Simplify Image Tag management for releases

[jlewi]

Right now we tag our docker images with v${DATE}-${SHA}. There are two problems with this approach

1. This format isn't very informative for users

2. It's really cumbersome to update images as part of an RC

   • We need to build the images
   • Then create a PR to update the images in the ksonnet configs.

I propose we tag images with v${RELEASE} e.g. v0.2.0 and v.0.2.1 and this work as follows

1. The tags v0.2.0 will be mutable until a release candidate is promoted to actual release

   • As we cut RC's the tag for the upcoming release will be updated to point at the most recent RC
   • Once an RC is promoted to the actual release the tag is no longer mutated

2. When we cut the release branch; we update all the images to use the tag of the next release e.g. 0.2.1

   • This means we don't have to update the configs each time we cut a new image; we can just move the release tag (v0.2.1) to the new image

3. To provide an audit trail we use a simple YAML that maps immutable sha based images to tags
   e.g.

images:
   - image: gcr.io/some-image/image:sha256@abcd...
        tags: 
            - gcr.io/some-image/image:v0.2.0

1. We can use scripts to apply the tags to a given image
2. To push a new image as part of an RC we just need to move the tags by
   i. updating the YAML file
   ii. rerunning scripts to reapply the tags

/cc @pdmack @kunmingg

[pdmack]

This sounds awesome!

[jlewi]

#1061 is an initial stab.

I think what we'll need some scripts to manage that YAML file for example

• Get the shas for the most recent tensorflow notebook images
• Apply some release tag to them
• Remove some release tag from some set of images.

[pdmack]

/cc @cheyang

[jbottum]

/area 0.4.0

[jlewi]

Closing this issue since we've committed a bunch of scripts to make this easier and there is no immediate follow on work.