$ hg clone https://bitbucket.org/eigen/eigen/
$ cd eigen
$ time hg grep CUDA > /dev/null
real 0m16.661s
user 0m16.097s
sys 0m0.531s
$ git clone https://github.com/RLovelett/eigen.git
$ cd eigen
$ time git grep CUDA > /dev/null
real 0m0.019s
user 0m0.035s
sys 0m0.057s
Never looked back. $ time hg grep --rev tip CUDA > /dev/null
real 0m0.075s
user 0m0.055s
sys 0m0.020s
Much better. Still much slower than git, which doesn't matter for this small repo but would definitely matter on a larger repo. But better. And hg lets you define aliases, so you don't have to remember to type "--rev tip".* Network effect. Almost any dev can use git, so it's our "common language". * Staging area. Mercurial doesn't have one, which I find is enough to make me never look back.
$ time taskset -c 2 hg grep CUDA >/dev/null
taskset -c 2 hg grep CUDA > /dev/null 25.04s user 0.46s system 99% cpu 25.514 total
$ time taskset -c 2 sh -c 'git grep --cached -l CUDA | xargs -n 1 git blame >/dev/null'
taskset -c 2 sh -c 4.85s user 0.13s system 97% cpu 5.097 totalWhy? When I do hg grep, it doesn't show me blame or anything; the output looks almost identical to the output of git grep. If it's doing work and then throwing it away, why is it fair to force git to do the same work and then also throw it away?
(I'll also point out that git is still 5x faster in this comparison. :)
First of all, if I just cloned the repos, they should be in disk cache. My machine has 64GB of RAM.
Second of all, if they're not in disk cache, what on earth is hg doing that it takes 16 seconds to read them all in? Eigen is not that big, and I have a fast SSD.
Third of all, you can run the tests yourself -- I gave you the exact commands to try. "I tried it and my results are X" is a much better contribution to the discussion.
Fourth of all, yes, I ran the benchmarks multiple times in a row to check whether disk cache was at play. It was not. :)
Sorry, the speedup was very great so it almost seemed like something else was causing it. Thanks for the benchmarks.
> Third of all, you can run the tests yourself -- I gave you the exact commands to try. "I tried it and my results are X" is a much better contribution to the discussion.
That's true; I'm lazy.
Actually, I was thinking more of the disk cache of loading hg, not the repo (Python programs tend to have a great speedup through the disk cache).
I do agree that reproducible differences in performance should be measured, not single tests.
(End Sarcasm).
Seriously? You're so productive that the cost of cloning a repository is a major impediment?
The benefits of a cleaner interface mostly trump the benefits of speed.
Read the commands more carefully. It is not the cost of cloning the repository, but rather the cost of searching it.
16s is not a big deal, but I'm assuming it would get much worse for larger repositories.
Personally, my searches are either:
1. Restricted to all files that end in cpp, h, py, pl, etc.
2. Explicitly exclude .hg directories.
I'll confess: I use a nice tool to do my searches than typing it out in grep. Much handier. And it remembers previously used filters (.cpp,.h) and is usually the first item in the drop down box. Much quicker than typing out a grep command.
Still, I feel if our team used git instead of hg, we'd lose a lot more time due to people not grokking git and making mistakes than the time lost in longer searches.
My point still stands: Things like search time are not a productivity bottleneck. If all our searches were 100x faster, there would be no difference in our output.
And I disagree with this point:
Things like search time are *not* a productivity bottleneck.
I maintain old codebases. I don't know them by heart, I didn't write them and the documentation is often poor (for the codebase, the specs are usually decent, but that's the system spec, not the code spec). Search is critical for me to get around these things. A 16s search would be incredibly annoying and bottlenecking compared to one that's nearly instant.Not having looked at the internals of either, I wonder if this is related to other performance differences in grep implementations across *nix systems.
EDIT: Exploring this, it seems they don't do the same thing? `git grep` searches, by default, only the most recent working tree or revision. It doesn't search every revision unless you tell it to. `hg grep` (don't have it installed) seems to search every revision by default, so restricting it to the current revision requires an explicit command line option specifying the revision range to search.
Things like search time are *not* a productivity bottleneck.
I would recommend, as a general rule, not generalizing from your development experience to everyone's development experiences.I happen to know that git grep is very carefully optimized. Someone cared a lot about making it run fast. Seems like a good guess that they did so because it was important to them or someone else, not just for funzies.
$ grep '^git grep' ~/.bash_history | wc -l
407
$ grep -w '^ls' ~/.bash_history | wc -l
37
$ wc -l ~/.bash_history
19421
"git grep" is 2% of all the commands I type. I type it 10x as often as "ls". So, yeah, its speed matters to me. :)I appreciate that - I know the frustration of tools being a bit laggy.
However, I think you are falling for the [Endowment Effect](https://en.wikipedia.org/wiki/Endowment_effect). You are accustomed to it, so the alternative seems horrible without you really thinking it through. If you were not already using grep, the search issue would not be a big deal.
I already pointed out how merely excluding ".hg" from your search will save all the time, and this benefit you speak of disappears (please correct me if I'm wrong).
On my shell, that's as simple as creating an alias for grep to always exclude ".hg".
Yes, I'll grant that searching Mercurial repositories is a lot slower. But you need to understand that the obvious solution is not "Switch to git". There are simpler solutions.
In that sense, this seems to be a very insignificant benefit of git, even for people like you who need to search often.
To beat a dead horse, if you had two developers: One using git and the other using mercurial, and both need to search heavily, the mercurial person will likely have already solved the problem in one way or another. It'd be a bit silly for an observer to look at them and say "Ah, Git is so much better because it did not require a shell alias for faster searches!"
What's interesting here is that it's the exact thing happening. Someone asked about how much hidden cost there is from the time people waste fighting to make git do what they need it to do, and it got derailed with a discussion of how `git grep` is faster than `hg` grep`.
I'm sure it's great for you that git's grep is fast. But the time you gain from how much faster it is than hg probably is strongly outweighed by the time other people lose to struggling with git's interface. Unfortunately there's no easy way to benchmark that.
Mercurial has a clean and simple UI, especially if you come from a SVN background.
But Git's internals are simpler and the concept is much easier to understand, especially if you start looking at a handful of the most popular Mercurial extensions, some of which will be "must have" sooner or later.
This is no excuse for an inconsistent UI in Git, but the damage is already done. There's no point in making major changes now (some renaming for consistency might be alright, though), it would be breaking existing workflows and scripts.
I find the best explanation of how Git works is... the first commit of Git[0] itself.
[0] https://github.com/git/git/commit/e83c5163316f89bfbde7d9ab23...
I'm only an occasional Git user, but am a heavy Mercurial user.
I have never needed to understand Mercurial's internals.
Are you suggesting a Git user will eventually need to understand it? That's a strike against Git already.
From my observation over the years, people who use Git get "stuck" more often, because they are trying to leverage the extra power Git seems to provide. With Mercurial, you often can do similar powerful stuff, but they make sure you have to go through hoops to do it. As such, unsuspecting non-power users do not shoot themselves in the foot.
The typical advice -- which you'll see in this thread and elsewhere -- is that anyone who uses git should learn git's internals, and preferably up-front, since apparently that will magically cause git to make sense.
I'm a bit put off by the idea, myself.
It's also worth looking at git internals because it's very elegant, clean and simple. In my opinion, one of the best designed software systems I've come across.
I encourage you to look at the README in the first git commit, it's very enlightening.
I think as engineers we should understand how our tools work.
I try to learn about git internals just for the pleasure of it.
There are A LOT of mercurial users which are big companies, they just aren't so vocal on the internal like people using GIT. Mercurial has an established community, great plugability system, an awesome feature like phases.
We see more and more people adopting Mercurial together with our product.
I wish to see more advocates of Mercurial, it's a shame it doesn't get more public attention nowadays.
The problem, that I don't think anybody even expected, is that there are strong network effects on the selection of your CVS. That gives the tool that can handle more use cases a big advantage, and bigger still if those features are needed on the projects with more developers.
Certainly, it's got a lot of different switches and stuff, but since its data model is so simple, you can always understand what it's doing to your data. So you can read a description of what a switch does, and then understand exactly what will happen when you use it.
(The exception to this is automatic merge conflict resolution with rerere, which is just magic to me. But it's a bonus feature anyway)
For me, as a long time git user, git's UI and UX is approximately a bajillion times better than something like SVN, which can barely even do half the things that I use every day in git.
Maybe I haven't just dug deep enough into Subversion's manual, but can you tell me how to do the equivalent of "git log -p"? Last time I checked (SVN 1.8) that's not very trivial. Or how to revert a commit? IIRC that requires a merge.
So at least compared to subversion, git enables a workflow that was previously too much work, and gives me a better tool for working with version-controlled text data.
Mercurial is another thing I find puzzling, and I'm honestly interested why it is that people find it so much easier to use than git. I'm so used to working with throw-away branches that hg's default branching model just feels needlessly difficult to me, and good support for managing branches is the most important feature of a VCS, since any change you ever make is automatically a branch from the base. In git, this feels natural, since branches are a feature of the underlying data model, not something treated specially.
svn log --diff, added in 1.9. Also doable with a little (est. < 10 lines) of shell scripting.
> Or how to revert a commit? IIRC that requires a merge.
git and svn disagree on whether history should be permanent or not.
Non-distributed VCS clearly don't cut it for many use cases, and of the other DVCS I've tried, none were really better than Git in my opinion.
Sure, it depends on your workflow. But if part of your workflow is authoring sequences of commits that make sense in hindsight, then I honestly haven't seen an alternative.
The point of Git is that it gets the underlying data model right, and just exposes that. Once you grok that, it is superior to everything else.
If you don't grok it, you have a problem. For this reason, I think GUIs are actually somewhat detrimental for Git use. git gui and gitk serve a purpose, and they could definitely need some polish, but my other explorations into Git GUIs have basically turned out to be badly done and leaky abstractions which hurt more than they help.
(Not to mention the fact that Perforce as I know it simply doesn't cover the distributed use cases that Git does. Obviously adding distributed features makes any VCS more complex, but that is the inherent complexity of the problem space as opposed to artificial complexity. If your problem is conceptually easier, you can make the solution appear easier.)
I would like to understand what's the yearly damage of such an important tool being so difficult to use. People committing the wrong stuff, unmergeable messes, people not being able to correct their mistakes, there must be thousands of Git rookies fucking up their Git repo or throwing away their work just as I am writing this.
What would be the cost? Millions of Dollars? Perhaps even billions?
It's about as bad as 0 being a legit value for a regular pointer.