Hitting every branch on the way down
rachelbythebay.com
rachelbythebay.com
-- std::string left = "\"";
-- std::string right = "\"";
-- if (use_system_include) {
-- left = "<";
-- right = ">";
-- }
-- return left + name + right;
++ return absl::StrCat("\"", basename, "\"");So if the chromium repo is representative of the state of C++ in rest of Google, they ditched it silently like this probably because it's so natural to them :)
> <> means system headers and "" means project headers (my naive understanding of the difference)
they just convert everything to a project header? That seems bonkers to me. It is intentionally removing useful information.
They also are not removing any info -- most of it IS project headers. To me that's the actual bonkers bit :)
With GCC/Clang, headers found in paths passed with -isystem are headers that are immune to compiler arguments like -Werror because they are, by definition, out of your control. In Google's case, ALL code is already checked in the project repo, including language stdlib. So none of them are system headers "per se".
I was curious about this, so I took a look at the list of protobuf releases[0] and they're...confusing, to say the least. Chronologically, the most recent tags at the time I write this comment are:
v5.27.0-rc1 v3.27.0-rc1 v27.0-rc1 v27-dev v26.1 v5.26.1 v3.26.1 v26.1 v26.0 v5.26.0 v3.26.0
Does anyone here happen to know what's going on here? As best I can tell, they're simultaneously supporting 3 major versions while keeping their minor and patch versions in lockstep, and then having one major version be implicit?
I can almost imagine a scenario where they started out with just major and minor version and then realized they wanted to make breaking changes, which led them down the path of adding a third separate number to the versions, but if they already were going down the path of assuming "wider" versions are newer , why not just stop using version numbers with only major-minor and instead just add a 0 or 1 to the front of all of the continuations of that branch? Also, why synchronize every single minor and patch version between all three major versions? I can understand why it might be useful to continue providing support for multiple major versions at the same time, but I'd expect that _sometimes_ there might be a bug or something in only one of them, and pushing out a release of the other two that don't contain any changes would be pretty strange.
But I suppose old version still receive bugfixes.
Perhaps someone wrote a library for that…
IIRC (5 yes ago), it was somewhat self describing with respect to struct offsets but had no (data format) versioning or type information. Never understood that design choice, even in mixed endian+architecture environments.
Ends up being horribly inefficient for small/one-time instances, and not all that great for client/server use.
We were looking at it to get away from C structs and handmade serialization — only to unexpectedly realize that our methods were still better. (We genuinely did not want this outcome!). JSON/yaml were not options for other reasons.
Protobuf is the modern realization of “the king has no clothes”.
But the fact that a merge can have arbitrary changes in it always bothers me!
This is a case for rebase over merge if there are conflicts.
You could have a merge of 2 empty repo parents where the result is the complete source of the latest version of Kubernetes!
Rebase is simple. Always. The end result is obvious and clear and can only be interpreted in one way.
Merge has lots of little sharp edges and surprises if you don't know every single tiniest detail.
Almost nobody knows it in that level of detail, so it's a terrible choice for interacting with anyone else. If you're on your own, sure, do whatever - many things are not built solo though.
With `git amend` and `git fixup` you can arrange your commits to be clean, properly documented and self explanatory (and maybe atomic but that's a little harder). It takes a little time but it is hugely beneficial to code reviews and bug investigation.
(Our workplace has a mix of github flow, which is one branch per PR: https://docs.github.com/en/get-started/using-github/github-f... ; Atlassian Gitflow https://www.atlassian.com/git/tutorials/comparing-workflows/... ; and the completely different Gerrit flow which ends up very rebase and amend heavy: https://gerrit-review.googlesource.com/Documentation/intro-g... )
Nobody should be able to force-push to master (or any public branch) except on specific occasion. In that case, someone is authorized, performs their specific action and then get de-authorized.
This is pretty basic.
Usually when using the commands directly they're more careful, but have the mindset "an IDE wouldn't intentionally break something so this button must be safe to click".
IDEs in particular should expose the details all the time, and show the command that's being run, so it can teach people passively instead of misleading them and leaving them stuck when it breaks.
making commits atomic is harder because we tend to just write code, without first breaking up the requirement into atomic pieces
Don't push half-baked work on other people! You waste their compute cycles needlessly, from now until the end of time.
I.e., git can retain two representations of a sequence of commits: the original sequence, and also a larger commit that (a) produces the exact same code change as the sequence, and (b) has its own commit message.
For viewing history you can use "git log --first-parent" to view only those merge commits, which should satisfy the people who love a linear history, without actually losing the intermediate commits.
Then I remember that I have enough trouble getting a few dozen people together to write well formed and understandable commit messages for one level of commit messages alone. This scheme would require people to extend more energy on constructing commits which is at best something very few care about.
Then there are tickets and other corresponding information, but they could rot for all I care, as they so often do, unless a decent commit log is in place.
(I'm surprised this got a downvote when that's how we got here: a situation in which a change was ""hidden"" in a merge commit that would have been explicit in a rebase workflow)
That can often help a lot to figure out why something ended up the way it did, but you also don't turn your entire history into a flat pile of disparate commits.
It's definitely how I prefer to review code (big changes broken up to isolated portions that are easier to validate, and the whole thing at once so you don't get lost in the trees), so it's how I would prefer to read it later too.
It does still have merge commits where stuff can hide though :/ and you've got to remember --first-parent :/ and all not-merge-focused things so have problems with it :/
So for example if you #include <foo> then the compiler (actually the preprocessor, but whatever) looks in the system's standard location, whereas if you #include "foo" then it looks in the local tree.
I am posting the summary here, although please do read the original if you have time:
The most portable thing to do is to use "" for including files within your project, and to use <> only for implementation supplied files.
(Disclosure: I was one of the contributors to the C FAQ).
Anyway, I thought the distinction was lost in later compilers in favor of a single include path and then just taking the first file found when looking at potential matches through include path.
It seems the author of that merge thought the same thing. So, the distinction is actually still used by compilers?
---
P.S. Any commit can have any change. Or no change.
A "commit" is a version...a message, a tree, some metadata, and 0 or more parents. In fact it's not even a change/diff/patchset per se. Though will often compare it against its assigned parents. If it has multiple parents, you'd have to choose which to compare. If it has zero parents, you can't compare against any parents.
Git log only shows one tree not parallel trees from the merge.
git log will show all ancestors.
And git diff shows any difference between two refs.
Nothing invisible unless you deliberately make it so.
It's only partially the fault of Git - the entire idea of a merge requires new concepts like 3-way diff, which are not needed for rebased commits. I'm not even sure that most software like GitHub can display such a diff.
I hate rebase on shared code I don't care how clean jt looks. Don't mess with history.
git reset --hard origin/branch-name> git pull --rebase
git config --global pull.rebase=trueYour two feature branches would be unprotected so you can merge away if you like. When one of you wants to commit something to master, that's when you'd check for dodgy merges.
Also, "git cherry-pick" is a good alternative to merging for this use case.
No, the feature branches need to be protected or something, to enforce that they only rebase locally and don't rebase the parts that I've merged into my branch (and vice versa).
> Also, "git cherry-pick" is a good alternative to merging for this use case.
No it isn't, it means you get multiple unrelated commits for the same change, which causes conflicts and can be disastrous if a commit is deliberately reverted.
- Having a developer on your team that rebases their own feature branch
- Then tries to "git push", only for it to be rejected since a force push is required
- Then performs a "git push --force", which will force-push all of their local branches, including feature branches from other developers that they may have checked out previously
Our team uses merges because they are safe from this kind of problem, although a rebase workflow would have cleaner history. I wish that "git push --force" would not push all branches by default, and just fail unless a (remote, branch) pair or --all is given.
This is (part of) why, for most common operations, I use a Git GUI (SourceTree). Force pushing all branches can only be done by very explicitly selecting them all and initiating a force push; the default when pushing is to push only the currently active branch.
It's also overall much clearer and more intuitive to use than the Git CLI. I use it when I have to—there are things that I can't do through SourceTree, and a few things that are complicated enough that I just want to be 100% sure I know exactly what's happening—but for 99% of the Git operations I do, it handles them perfectly and without any worry that I've mistyped something or forgotten to specify a branch.
And only on working branches. I do this every single day.
To collaborate more closely and reduce (or get ahead of) conflicts. The whole point of using git at all is to be able to base your work off other people's in-progress work; if you're not interested in doing that then Subversion works better.
Assuming "H" is the hash of the current state of the repository content, consider this initial state of the repository (most recent first):
H(3) Implement feature B
H(2) Implement feature A
H(1) Initial commit
H(5) Shiny feature, improvements.
H(4) Shiny feature, initial implementation.
H(3) Implement feature B
H(2) Implement feature A
H(1) Initial commit
Then you `git pull --rebase`, and your history looks like this:
H(10) Shiny feature, improvements.
H(9) Shiny feature, initial implementation.
H(8) Pulled commit C
H(7) Pulled commit B
H(6) Pulled commit A
H(3) Implement feature B
H(2) Implement feature A
H(1) Initial commit
With the usual pull request flows, `H(9)` (i.e. anything between your new "base" and your most recent commit) usually stays untested, entirely ignored by the developers, and you would only ever find out if you ever need to bisect.
Not usually a problem, unless you have a rule of "every commit should be verified/tested" and the untested commits have a change that doesn't prevent a build but still causes issues (e.g. something that's only visual, or a new config file was added to a "conf.d" directory and its presence changed some behavior, stuff like that).
Everyone else that had a copy of the repo would have had a copy of the "lost" commits.
I really cannot imagine how many things would have to go wrong for weeks of work to be lost.
- Person who destroys git history
- Person who hates destroying git history
- Person who knows how to recover "destroyed" history
- Person who knows how to truly destroy git history
… tell me more!
Stuff "leaks" so much in git, that it's really hard to lose work. The only way I could see someone losing work is if they never commit or if they never push. But even if you don't push and just rebase, you're not losing work. You would have to go out of your way to delete git history locally.
You just need to find an old commit ID somewhere, normally the reflog.
The old stuff will go away on its own eventually due to git's self-maintenance procedures removing unreachable commits, or it can be done forcefully by adjusting the gc parameters to get rid of it.
The Gitsatz Haderach
So we released at the end of our cycle. It gets used for, if a vague memory serves, end of months billing. Meanwhile someone on another team "merged" our code, and actually randomly dropped a big chunk of our work. A week later they release. Some (but not all) of our features disappear. We pull that and don't notice because we're on a new sprint. Wait until the end of the month, users try to do billing. "Hey, why did you take away those features you built for us a month ago?" "What, we never...?"
We had no clue what happened.
This got to repeat a couple of times before we figured out what must be happening. We made changes to the release process so we could track what was actually released each time, with its history. We tracked down who we thought was making the mistake, but didn't have enough evidence to prove it to his manager. That didn't stop the idiot from making the mistake, but it did streamline the process of recovering it. Meaning we had the version with our feature, we had current code, and "just" had to sort out conflicts rather than rewrite from scratch.
Now that you've heard the story, can you see how weeks of work could be lost before we figured it out? And can you understand how we could have lost history?
This was a decade ago. At my next job we had more competent people. But there we had a huge debates between rebase and merge people. There are arguments on both sides. My conclusion was that about 90% of the time, rebase makes things simpler and easier. But that remaining 10% of the time makes the 90% not worth it.
Just learn how to merge properly.
What you're describing does sound awful, but I'm pretty sure that idiot could have found a way to mess up a merge. The entire workflow sounds completely fucked, I'm not convinced it's entirely fair to blame rebase in that case.
> Just learn how to merge properly.
I know how to merge and rebase properly. My favorite PR merge strategy is rebase + merge --no-ff. So your master branch is nice and linear, but you can still see where your PR merges came in. Let's you have a "all PRs get squashed" view of the world by just adding '--first-parent' to your git commands, but also lets you have the inner details for when you're git bisecting or spelunking trying to figure out why a certain line exists.
Most people hate what I describe though, similar to mixing spaces and tabs.
I agree that the workflow was a mess in multiple ways. A lot of which were organizational decisions that I was in no position to influence.
Your favorite PR strategy is fine if you're doing it locally. However when it is done on master, you're going to have to get master again by force. Because changed history creates conflicts. Which means that you're going to have to hope that everyone only did it your way, and no idiot created conflicts in some other stupid way that you'll suffer for later.
I'd prefer to merge to head early. Merge to head often. Merge from head often. Don't have long-running shared branches. This does take some other forms of discipline though.
Master is always locked down anyway by "something" - no idea what the technical term for Github/Gitlab/Bitbucket is. Stopping people from force pushing to master prevents the sort of stuff that happened to you. Even if you don't have any "idiots", you really don't want a poor intern accidentally slightly pissing off everyone.
> I'd prefer to merge to head early. Merge to head often. Merge from head often. Don't have long-running shared branches. This does take some other forms of discipline though.
I agree with everything there, except I rebase instead of merge. So when I merge my branch to master, it's a nice neat little package that sits on top of master. It doesn't have the history of 10 merges I did while I was developing because I don't see the value in those merges.
But hey, to each their own. When I was younger, I used to get into heated debates about why I was right, now I don't really care. I'm either in a branch of my own and can do whatever I want, or working with someone and then I'll just copy whatever they do to not confuse them.
Rebase is a necessary part of a workflow even if you like merge commits. You're severely missing out if interactive rebases are not part of your toolbox.
This is exactly what backups are for.
That one's obvious, you can type <> and you can't type “”.
I have to admit that I never really understood the advantages of rebase, and what I mean by this is they I actually don’t understand how the dangers of rebase out-weighs any form of advantages. Especially because on of the major advantages of merge is that you can squash your local commit history when you submit it to your main branch.
What we do is that we tie every pull request to a relatively small feature task, and because we do this, we genuinely don’t care about the individual commits developers do. Which means they can commit really silly messages if they are heading to a meeting or if they are just tired at the end of the day. It also helps with them merging main into their branch often, because it doesn’t taint the history.
The biggest advantage we’ve seen, that maybe we didn’t expect, is that nobody ever fucks up our tree in a way that needs someone who actually understands git to solve. We’ve also locked down the use of force push so that is not available to anyone unless it’s absolutely needed. Part of the reason I set this up initially was to protect myself from me, but it’s been a good thing since.
But I’m actually curious if it’s wrong.
Squashing is in no-way limited to merging and is actually done by doing an interactive rebase. Nothing is stopping you from squashing without creating a merge commit. It's entirely separate.
If you're squashing everything anyway, what does merging even give you? Is your main branch just:
* merge B
* squashed commit B
* merge A
* squashed commit A
If you didn't merge, you'd have:
* squashed commit B
* squashed commit A
> What we do is that we tie every pull request to a relatively small feature task, and because we do this, we genuinely don’t care about the individual commits developers do.
Except eventually there is a large feature task and then you end up with a giant commit that is annoying when git-bisecting.
But at the end of the day, these things only matter if you are spelunking through git history and/or using things like git bisect. If your git history is "write-only & rollback", then none of this stuff matters.
No, it's the other way around. Squashing is a type of rebase.
Most workflows involve both. Merges can also be fast-forward merges, which are indistinguishable from rebases. Choosing between a rebase and a merge operation is often the wrong question to ask. The question is what state you wish the repository to end up in.
> I’m actually curious if it’s wrong
Look at "git log". It is readable and easy to understand? It is obvious why each commit was made, and why alternative solutions were turned down?
Are you able to use "git bisect" to track down problems?
Then you're doing it right. If not, think about what a functional commit log would look like and how you would get there. Working together is culture, and what type of merges you decide to use is just a tiny part of that culture.
A "merge commit" is nothing more than a commit claiming any number of parents greater than one. It is still its own file tree reference that decides how the tree looks, and nothing dictates that it should be related to the parents.
Multiple initial commits are a bit rarer, usually stemming from merging in entirely different git repos with their own separate history as part of consolidation.
Git is more than the data structure backing it. And many parts of git make all sorts of assumptions that treat things that are more or less identical in the data model as actually being different. Tags are not the same thing as branches, for example, even though they are stored in virtually the same way.
git log -p is a convenience tool that tries to show code progression, and so it comes up with these simple diffs. Showing a graph of N-way conflict resolutions would not help the user trying to get an overview. Other tools exist to track the origin of a particular change, such as git blame.
It is important to understand what a git commit actually is exactly because of the caveats of such convenience interpretations. Otherwise you'd have no idea where a conflict resolution is stored, and you'll run into the surprises mentioned here.
In my opinion, git also becomes a lot easier to work with once you understand the low-level details, as you realize which high-level tools are similar, compatible, or fit or unfit for a a specific purpose.
Commits are never diffs. Commits are snapshots, and sometimes git computes a diff between two commits. Commits are also never delta compressions, but can be stored within a delta-compressed packfile.
Whether you like it or not, git is primarily its low level details. The porcelain stacked on top changes, and differs depending on the user's client (e.g., a GUI using libgit2). However, that "git log -p" is "part of git" that git log -p is not trying to convince you that commits are diffs and show you a true chronicle. It instead assumes that you know what commits are, and that you are asking for an easy to read overview of what has been going on.
Accepting that commits are always solely snapshots will make the issues you run into when working with the porcelain easier to understand, especially when exposed to more than one client.
(Knowing about packfiles and delta compression can also be useful when looking into performance/resource utilization.)
After that xy thing where they were trying to install a back door having changes that are hidden like this is a big red flag.
In fact changing include <something.h> to include "something.h" with a hidden commit like this isn't a red flag it's a big rotating alarm with a siren. Someones trying set things up to include malicious code via a faked system lib.
$ git show d85c9944c55fb38f4eae149979a0f680ea125ecb | wc -l
11067
$
Presumably the author would have been happier using the -m-flag in addition to -p.
The approach described on that site doesn't strictly rule out "git merge", but it emphasises short-lived branches and unidirectional commit flow. If you do things that way you find you just don't really need merges. The next step is to think "merges are rarely useful and sometimes dangerous, so let's just avoid them completely".
Gotta say, I find that horrifying. What about peer reviews? What about breaking up a change into smaller commits, none of which make sense until they're all together (changing the signature of a method, then changing the places that call that method, etc)?
It's worth noting that is mentions that work on a branch and the use of PRs are acceptable, but... the two statements appear to contradict each other. Why say "only do x" and "do thing that isn't x" on the same page?
> What about breaking up a change into smaller commits, none of which make sense until they're all together (changing the signature of a method, then changing the places that call that method, etc)?
The general-form solution to this is to make a three-part change: 1. add the new code, 2. migrate all the callers, 3. delete the old code in three separate PRs (or more, if migrating the callers takes several PRs, or some of the old code can be deleted earlier than the rest). I believe arbitrarily large changes can be made this way, and as your origination grows, eventually all large changes _have_ to be made this way.
Isn’t that a lot of extra work? IME it’s a lot less work (and risk!) than resolving a massive merge conflict.
The problems with long-lived branches all derive from the basic problem that eventually the complexity of maintaining two parallel implementations affects the work of everyone at the company.
New person joins before `Big_Refactoring` is merged? You’re either onboarding them twice into two branches, or they’re getting nothing done while they wait for the merge so they don’t have to learn a bunch of code that’s going away soon anyway.
Someone else wants to make a significant change? They either carefully patch each PR into both branches, or they decide “screw it” and do all their work in `Big_Refactoring` anyway, diverging the branches _even more_ and creating more risk in the ultimate merge and more of an incentive for others to start developing in `Big_Refactoring` and make the problem even worse. Soon the feature branch is a de facto main with failing tests while there’s incredible pressure to just ram the merge through so all these changes can go out.
The only way to make it work is to demand that only one person develop in `Big_Refactoring` and everybody else manually cherry-pick their changes into both branches (which quickly just means implementing them twice). IME everyone finds this so annoying that small branches, feature flags, and three-part changes (which makes code sharing between the old and new implementations much easier) become broadly preferred anyway.
As far as PRs, I can’t speak to the linked article, but everywhere I’ve worked that implemented short-lived branches still did PR review. But the PRs had to be small and quick to review (which IME actually helps catch bugs too)
This seems to imply that there's a choice between A) committing to the main branch, and B) long lived branches. I always work on branches, almost always with multiple commits, and then merge it into the main development branch once the feature is complete. I almost never have to deal with complicated merge conflicts.
That being said, you're talking about short lived branches. The article talked about committing directly in the main trunk/master branch; which is what horrified me.
i use rebase frequently, but i never remember which direction the operation goes in. do you need to checkout the source branch or target branch? truly, it is unknowable. my workflow is to type `man git rebase` and hit space to page through the manual until the first ascii tree surgery diagram appears. then i stare at it until i remember that i need to have checked out my feature branch and am meant to type `git rebase main`. i have trained myself to read the man page every time, perform the operation correctly, then immediately forget.
It's essentially a "cp thing-i-want ."
Combined with git reflog your repo becomes as understandable as a floppy disk.
When you make a merge commit, the merge commit contains all the changes. It's what happens when you fix a merge conflict. The fix for the conflict only exists in the merge commit. Similarly you can just add whatever you want in the merge, and it won't appear in any other commit.
They obviously changed their version numbering convention at some point, and this is protobuf 2.4.something.
Honestly, I see the shit she tripped on all the time. It doesn't even register anymore.
It wasn’t until I just read her article that I’d even considered some systems/distros doing weird things like rewriting C include syntax for questionable reasons.
What a terrible thing to deal with, simply frustrating.
I don't know how "common" it is, but when dealing with third party code, I run across <> / "" confusion quite commonly, so when that's a significant part of your job, you'll probably stumble upon this flag eventually.
> There's no "body" to this commit. It's just a "Merge:" and two other commits
Commits are snapshots of repository state, and merges "obviously" have differences from its parents. So not having "body" for a commit is bit nonsensical in git (yes, technically you can make empty commits but that's a special case). These sort of things are where having good mental model of git is useful.
As I have my share of hairy merges, it is pretty intuitive that merge commits can, and in many cases need to, have changes that are not part of either parent.
Maybe something like pijul (/darcs) would handle things differently here, but I believe that merges are fundamentally difficult problem.
It actually escapes me why Linus decided to make git a merge-first VCS in the first place. There aren't many projects which are more linear in nature than Linux kernel.
This isn't a problem with the author or her mental model, it is a problem with `git log -p`. The output she is describing is exactly how merge commits show up there, with no other flags.