Notes for a Git fireside chat

This is simply an outline, with no actual content.

How Git works

  • Recommendation to read Pro Git
  • Fundamental model: the object database
    • Blobs, trees, tags, commits
  • Extra metadata
    • Branches
    • Lightweight tags
    • HEAD (and other symbolic refs)
    • Remotes
  • Mention of performance optimisations such as packfiles
  • Merge algorithms
    • recursive
    • ort
    • ours
  • Useful commands
    • git rev-parse
    • git merge-base
    • git symbolic-ref
    • git clone --depth=1
    • git commit --amend
    • git grep (although ripgrep is generally faster anyway)
    • git log -p (for patch) and -S function_name
    • git blame
  • Brief mention of what a submodule is
  • Brief mention of what git rebase is
  • How-tos
    • Stacked PRs in a squash-merge-to-main world
  • Case studies
    • $INTERNAL_WORKSPACE_TOOL push-files
    • my internal Git-related dotfiles
    • git-appraise

Problems with Git, and why Pijul is interesting

Merging

Git works on snapshots => merging is impossible in principle. Here’s the textbook example:

Start:

1
2

A
B

Branch 1:

1
2
3

A
X
B

Branch 2:

1
2
3

G
A
B

then

1
2
3
4
5

A
B
G
A
B

When merging branch 1 and branch 2, the snapshot-based merge algorithm can’t know which AB pair the X was added in between, so it must choose arbitrarily.

Conflicts

Conflicts are not modelled at all: a conflict is only a property of the working tree, and you can’t “commit it”. That means you can never tell Git how to resolve a conflict in such a way that it can reliably resolve the same conflict again in the future. (git rerere is a hack that in my experience frequently does not do the right thing.)