Hacker News Reader

Comment by howinteresting | Hacker News Reader

You're not alone. Git is an embarrassment to our profession.

No, it isn't. It's just a tool designed to work with complex software development workflows, which 90% of software projects do not need. 8 commands (pull, fetch, merge, push, add, commit, checkout and branch) are all you need to get going for git, which is like 2-3 more than what you need for svn/cvs. It's like getting upset that cars that are manual have an extra pedal (clutch) compared to automatic cars with their 2 (brake, accelerator).

Why does any serious workflow require you to be in the middle of an interactive rebase all the time?

None of my serious workflows require rebases, much less interactive ones.

Not that I don't do interactive rebases sometimes in local branches, but that I absolutely do not make them a part of any more general workflows. I rather try to avoid rewriting shared history where I can, and I'm happy with DAG navigation tools like --first-parent.

How do you manage stacks of commits for code review? (I'm guessing you just don't -- a lot of git users have really messy, unprofessional repository histories. The fact that git doesn't have great tooling for history management is also an embarrassment to our profession)

As I stated in the comment you are replying to:

> Not that I don't do interactive rebases sometimes in local branches

I don't have a problem with rebasing in local branches before code review. I just don't allow rebasing once a branch has started code review, and I especially don't allow rebasing in a public integration branch ever.

(I also rarely feel the need to rebase, personally, even in my own local branches. Working in better tools than git in a past life has given me a lot of the discipline I feel that I need to build good commits as I work. I make much more use of the git staging area than I see most devs generally do and use `git add --patch` and `git add --interactive` far more than `git rebase`. `git add --patch` especially is an under-sung hero that I think a lot of junior developers should learn well in advance of rebase.)

Also, I'm just not that bothered by messy histories in code reviews when I review other developers work. If the idea comes across even if the commit messages don't have the greatest narrative flow or are perfect, I'm perfectly fine with that. I make sure all merges are --no-ff and after that I can browse a "clean" history post code reviews with DAG traversal tools such as --first-parent. But that also gives me an opportunity to dig back into the mess months or years later after the code review if I need to research a bug or a regression. That sort of "sewage archeology" isn't "fun" by any means, but the number of problems I've been able to solve with it is high enough I appreciate keeping that mess around in source control. It's source control's job to keep the mess for me and present me pretty views for more regular work. That's why I prefer to trust tools like --no-ff and --first-parent over --squash, because that's what source control is built to do, control old histories messy or not.

What can I say. You're locked into a significantly worse maximum than is possible (and exists) using better tools. In a normal workflow your commits are small enough that you don't have to do "sewage archeology".

I don't think you are reading me correctly. That is a rare occurrence, but it is possible when using merge commits to correctly describe the DAG. It's entirely impossible if you overly rebase/squash. Those rare times that I've needed to do it it have "saved" a project a lot of time from a major bug or regression. My normal workflow I look at a "clean" faux linear view, but it's just the "--first-parent" view of merge commits, which gives me a "PR focused" or "integration focused" view of source control first.

If your average commit size is 30-40 lines of code, this isn't an issue. Stacks of small commits is a time-tested workflow, used by the Linux kernel among many other projects.

I think we are talking about different things at this point, but I may be losing the thread of your argument here.

I love a code review to consist of lots of small commits. I'm mostly fine if even some of them from developers more junior than me are just named "commit" if they are small.

When I finish the code review I prefer to only `git merge --no-ff`. All those small commits stay in the code history in the exact form they were reviewed under.

If I'm looking at history I'm most often using something like `git log --first-parent`. In my --no-ff integration branch that just shows me an integrated change list (PR list). Git gives me a "straight line" view down the DAG and hides irrelevant information I don't want in that moment like all the small "commit" commits.

If for some reason I find a bug or issue in some code, I may need to drill down into specific small commits including the "commit" commits, and I have that ability because all those commits are still in the DAG.

Huh? You only rebase when you're either:

* Trying to shorten commit history

* Trying to merge feature branches together (main or other feature branches)

* Trying to remove feature branches

None of which are necessary in a lot of development models. If you have that many feature branches that you have to rebase constantly, it doesn't sound like feature branches as a concept are properly understood.

The idea behind feature branching is one of 2 models:

1) either the main branch always works, and development is done in a feature branch, only to be rebased into main when development, testing, and sign off occurred, or

2) serious bug fixes occur outside the main branch, to be re-incorporated back into the main branch once it's been proven to be effective.

The 1st model should only require a rebase when the feature is complete. The 2nd model MAY require a couple of rebasing, dependent on how extensive the bug fixes are, and how fast the main branch is moving along, but not to the point where constant interactively rebasing your branches is the norm.

If you're writing a local stack of commits, you need to do an interactive rebase to edit the commits in the middle. You can't just check out and amend earlier commits in the stack, since git has a branch-first rather than commit-first model (one of its biggest UI failures).

No, I don't. Again, this seems to be a personal design/preference rather than a behavior git is forcing upon you, which is my point (This isn't a git issue, this is a development workflow that seems needlessly over complicated). When I do development locally, I only commit changes I want to keep. If there is something wrong, or code that needs to be modified, I make those changes, and then commit. When the work is done, if pressed upon me by the team, I rebase the local commits to a single commit before submitting a merge request.

It's the same development workflow used by the Linux kernel? You know? The project that Git was invented for?

Are you working on something as complex as the Linux kernel?

There are tons of different workflows used by Linux Kernel maintainers. Some of them don’t even use Git except to submit patches.

Whose workflow are you trying to emulate? (followup question: and why?)

Finally someone comes out and says it - I totally agree. I don't recall ever having to have a "tutorial" on Subversion. Git solves a problem the average dev really doesn't have in the first place.

It was painfully difficult to branch in subversion. And remember the agony of asking someone to check in a file so another developer can check it out?

Branching was just a copy? What was so hard about that?

As for check-out locks. you only had to ask them to do that if they'd requested a lock on the file which was not the default. I hope people weren't doing that too often at your workplace.

To be fair, Subversion had plenty of serious limitations. Local commits are a really good thing. It's just that Git itself is a terrible UI around a bunch of decent ideas.

Sure. Absolutely but there were some benefits to its simple model that are still attractive for a corporate env.

https://www.bitquabit.com/post/unorthodocs-abandon-your-dvcs...

One of my fav rants on this subject.

Merging is an operation which takes three inputs: left, right, and common ancestor. This is a very old algorithm, the diff3 utility on Unix is at least from the 70s I think? The reason history is a graph is to walk the graph and find common ancestors. Subversion's model cannot do this (screw off, pedants, I know they added an attempt at detecting common ancestors after Git had long since supplanted Subversion in relevance).

And if you can't merge, you can't branch. And if you can't branch, you need locks.

I'm just saying, as someone who has used subversion quite a lot over the years, the situation he is describing where you have an exclusive lock where they have to check in before you can, does not happen. Exclusive locks are very rarely used in subversion. If you and your coworker are working on a file, and he checks in before you, you will simply have to update and merge. Similar to if you were both working on a local git or mercurial tree.

Oh, and in that situation the 3 way merge is between the version he checked into the repo, your local version you are trying to commit, and your parent revision.

Well put! And it demonstrates why people (like the HN'ers who think they're 10xers) have trouble with git. They fundamentally don't understand the problem and how git solves it.

This is correct, and yet Git is an embarrassment to our profession. That's how bad it is.

Mercurial should have won, it's superior in every way. For pesonal projects I use Fossil, which is fantastic.

I organized the migration from CVS -> Mercurial at my last job, and finally after several years when it became clear Git was where industry was headed we switched from Mercurial -> Git. After resisting it for a long time and then being forced to use it for a collaborative project with another firm, I have to say git is way better at being a DVCS than Mercurial.

I think, as other have mentioned, that once you truly "grok" git you begin to understand the elegance and simplicity of the system and how its layered. Ultimately git's use of branch pointers was a way better solution than encoding branch data in the commit itself. To it's credit Mercurial did attempt to rectify this later with bookmarks, but ultimately DVCSs are about branch management, and git has always excelled at it.

Mercurial's branch management makes more sense. If you want to know which branch a commit originated from in git, good luck.

Many things in git, are frankly anti-patterns in any other sane versioning system - comitting partial changesets because of the index? gross. Rewriting history and force pushing? absolutely insane.

The fact that I don't have to "grok" mercurial like you need to git is actually a plus. Mercurial can do everything that git does (and more: check out changeset evolution in mercurial, it blows anything you can do with git rebase out of the water) and it doesn't require Ph.D. level knowledge in directed acyclic graphs to do it either.

> Mercurial's branch management makes more sense. If you want to know which branch a commit originated from in git, good luck.

`git merge --no-ff former-branch` by default includes "former-branch" in the auto-generated merge commit name. When it matters what branch a commit "originated" on, it's very easy to use `--no-ff` (which doesn't rewrite history and won't need any sort of force push!) and most PR tools used to default to that before "rebase" and "squash merge" became some people's favorites because of some weird (in some cases OCD) wish to see SVN-like straight lines only in their commit history.

It isn't a "weird ... wish", a linear history is the correct default. Merges have always been, and should continue to be, a tool of last resort. In general it is good to minimize complexity in life.

Merges are the default state of life in any source control system with more than one person or machine involved. They are a regular, required need. Sweeping them under the rug, pretending that they don't happen, acting like they are rare is exactly the sorts of things that added complexity to systems such as CVS, SVN, TFS, Perforce and much more. It is exactly what adds complexity to so many of these rebase-only and squash-only workflows people like to constantly post about using in git.

Linear history is a punchline that ignores the reality of source control. The only truly "linear" history I'm aware of in source control is systems built on patch theory/operational transforms/CRDTs and that form of linear is subject to reordering and still needs explicit merge commits with (implicit) multi-commit dependencies even if there are few stated needs to visualize that as a DAG.

I agree with you, in general it is good to minimize complexity in life, which I why I eliminate rebase workflows and squash merges from mine. I prefer to see the evidence of where merges happened (and they should happen early and often), and graph data structures are very familiar to me as a software developer, including traversing them. (And git offers useful traversal tools such as --first-parent for when you want to view the DAG more like a linear history and ignore some of the noise. There's no reason to artificially simplify the DAG when DAG traversal is easy and accurate.)

Reply on news.ycombinator.com