title: “Design Doc - Push Reviews - Use Cases” permalink: design-docs/push-reviews-use-cases.html hide_sidebar: true hide_navtoggle: true toc: false folder: design-docs/push-reviews

Push Reviews - Use Cases

Objective

Enable code review for direct updates (pushes that update/reset the target branch directly to the pushed commit, including non-fast-foward updates).

Background

Gerrit's primary purpose is to support reviewing code changes, but there are some updates that currently cannot be reviewed:

  1. branch creation (ref creation + fast-forward from zero to the inital commit)
  2. branch fast-forward
  3. branch rewind (reset to a previous head)
  4. branch rewriting (rewind + fast-forward)
  5. branch deletion (rewind to zero + ref deletion)
  6. branch renaming (this is not equivalent to a create + delete since git actually renames the reflog when this happens which helps to preserve the history of the ref)

Direct Updates

The term ‘direct update’ is understood to mean an update of a branch that sets the branch to exactly the commit that was pushed by the user:

  • branch fast-forward
  • branch rewind (reset to a previous head)
  • branch rewriting (rewind + fast-forward)

IOW a direct update can be a fast-forward update (the commit that was pushed is a successor of the head of the target branch, aka ‘git push <SHA1>:refs/heads/<branch>’) but also a non-fast-forward update (reset to an arbitrary commit, aka ‘git push -f <SHA1>:refs/heads/<branch>’).

Problem Statement

Having valid use-cases (see below) that require doing direct updates with bypassing code review is a security problem because it means that the involved users need to have permissions to do direct pushes. If their account is compromised, an attacker could use these permissions to do malicious updates without involving another person (see also the SLSA Two-Person Reviewed requirement).

Also during normal operations being able to review direct updates increases safety, since reviews make erroneous updates less likely.

Use Cases

As a project owner I would like ...

  • ... to rebase a series of internal patches on top of a branch that is fetched from an open-source project (when this branch has changed) ...
  • ... to retroactively split up a repository (e.g. move a subfolder into a new repository while maintaining a rewriten history in both repositories) [1] ...
  • ... to import arbitrary commits without changing the SHA1 (e.g. code drops that were received from external partners) ...

... without needing to bypass code review.

[1] requires using git filter-repo to rewrite the history of the repository/branch.

As a user I would like ...

As an administrator I would like ...

  • ... to be able to enforce code review for all updates without breaking use cases that require project owners to do direct updates today.
  • ... to make sure that any commit that potentially went into a build of an artifact is preserved, so that any time it can be proved how binaries were made.

Non-Goals

  • Support code review for updates that completely eliminate commits from the repository by erasing history (e.g. removing large files from the history of the repository, removing commits that have leaked data).
  • Support code review for submodule subscription updates.
  • Support GitHub-like pull requests.
  • Support code review for branch creations.
  • Support code review for branch deletions.
  • Support code review for branch renames.

Nice-to-have Use Cases

Supporting the following use-cases is out of scope for this design, but we see them as nice-to-have. The implementation of push reviews should try covering these use cases if they can be achieved with low additonal effort.

  • Improve working with direct updates that bypass code review:
    • Make open changes show a conflict when a non-fast-forward update has been pushed that removed the base commit of the change from the history of the target branch.
    • Reject uploads of commits that are based on commits that have been removed from the branch history by a non-fast-forward push for this branch without requiring the administrator to ban the removed commits manually (this is to prevent that changes are accidentally created for all commits that have been removed from the history of the target branch).

Acceptance Criteria

Reviewing direct updates is possible:

  • A direct update can be pushed for review.
  • The change shows the diff between the head of the target branch at upload time and the pushed commit.
  • The change becomes unsubmittable when the target branch is updated so that the base commit of the change no longer matches the current head of the target branch (IOW the change is only submittable when the head of the target branch at upload time matches the current head of the target branch).
  • Changes with large diffs are handled in an acceptable way, changes with unreasonable large diffs are rejected (changes with diffs that are too large for Gerrit to handle should be rejected so that malicious users cannot use super large push reviews to cause failures in Gerrit).
  • Submitting a change that applies direct updates requires more/other permissions than submitting normal changes if the submission results in:
    • a fast-forward update that adds more than 1 commit to the history of the target branch,
    • a non-fast-forward update that rewinds the branch or
    • a non-fast-forward update that rewrites the branch (rewind + fast-forward update).
  • On submit of the change the target branch is set to the commit that has been pushed.
  • The previous state of the branch is preserved for audit in the code review, so that users can inspect retroactively how the branch was changed.
  • Previously merged changes and old patch sets stay intact when a push review has been submitted.
  • Open changes show a conflict when a non-fast-forward push review has been submitted that removed the base commit of the change from the history of the target branch.
  • Uploads of commits that are based on commits that have been removed from the branch history by the submission of a non-fast-forward push review are rejected for this branch unless they are push reviews themselves (this is to prevent that changes are accidentally created for all commits that have been removed from the history of the target branch).