For non-trivial reviews, when there are files with several changes, I tend to do the following in Git:

  1. Create local branch for the pr to review
  2. Squash if necessary to get everything in the one commit
  3. Soft reset, so all the changes are modifications in the working tree
  4. Go thru the modificiations “in situ” so to speak, so I get the entire context, with changes marked in the IDE, instead of just a few lines on either side.

Just curious if this is “a bit weird”, or something others do as well?

(ed: as others mentioned, a squash-merge and reset, or reset back without squashing, is the same, so step 2 isn’t necessary:))

  • Kache@lemm.ee
    link
    fedilink
    arrow-up
    25
    ·
    9 months ago

    If you were reviewing a “non-trivial” PR from me, I’d recommend not squashing because I would’ve broken it up into readable atomic commits.

    • fhoekstra@programming.dev
      link
      fedilink
      arrow-up
      3
      ·
      8 months ago

      This should be much more wide-spread. The hardest part of programming is reading someone else’s code.

      More people should learn to do git rebase -i, it’s a simple way to re-organise your commits to make sure that they tell a story to someone going through the PR commit by commit. It only takes a minute and can save your colleagues so much time and increase the quality of the review process.

      • Kache@lemm.ee
        link
        fedilink
        arrow-up
        2
        ·
        edit-2
        8 months ago

        Unfortunately it’s uncommon now that GitHub’s PR workflow dominates, so people think in terms of (often squashed) PRs and talk about “stacking PRs”. At least GitHub supports viewing PRs commit by commit.

        If PRs are just how it’s going to be, I wish GitHub could auto cut stacked PRs from a linear branch of commits.

        • tatterdemalion@programming.dev
          link
          fedilink
          arrow-up
          1
          ·
          8 months ago

          It’s surprising how Github is sorely lacking in git features and even encourages bad practices.

          1. There is only a linear history view of commits without any child/parent relationships visualized. I have coworkers that tell me it’s annoying how I use merge commits because they just see a big list of small commits in the history, and they’d rather see one commit per PR. Well… I sympathize, but also, fuck that! I’m not squashing my carefully crafted commits because Github has a shitty UI. Use git log --first-parent or something like lazygit. It’s sad that many open source projects have adopted a “squash and merge” mantra for PRs. All of the merge methods have a valid use case, but people are so beholden to the damned Github history view.

          2. PRs have no concept of what a patch is. If you rebase, the entire history disappears. Reviewers cannot see changes made to individual patches unless they are applied first in new commits and then manually squashed into old patches after the review is done. Phabricator does this better IMO; each patch has its own history. I’ve even heard devs at other companies tell me they never make PRs bigger than 100 LOC. That just seems like a huge waste of time to me, unless you have some special tools to support that bizarre workflow.

          3. Merge queues only support a single merge method, one of merge commit, rebase, or squash. So we just don’t use this desirable feature because of this limitation.

          • Kache@lemm.ee
            link
            fedilink
            arrow-up
            2
            ·
            edit-2
            8 months ago

            Ooh yeah PR as patches, persistent despite rebases, would be nice.

            Many git operations fundamentally have three SHAs as parameters (tree operations after all), and GitHub’s model simplifies it down to two.

      • tatterdemalion@programming.dev
        link
        fedilink
        arrow-up
        2
        ·
        edit-2
        8 months ago

        Or use a tool like StackedGit which makes the atomic commit workflow incredibly simple. Build atomic commits as you go instead of after you’ve written all of the code.