Hacker News Reader

Comment by mannykannot | Hacker News Reader

Could you give an example of where it is necessary (in principle, not just because that's the way the tool works) to rewrite the detailed history in order to get the history you want?

I realize that we generally have to work with the available tools, but it is also useful to think about how things would optimally work. When I wrote that frequent rebasing looks like a process smell, that could because the tool is not optimal.

View on HN

I rewrite history (not in the upstream) every day, multiple times a day. I do it to split commits. I do it to squash commits. I do it to reorder commits. I do it to make my commits easier to review by whoever is doing code reviews for me. I do it make my commits logical: bug fixes get their own commits, features get their own commits, tests get their own commits if that's what the upstream wants, ...

It's the only way to do things that yields useful history in the upstream. What is useful history in the upstream? It's history that others can read (linearly!) that is informative and makes sense and makes it easy to bisect, git blame, and so on, to find bugs, to understand changes in the smallest logical units.

This is impossible to do without rebasing.

If you find yourself copying changes to another clone to commit them one at a time, then you are rebasing, and you just didn't know it.

History in the upstream is sacred. Unpushed history is absolutely not.

Do you do code review using a formal tool, like Gerrit or Phabricator? If so you already have a "code review history", separate from the repo history. The code review history is at times interesting to review, because it contains discussions, tradeoffs, etc.

Given that we have this secondary history, why require a completely different tool to track and access it? That's just pointless duplication. We should track all the history using the same tool.

Now to your point, it's still useful to call out "final" commits for bisecting, blame, etc. So it would be good to group commits, and hide the detailed history by default.

I very much like features like built-in wiki (which is trivial to do in Git anyways, using either a separate branch or a separate repo with a named derived from the base repo), built-in issue tracker (this is harder to do in Git, though there exist projects that do it), built-in code review, ...

Still, I've worked with codebases sized in the hundreds of millions of lines of code. To deal with that level of complexity one needs things like OpenGrok, cscope, and so on, to find one's way around. And when it comes to history, I could not care less about past code reviews or history internal to a feature branch. When I need to do `git blame` or look through commit history or a large codebase, I want to see clean history with a high signal-to-noise ratio. The more noise, the slower I'll make progress on understanding whatever code/history I'm trying to understand, therefore the slower I'll make progress on bug fixing or feature development -- I might even give up on history, and lose a lot of important information, if the noise level is too high.

For me the ability to rebase, and to require clean history, trumps all the great things in Fossil -- each and every one -- that Git lacks. And this even though I love Fossil's design.

Is it really true you never want to refer to a code review history? It can provide important context missing from even a well-commented commit.

Regardless it's possible to have both. An example is hg's changeset evolution. With changeset evolution, each commit has two histories: the repo history and the changeset history. Commands like `blame`, `log`, etc. show only the repo history; a separate set of commands accesses the changeset history.

An example where this is useful: sometimes rebasing can inadvertently produce bugs, such as collapsing two identical lines which ought to have been duplicated. `git blame` cannot check if that happened. But the changeset history, by tracking the rebase, can tell you that.

Yes, it is. In my many years in this business I have never gone back to any code review -- of my code, my reviews of others' code, or anyone's review of anyone's code.

EDIT: I suppose I might look at past code reviews when evaluating a candidate for employment. Still, there is no need to store those along with code. And if a code review comment needs to be recorded for posterity, it gets recorded in the code or in the commit comment.

No. I worked for years with software that had no way of changing history. It was never required.

It was required at Sun, in Solaris engineering. Merge commits were absolutely verboten (which prohibition was enforced by tooling) -- therefore pushes had to be linear. Commit commentary had a very specific required format.

Clean, linear history has never been required anywhere else I've worked, but I've done it ever since Sun taught me to. Just because it's not required doesn't mean it's not a good idea, and it can't be forbidden ("what happens on my dev instances, stays in my dev instances", and the only thing seen in the end is what I choose to publish, and it's going to be clean and linear).

Large projects at Sun used a rebase-heavy / rebase-only workflow like so:

                +-----------------+
                | Upstream "gate" |<------+
                +-----------------+        \
                  /     \                   \
                 /       \                   \
                v         v                   \
    +--------------+    +--------------+     +--------------+
    | Project gate |    | Project gate |     | Project gate |
    | (re)based on |    | (re)based on | ... | (re)based on |
    | build N of   |    | build N+1 of |     | build N+M of |
    | upstream     |    | upstream     |     | upstream     |
    +--------------+    +--------------+     +--------------+
            ^             ^
            |            /
            |           /       ...
            v          v
      +-----------------+
      |   dev clones    |
      |                 |--+
      | periodically    |  |
      | rebased --onto  |  |--+
      | next rebasing   |  |  |
      | of project gate |  |  |
      +-----------------+  |  |
        | ...              |  |
        +------------------+  |
           | ...              |
           +------------------+

In large projects individuals tracked a project fork of the upstream, which forks were periodically rebased onto the latest "build" of the real upstream, and the individuals' forks of the project "gates" were rebased onto the latest project fork as needed. At the end, when all the i's were dotted and t's crossed, the tech lead would push the project gate's linear history additions to the upstream.

We did this with early 90s tech known as Teamware, with lots of scripting on top. We later did it with Mercurial (again, with lots of scripting on top). Mercurial was a mistake. Git is much, much easier to use this way than any other VCS I've ever worked with, which for me includes: CVS, Clearcase, PRCS, Subversion, Mercurial, Git.

Mercurial was a mistake.

Anecdotally, as a fellow Sun Alumni, I strongly disagree (and I know many Sun Alumni would as well). I greatly miss using Mercurial instead of git now that I work elsewhere. The cadmium extension we had in-house at Sun more than made up for any plausible deficiency and Mercurial phases let me mostly drop use of cadmium.

Reply on news.ycombinator.com