|Version 10 (modified by hvr, 20 months ago) (diff)|
The Gitolite Switch
Currently, the developer setup on ghc.haskell.org is a bit complicated and unfortunate. Most importantly, it hosts all of the Git repositories we use. When a developer for GHC is inducted, we:
- Create them a user (i.e. a full shell account) on ghc.haskell.org
- Add their SSH public key
- Add them access to the darcs group (which owns the canonical, public facing Git repositories.)
Then, that user can clone from the server over SSH, and also push to the repositories directly with their new permissions.
This unfortunately has some downsides:
- Every user needs a full shell account.
- While we probably won't be forkbombed by someone, few people actually need a full shell account, and the principle of least privilege applies here. Really, everybody is just pushing to Git.
- Because of the previous point, group and access permissions on the repositories regularly get screwed up, causing situations where people don't have access (and can't push,) or we have to have post-receive hooks that modify the permissions. Both of these suck (this doesn't affect e.g. the Linux kernel developers, who have pull-based development models, because they can afford to.)
- Due to the current permission scheme, all users in the darcs group are effectively at the level of Trac admins, which can manipulate the trac.db database (this is needed by the Git hooks to update the Trac tickets). Again, the principle of least privilege should apply here.
- Moreover, people invariably fix this but it's always slightly patchy, and so the repositories that need 'fixing' for things like permissions are inconsistent, and it's hard to keep track of what needs to be maintained.
- All users can willy nilly create (and delete!) tags and branches, and perform some risky Git operations. Ideally, only release maintainers should have permission to do things like cut a release tag.
- Git admins need to perform risky direct manipulations in the file-system, even for the simplest Git repository administration tasks.
The proposed remedy is to use the Git-access wrapper ''Gitolite'' which provides an authorization layer on top of Git, and is executed as a separate system user (thus accessing the git repositories with only one Unix UID). Users accessing git repositories via ssh are discriminated by their ssh public key.
Gitolite also greatly simplifies user management, as a user management is little more than adding/removing a file containing the user's public key, and pushing that the administrative "gitolite-admin" git repository (Gitolite is administrated via git itself!). Similarly, adding a new git repository comes down to adding a few lines to the central Gitolite repository config file and pushing that file to the "gitolite-admin" Git repository.
Below are some notes about how we (Austin & Herbert) would like to go about doing this.
Ideally, most of the new setup can occur concurrently with the normal one undisturbed. Presumably 'the big switch' can happen in an hour or so downtime, in which we take the old URIs offline, bring Gitolite online and tell people this is the time to fix your push URLs.
For developers (with push permissions) who have already checked out repositories, the only change needed is to go over their repositories and update their git uris from
This can be accomplished either by editing their '.git/config' manually or using a Git command like git remote set-url origin NEW-URL.
The server currently isn't setup, but Austin can easily make it so.
- Tangential: should we deprecate the darcs.haskell.org URL? Who uses it? The name was known to be a funny misnomer from the Git switchover times, but As Far As Austin Knows, only GHC developers really use it these days. Perhaps we could just retire it.
- Austin notes that both nhc and yhc use it, so Malcolm and Neil will need to be asked.
- We can leave darcs.haskell.org as-is for legacy reasons, while creating new VHOST on http://git.haskell.org/ which would expose only the Gitolite-owned git repositories (and maybe also a Cgit front-page as e.g. on http://git.kernel.org)
- Who's actively committing, and does anybody beyond that actually need a shell account? It's unclear who uses ghc.haskell.org for what at the moment, beyond push access.
- Tangential: The current directory setup is a total mess on darcs.haskell.org, especially since the old darcs repos hang around there. Maybe we should clean it up if we're going to use it for a browseable directory.
Austin Seipp (thoughtpolice) and Herbert Valerio Riedel (hvr) can be contacted about details or specifics.