Re: [PATCH v2 3/3] Documentation/git-bundle.txt: discuss naïve backups

From: Junio C Hamano <gitster@pobox.com>
Date: Nov 2, 2024, 01:39 UTC
Message-ID: <xmqqzfmiza69.fsf@gitster.g>
In-Reply-To: <c50f9d405f9043a03cb5ca1855fbf27f9423c759.1730234365.git.code@khaugsbakk.name>

kristofferhaugsbakk@fastmail.com writes:

> +A naive way to make a full backup of a repository is to use something to
> +the effect of `cp -a <repo> <destination>`.  This is discouraged since
> +the repository could be written to during the copy operation.  In turn
> +some files at `<destination>` could be corrupted.

There is nothing wrong in the above, but ...

> +This is why it is recommended to use Git tooling for making repository
> +backups, either with this command or with e.g. linkgit:git-clone[1].

... "git bundle create - --all" in (or "git clone --mirror" from) a repository cannot be an equivalent to "cp -a" in a quiescent repository. You'd lose what is staged, you'd lose per repository configuration, you'd lose hooks, you may be losing other things I am not remembering.

IOW, those who uses what you call "a naive way" may be doing so with a valid reason.

So the patch is not quite correct as-is, even though it is going in the right direction. It discourages apples and recommends oranges.

Let me think aloud to see if I can come up with an improvement.

    It is tempting to do "cp -a <repository> <elsewhere>" or "tar cf
    - <repository>" to take a snapshot of a repository, but unless
    the repository is quiescent, you will get an incomplete snapshot
    with inconsiistent state.  If all you care about is the state of
    the objects and the refs that point to objects, you can instead
    use "git clone --mirror <repo> <elsewhere>" or "git bundle
    create <bndl> --all" to take a consistent snapshot.

    Note that you will lose other local states of the repository,
    such as the contents of the index, working tree, per-repository
    configuration, hooks, etc.

or something like that, perhaps?