Re: [RFC PATCH 12/12] commit-graph: update design document

Derrick Stolee <dstolee@microsoft.com> writes:

See comments below, but this looks good to me.

What is missing from the "Future Work" section (and which was somewhat implied by now removed "When this feature stabilizes enough to recommend to most users") is safety against history [view] changing features: git-replace, shallow clone and grafts file. I wrote about this in "Re: [PATCH v8 04/14] graph: add commit graph design document" https://public-inbox.org/git/86vacsjdcg.fsf@gmail.com/

JN> The problem in my opinion lies in different direction, namely that JN> commit grafts can change, changing the view of the history. If we want JN> commit-graph file to follow user-visible view of the history of the JN> project, it needs to respect current version of commit grafts - but what JN> if commit grafts changed since commit-graph file was generated? JN> JN> Actually, there are currently three ways to affect the view of the JN> history: JN> JN> a. legacy commit grafts mechanism; it was first, but it is not safe, JN> cannot be transferred on fetch / push, and is now deprecated. JN> JN> b. shallow clones, which are kind of specialized and limited grafts; JN> they used to limit available functionality, but restrictions are JN> being lifted (or perhaps even got lifted) JN> JN> c. git-replace mechanism, where we can create an "overlay" of any JN> object, and is intended to be among others replacement for commit JN> grafts; safe, transferable, can be turned off with "git JN> --no-replace-objects <command>" JN> JN> All those can change; some more likely than others. The problem is if JN> they change between writing commit-graph file (respecting old view of JN> the history) and reading it (where we expect to see the new view). JN> JN> a. grafts file can change: lines can be added, removed or changed JN> JN> b. shallow clones can be deepened or shortened, or even make JN> not shallow JN> JN> c. new replacements can be added, old removed, and existing edited JN> JN> JN> There are, as far as I can see, two ways of handling the issue of Git JN> features that can change the view of the project's history, namely: JN> JN> * Disable commit-graph reading when any of this features are used, and JN> always write full graph info. JN> JN> This may not matter much for shallow clones, where commit count JN> should be small anyway, but when using git-replace to stitch together JN> current repository with historical one, commit-graph would be JN> certainly useful. Also, git-replace does not need to change history. JN> JN> On the other hand I think it is the easier solution. JN> JN> Or JN> JN> * Detect somehow that the view of the history changed, and invalidate JN> commit-graph (perhaps even automatically regenerate it). JN> JN> For shallow clone changes I think one can still use the old JN> commit-graph file to generate the new one. For other cases, the JN> metadata is simple to modify, but indices such as generation number JN> would need to be at least partially calculated anew.

Note that in all cases one can simply discard generation number information (treating it as if it was ZERO), and use commit-graph as values before applying history-changing feature: replacements, grafts or shallow.

Well, at least for shallow you can do that for generation numbers: using grafts (deprecated) or replacements to join repository with historical one would mean that we are no longer have commit-graph transitively closed under reachability.

All right (though "verify" mode is actually done via "check" command).

I'm not sure if this paragraph should be deleted as a whole; it depends on whether we decide that using git-gc to do automatic graph writes is enough (git-gc should be ran automatically by git if we get many new objects anyway, so this gives us almost "compute a graph on 'clone', 'fetch', or 'repack'").