Switch analyzer and restorer phases

[lukasberger]

• Analyzer restores metadata for cache and app layers in /layers:
  • Restores metadata and sha for all cache=true layers.
  • Restores metadata and sha for all launch=true layers.
  • Does not restore metadata for cache=false and build=true layers.
• Restorer attempts to restore all cache layers in /layers.
  • Restores data for all cache=true layers that are in the cache.
  • Removes metadata for all cache=true layers that are not in cache.
  • Leaves all cache=false metadata as they are.
• Remove obsolete analyzer testdata.
• Remove unused classifiyCache.

Part 2 of the implementation of buildpacks/rfcs#21.

[ekcasey]

This looks great! I do think we need to handle the no-op cache case in analyzer before we can merge. restorer can be skipped when there is no cache but analyzer should always run and a platform should be able to use the lifecycle w/o supporting providing a cache.

[ekcasey]

@lukasberger It occurs to me that there is a race condition here when an image is used as a cache. The image digest at the given tag can change between the execution of analyzer and restorer.

Once upon a time there was a similar race condition between analyzer and exporter. The "previous image" at a given tag may have changed during the time interval after it was analyzed and before we attempt to reuse layers. We solved that race condition by putting the digest reference to the previous image in the analyzed.toml file and using the immutable image at the digest reference from which to consume layers during export.

We may want to store the digest reference of a cache image during the analyze phase so that the restorer can use the exact same image. Otherwise, there could be some wacky edge cases where the metadata doesn't accurately describe the restored layer contents.

EDIT: retracted b/c of sha comparison

[lukasberger]

@lukasberger It occurs to me that there is a race condition here when an image is used as a cache. The image digest at the given tag can change between the execution of analyzer and restorer.

Once upon a time there was a similar race condition between analyzer and exporter. The "previous image" at a given tag may have changed during the time interval after it was analyzed and before we attempt to reuse layers. We solved that race condition by putting the digest reference to the previous image in the analyzed.toml file and using the immutable image at the digest reference from which to consume layers during export.

We may want to store the digest reference of a cache image during the analyze phase so that the restorer can use the exact same image. Otherwise, there could be some wacky edge cases where the metadata doesn't accurately describe the restored layer contents.

Since the restorer verifies the metadata sha restored by analyzer matches the layer sha in the cache image and deletes the layer metadata if the two don't match, is this really an issue? Presumably if the layer sha still matches, we don't care if the cache image changed since the layer data is the same.

[ekcasey]

@lukasberger good point. The previous case was different b/c there was no sha to compare. I retract my concern.