cache_from parameter does not work

When using cache_from and providing a docker image with layers that do not need to be re-built, Concourse will re-build all of the layers every time instead of using cached layers. (It is the same behavior as if you did not use cache_from at all.)

Here is an example pipeline:

---
resource_types:
- name: pull-request
  type: docker-image
  source:
    repository: jtarchie/pr

resources:
- name: pull-request
  type: pull-request
  source:
    repo: quotecenter/app-thunder-storefront
    access_token: ((github-token))
    username: x-oauth-basic
    password: ((github-token))

- name: docker-image
  type: docker-image
  source:
    repository: gcr.io/np-quotecenter/app-thunder-storefront/pull-requests
    username: _json_key
    password: ((gcr-key))

- name: docker-image-master
  type: docker-image
  source:
    repository: gcr.io/np-quotecenter/app-thunder-storefront/master
    username: _json_key
    password: ((gcr-key))
    tag: latest

jobs:
- name: build
  plan:
  - aggregate:
    - get: app-thunder-storefront
      resource: pull-request
      trigger: true
      version: every

    - get: docker-image-master
      params:
        save: true

  - put: docker-image
    params:
      build: app-thunder-storefront
      build_args:
        NPM_TOKEN: ((myget-token))
      tag: app-thunder-storefront/.git/head_sha
      cache_from: [ docker-image-master ]

The docker image docker-image-master contains the exact same docker image that this PR is building - I have verified this several times.

The resulting output shows that it builds the entire image from scratch without using any caching:

Reproduce the behavior with the example pipeline, especially the saved docker-image-master input and its cache_from parameter. Trace how the docker-image resource handles cache_from during the image build and verify the result against the expected reuse of unchanged layers. Done means the build reuses matching cached layers instead of rebuilding the image from scratch.