image for releases.

[jlewi]

• Update our E2E test to use ksonnet not helm to deploy

  • As part of this we have a slightl loss in test coverage because our
    helm test provides some verification that our python tests don't.

• Create a releasing environment for workflows that has parameters set
  as needed to build using our release cluster.

Related to
#400 Use Argo for releases
#373 Improve our test harness

---

This change is 

[coveralls]

Coverage remained the same at 31.868% when pulling 93c2234 on jlewi:releases into ab1cb4d on kubeflow:master.

[jlewi]

/test all

[lluunn]

Reviewed 2 of 35 files at r1, 4 of 11 files at r2.
Review status: 6 of 37 files reviewed at latest revision, 2 unresolved discussions.

---

py/deploy.py, line 72 at r3 (raw file):

  util.run(["ks", "env", "add", env], cwd=app_dir)

  for k, v in params.iteritems():

params might be None right?
Should check it first.
Also component should not be None, better check it too.

---

py/deploy.py, line 145 at r3 (raw file):

    logging.info("Using GCP account %s", account)
    util.run(["kubectl", "create", "clusterrolebinding", "default-admin",
              "--clusterrole=cluster-admin", "--user=" + account])

Why does this have to be run every time?

---

Comments from Reviewable

[gaocegege]

There are some conflicts now, please resolve them. And it seems that the coveralls does not report failure now, we could close the issue #406

[jlewi]

Review status: 3 of 37 files reviewed at latest revision, 2 unresolved discussions.

---

py/deploy.py, line 72 at r3 (raw file):

Previously, lluunn (Lun-Kai Hsu) wrote…

params might be None right?
Should check it first.
Also component should not be None, better check it too.

params should never be None; it could be an empty dictionary.

Added a check for component although it too should always be uspplied.

---

py/deploy.py, line 145 at r3 (raw file):

Previously, lluunn (Lun-Kai Hsu) wrote…

Why does this have to be run every time?

It needs to be run once per cluster. The TFJob operator test creates a new cluster for each test run. We don't reuse clusters.

---

Comments from Reviewable

[gaocegege]

LGTM

Do we already use ksonnet instead of helm in our e2e test after the PR?

[jlewi]

@gaocegege Yes we use ksonnet not helm after this PR to deploy the operator.