Skip to content

Latest commit

 

History

History
125 lines (91 loc) · 4.94 KB

rebase.md

File metadata and controls

125 lines (91 loc) · 4.94 KB
Raw

This code is experimental and might break you if not used correctly.

The rebase command efficiently rewrites an image to replace the base image it is FROM with a new base image.

rebase visualization

(link)

This is not safe in general, but it can be extremely useful for platform providers, e.g. when a vulnerability is discovered in a base layer and many thousands or millions of applications need to be patched in a short period of time.

A commonly accepted guideline for rebase-safety is ABI-compatibility, but this is still imperfect in a handful of ways, and the exact contract varies between platform providers.

Rebasing is best suited for when rebuilding is either impossible (source is not available) or impractical (too much work, too little time).

Using crane rebase

For purposes of illustration, imagine you've built a container image my-app:latest, which is FROM ubuntu:

FROM ubuntu

RUN ./very-expensive-build-process.sh

ENTRYPOINT ["/bin/myapp"]

A serious vulnerability has been found in the ubuntu base image, and a new patched version has been released, tagged as ubuntu:latest.

You could build your app image again, and the Dockerfile's FROM ubuntu directive would pick up the new base image release, but that requires a full rebuild of your entire app from source, which might take a long time, and might pull in other unrelated changes in dependencies.

You may have thousands of images containing the vulnerability. You just want to release this critical bug fix across all your apps, as quickly as possible.

Instead, you could use crane rebase to replace the vulnerable base image layers in your image with the patched base image layers, without requiring a full rebuild from source.

$ crane rebase my-app:latest \
  --old_base=ubuntu@sha256:deadbeef... \
  --new_base=ubuntu:latest \
  --tag=my-app:rebased

This command:

  1. fetches the manifest for the original image my-app:latest, and the old_base and new_base images
  2. checks that the original image is indeed based on old_base
  3. removes old_base's layers from the original image
  4. replaces them with new_base's layers
  5. computes and uploads a new manifest for the image, tagged as --tag.

If --tag is not specified, its value will be assumed to be the original image's name. If the original image was specified by digest, the resulting image will be pushed by digest only.

crane rebase will print the rebased image name by digest to stdout.

Base Image Annotation Hints

The OCI image spec includes some standard image annotations that can provide hints for the --old_base and --new_base flag values, so these don't need to be specified:

  • org.opencontainers.image.base.digest specifies the original digest of the base image
  • org.opencontainers.image.base.name specifies the original base image's reference

If the original image has these annotations, you can omit the --old_base and --new_base flags, and their values will be assumed to be:

  • --old_base: the base.name annotation value, plus the base.digest annotation value
  • --new_base: the base.name annotation value

If these annotation values are invalid, and the flags aren't set, the operation will fail.

Whether or not the annotation values were set on the original image, they will be set on the resulting rebased image, to ease future rebase operations on that image.

crane append also supports the --set-base-image-annotations flag, which, if true, will set these annotations on the resulting image.

Caveats

The tool has no visibility into what the specific contents of the resulting image, and has no idea what constitutes a "valid" image. As a result, it's perfectly capable of producing an image that's entirely invalid garbage. Rebasing arbitrary layers in an image is not a good idea.

To help prevent garbage images, rebasing should only be done at a point in the layer stack between "base" layers and "app" layers. These should adhere to some contract about what "base" layers can be expected to produce, and what "app" layers should expect from base layers.

In the example above, for instance, we assume that the Ubuntu base image is adhering to some contract with downstream app layers, that it won't remove or drastically change what it provides to the app layer. If the new_base layers removed some installed package, or made a breaking change to the version of some compiler expected by the uppermost app layers, the resulting rebased image might be invalid.

In general, it's a good practice to tag rebased images to some other tag than the original tag, perform some confidence checks, then tag the image to the original tag once it's determined the image is valid.

There is ongoing work to standardize and advertise base image contract adherence to make rebasing safer.