On 2020-09-23 at 12:34:24, The Sharp Ninja wrote: > # Proposal for Feature Sets and Feature Dependency Sets > > I believe a very useful feature for Git would be the addition of > Feature Sets and Feature Dependency Sets. > > ## Feature and Feature Set > > A Feature is a defined domain of functionality, such as a specific > service of an API. Features can be managed as a subset of the > repository, allowing for segmenting a repository into working groups > that have their own set of permissions for the pool of users with > access to the repository. As an example, User A may not have > permissions to merge within the repository, but within Feature A he is > allowed to do so. This would streamline integrating new users into > large repositories and teams. Git doesn't natively provide any functionality for restricting access to a repository more than by read-only or read-write permissions. This is intentional. Restricting merging to a set of files or even a directory causes problems detecting what files those are in the context of renames and copies. Plus, should Git consider the paths as they were in the merge base, or one of the heads? If one of the heads, which one? Currently, merges are symmetric except for the order of parents in the commit object, and this would break that property. This does seems like functionality you could add on using your CI system, a pre-receive hook, or features of your hosting platform. For example, your CI system could run a check that the user is operating only within their domain, and reject the operation otherwise. At $DAYJOB, we do this by requiring reviews by certain teams for certain files. If I want to merge code with approval only from my team, I must restrict myself to files my team owns. If I touch other teams' files, I must seek their approval. Without the relevant approvals, I cannot merge or deploy any code. All of this is built on top of Git and Git knows nothing about it. In addition, your proposal is very specific to your workflow where teams must touch only their code. In other organizations, such as mine, teams may touch any code with approval. There's no reason to think that one or the other approach is better or should be more privileged in Git itself, especially when both approaches can be effectively built on top. > ## Changes to Workflow > > Once a user has cloned a repository, he may chose to scope his work to > the Feature Set. All Fetch/Pull/Push operations are limited to the > Feature Set. Because a blob can be shared across paths and is only stored once, it's not possible to determine efficiently whether a blob is accessible only under a certain path. As outlined in gitnamespaces(7), trying to restrict access to only a subset of a repository can't be done securely, because attackers can make deltas against other objects. So this approach would be both inefficient and insecure. > ## Feature Dependency Sets > > Artifacts that are not part of a Feature, but directly affect the > Feature can be added to a Feature Dependencies Set. When scoping > actions to the Feature, artifacts related to the Feature by way of > inclusion in a Feature Dependency Set are also included in the > activities. If a commit includes changes to items in the Feature > Dependency Set then creating a pull request will generate two PRs, the > first a normal PR with normal permissions that includes only the > changes in the Features Dependency Set, to be adjudicated via the > normal workflow, and the second PR scoped to both the changes in the > Feature Set and Feature Dependency Set in the Feature's user-level > branch. Git has no concept of pull requests. Those are a feature typically built on top of hosting platforms like GitHub, GitLab, or Bitbucket. Some projects, such as Linux, don't use pull requests at all. -- brian m. carlson: Houston, Texas, US
Attachment:
signature.asc
Description: PGP signature
