Having gone through the same process, for a small company, that "one extra level" is the key. Now I can check my code in early and often without worrying about stepping on anybody else's code. That encourages me to use SCM even more.
You can do the same thing with a personal branch in SVN. There was sort of a competing line of thought that keeping it local is a bit of "cave-hiding" behavior.
http://blog.red-bean.com/sussman/?p=20
While there may be some equivalent in SVN, it gives you nowhere near the flexibility. I can't for example, create an idea branch for one feature I'm working on, try another idea on another branch, delete the first branch, switch back to the second, create another branch off of that and stop midway to checkout master (trunk) to work on something urgent and the then continue development and merge back into master when I'm done.
The equivalent in svn would make me groan, for complexity and for the performance.
Also, the same rules apply when 'cave-hiding'. You should still be committing early and often, but in gitland, 'committing' really means syncing up with an authoritative repo. You can 'cave-hide' in svn by just not committing anything and keeping your changes in your local working copy.
Agree on 'cave-hiding' being just as possible in svn as git, but I disagree that your workflow is particularly complicated or poorly performing in svn.
Creating branches in svn is cheap, switching branches is cheap, checking out a new working copy if you want to work on two changes at once is also cheap, merging is cheap. There's no reason not to do what you're suggesting in svn.
One of the things that git does better than svn is merging branches when files have moved, but I don't think your use case addresses this directly. Git does some things better than svn, but svn does a good enough job that you have to be a little careful when making the case against it.
>Creating branches in svn is cheap, switching branches is cheap, checking out a new working copy if you want to work on two changes at once is also cheap, merging is cheap.
This is just misinformation. Merging is absolutely not cheap, how incredibly expensive it is in SVN is the reason some of the DVCSes of today even exist. Try doing some of the common Git work flows [1] in SVN and let us know how well that works out. I've done it. Once. After that we adopted a methodology to ensure we'd never have to do it again.
[1] Pretty much anything where you are moving change sets between multiple branches before pushing back to main. Trivial in a DVCS, a nightmare in SVN.
> Creating branches in svn is cheap, switching branches is cheap, checking out a new working copy if you want to work on two changes at once is also cheap, merging is cheap.
I think then the problem is our definitions of 'cheap'.
A typical switch command on a working copy of our codebase at work (20,000 + files) takes 1-2 minutes, and its a lot longer for creating a new branch. In a git version of the same repository both these operations are comparatively instantaneous. YMMV.
Also, regarding merging, I believe that in new versions of the svn client there is mergeinfo, however older repositories without this metadata will give you issues when you for example create a branch, merge changes from trunk into your branch and then subsequently attempt to merge that branch back into trunk.
This is the classic 'svn sucks at merging' problem and to me, its complexity overhead that would deter me from branching, i.e. not 'cheap'. I normally get around this by remembering the revision number of the last 'rebase' from trunk (or give it an easily greppable commit message) and then when I do the merge back into trunk I specify that I only want changes from the last rebase to HEAD.
The internal model of git makes this a non-issue, i.e. each commit has a parent commit, and for any given commit on two separate branches, git just walks the tree up to the nearest common ancestor.
Any version from 1.5, which was released in June 2008, has merge tracking. If you're using svn and don't understand how mergeinfo works, then you're not in a position to make a fair comparison on merging. A lot of the comparisons between svn and git/hg/etc. are made by people who don't seem to understand recent versions of svn well.
Yes, old versions of svn "suck at merging", but the workflows with merges between multiple branches you describe work well in modern versions of svn with merge tracking. Your "remembering the revision number of the last 'rebase'" is unnecessary in modern svn.
I grant you that switch can take a while on a large code base, as can a checkout of a new working copy. Creating a new branch is contrary to what you say instantaneous in svn due to the cheap copies, however getting that new branch in a new working copy to work on I admit might take a while. A large part of this problem is mitigated if you just have multiple working copies checked out, but not all.
The details on why branching is cheap in svn are described here:
Part of the problem with svn is that the model is too flexible compared to git: the merge tracking supports tracking partial merges between branches, which is arguably an unnecessary complexity - the svn developers have expressed concern about whether this feature is valuable given the complexity it adds.
For general merge scenarios if you're using mergeinfo correctly, follow the practices laid out in the svn book, you shouldn't find merging particularly difficult or expensive in svn, even if you're doing relatively complicated workflows.
I'm not saying this is easier or works better than svn than git, just that it's doable and not as bad as many people make out. The merging in svn is (too) complex, but once you figure out how it works then it works well. With your "I believe...there is mergeinfo" and loewenskind's "I've done it. Once." I get the impression that neither of you have put the time in to really figure out how it works, but are still arguing against it.
Personally I prefer the way git works. But as someone who uses it everyday, for scenarios like the ones you say are nightmares, I don't think svn is as bad as you make out. There is an argument to be made for using git, but neither you, loewenskind nor the article make it as well as you could.
We’re self-hosting SVN on AWS…on our webdev server. Having all of our revision history and development code on one machine, especially one as potentially ethereal as AWS, is terrifying.
You might want to investigate how to to set up an EBS backed persistent root filesystem on AWS as well as how to do S3 snapshot backups. EC2 instances fail. EBS volumes fail. You need to be ready for that regardless of the revision control choices you make.
Is your backup culture really that sketchy? Your source code is probably just one of several digital assets the loss of which would cripple your company. If you don't consider an SVN repository reliable, it might be a good time to check your database backup procedures.
Also, SVN stores a local copy of the codebase as well. If you're working off trunk and the SVN server detonates, all you lose is the history and any branches you didn't have checked out.
We do use EBS, but AWS has a flag for "delete on terminate" where the EBS is automatically purged when the instance is terminated. Not 100% sure of the conditions for this to happen, but when you have a single boolean for something like that, and the only way to configure it is through their command line tools/API, and it's poorly documented, it makes me pretty uneasy.
Edit: We also have a script that takes a daily snapshot of all our EBSes, but we haven't tried to recover them (mostly due to lack of time to experiment).
I know, it does seem the default for "delete on terminate" is wrong. Most people use EBS root volumes in an attempt to gain persistence and a level of protection against accidental termination.
Nice story, but there's no need to categorize a switch from svn to git as optimization, especially as part of the phrase "premature optimization".
That refers to exchanging clarity or simplicity for performance. If a technique is fundamentally superior then it's just the way you should do things, and few things in tech are as unilaterally superior as git over svn.
In this case optimization refers to trying to pick the best option. PG talks about it a lot: we could spend hours trying to pick the most efficient desk layout, the best apartment for getting work done during YC, or a hundred other things. As a startup, the right thing to do in most cases is to shut the hell up, pick something, and get back to work. That was SVN for us.
But if you're really good you might not be willing to spend a big chunk of your day fighting with obsolete technologies. Unless the money makes up for it of course.
I'm not fighting when I use svn, and it's relevant to the client. They choosed it because it's simpler for them, and do all that is required perfectly.
I would totally push git (I pushed for git in some places if you wonder) if I thought it would be a good fit for them, but it's not.
In any team larger than 3 actively working on code, git is a far superior local client than svn. The speed of checkin, pushing, and merges are phenomenal. Even if you use svn as your backend server, git-svn is still a stupidly good local client to use and gives you a huge number of the benefits of using a pure git solution.
Now remote git repositories (akin to svn repos) do not have a good locking large binary file solution. But other than large binaries, git is very nice in comparison (You could do hg or bzr if you prefer and gain the same, although their git/svn bridges are not known to me).
It seems like this article is premature. I would like to have read about their experiences a few more months down the line when there's been a bit more real world use.
I just attended a DVCS talk today (http://worldtour.fogcreek.com/) and one simple difference between SVN and DVCSs is still fresh in my mind:
When you are pushing your changes back to trunk, how safe is that process?
In a normal SVN workflow, when you're ready to commit your changes to the central repository to which a team member has also committed changes, you'll end up running 'svn up' followed by 'svn commit.' If you have substantial code changes and you screw up the merge during 'svn up,' there's little record of your pre-update code. In the case that you accidentally select the trunk version of a changed file for the merge, your own recent code may be overwritten.
With Git/Hg/etc, you are constantly committing your changes to the VCS. And you commit before you merge your working copy with any changes pulled from the "central" repository. If you screw up the merge, the DVCS has a record of your local work-in-progress code changes.
You could argue with SVN that any sufficiently large feature transaction should be done in a branch. Then the merge back to trunk will be safe with the branch's commit history. Pragmatically speaking, however, many people - especially small teams - end up doing all of their SVN-based development simply using 'svn up' and 'svn commit.' I know I have.
The safety that comes with a detailed commit history, especially for day-to-day file merges, is one big advantage of DCVSs.
We moved from CVS to Git over the summer. It was a pretty easy transition (we only have seven employees) though it took some explaining about the local-vs-remote copies.
The founders originally wanted to upgrade from CVS to SVN. I pointed-out that not only would they have transactional commits in Git, but that they could recover from anyone's copy of the repository, since everyone has a local clone. That, plus the first check-in (which ran blazingly fast) ended the argument.
31 comments
[ 0.26 ms ] story [ 97.2 ms ] threadThe equivalent in svn would make me groan, for complexity and for the performance.
Also, the same rules apply when 'cave-hiding'. You should still be committing early and often, but in gitland, 'committing' really means syncing up with an authoritative repo. You can 'cave-hide' in svn by just not committing anything and keeping your changes in your local working copy.
Creating branches in svn is cheap, switching branches is cheap, checking out a new working copy if you want to work on two changes at once is also cheap, merging is cheap. There's no reason not to do what you're suggesting in svn.
One of the things that git does better than svn is merging branches when files have moved, but I don't think your use case addresses this directly. Git does some things better than svn, but svn does a good enough job that you have to be a little careful when making the case against it.
This is just misinformation. Merging is absolutely not cheap, how incredibly expensive it is in SVN is the reason some of the DVCSes of today even exist. Try doing some of the common Git work flows [1] in SVN and let us know how well that works out. I've done it. Once. After that we adopted a methodology to ensure we'd never have to do it again.
[1] Pretty much anything where you are moving change sets between multiple branches before pushing back to main. Trivial in a DVCS, a nightmare in SVN.
I think then the problem is our definitions of 'cheap'.
A typical switch command on a working copy of our codebase at work (20,000 + files) takes 1-2 minutes, and its a lot longer for creating a new branch. In a git version of the same repository both these operations are comparatively instantaneous. YMMV.
Also, regarding merging, I believe that in new versions of the svn client there is mergeinfo, however older repositories without this metadata will give you issues when you for example create a branch, merge changes from trunk into your branch and then subsequently attempt to merge that branch back into trunk.
This is the classic 'svn sucks at merging' problem and to me, its complexity overhead that would deter me from branching, i.e. not 'cheap'. I normally get around this by remembering the revision number of the last 'rebase' from trunk (or give it an easily greppable commit message) and then when I do the merge back into trunk I specify that I only want changes from the last rebase to HEAD.
The internal model of git makes this a non-issue, i.e. each commit has a parent commit, and for any given commit on two separate branches, git just walks the tree up to the nearest common ancestor.
Yes, old versions of svn "suck at merging", but the workflows with merges between multiple branches you describe work well in modern versions of svn with merge tracking. Your "remembering the revision number of the last 'rebase'" is unnecessary in modern svn.
I grant you that switch can take a while on a large code base, as can a checkout of a new working copy. Creating a new branch is contrary to what you say instantaneous in svn due to the cheap copies, however getting that new branch in a new working copy to work on I admit might take a while. A large part of this problem is mitigated if you just have multiple working copies checked out, but not all.
The details on why branching is cheap in svn are described here:
http://svn.apache.org/repos/asf/subversion/trunk/notes/subve...
Part of the problem with svn is that the model is too flexible compared to git: the merge tracking supports tracking partial merges between branches, which is arguably an unnecessary complexity - the svn developers have expressed concern about whether this feature is valuable given the complexity it adds.
For general merge scenarios if you're using mergeinfo correctly, follow the practices laid out in the svn book, you shouldn't find merging particularly difficult or expensive in svn, even if you're doing relatively complicated workflows.
I'm not saying this is easier or works better than svn than git, just that it's doable and not as bad as many people make out. The merging in svn is (too) complex, but once you figure out how it works then it works well. With your "I believe...there is mergeinfo" and loewenskind's "I've done it. Once." I get the impression that neither of you have put the time in to really figure out how it works, but are still arguing against it.
Personally I prefer the way git works. But as someone who uses it everyday, for scenarios like the ones you say are nightmares, I don't think svn is as bad as you make out. There is an argument to be made for using git, but neither you, loewenskind nor the article make it as well as you could.
You might want to investigate how to to set up an EBS backed persistent root filesystem on AWS as well as how to do S3 snapshot backups. EC2 instances fail. EBS volumes fail. You need to be ready for that regardless of the revision control choices you make.
You can just train the "new tech tolerant ones" in git-svn and don't worry as well.
Also, SVN stores a local copy of the codebase as well. If you're working off trunk and the SVN server detonates, all you lose is the history and any branches you didn't have checked out.
Edit: We also have a script that takes a daily snapshot of all our EBSes, but we haven't tried to recover them (mostly due to lack of time to experiment).
That refers to exchanging clarity or simplicity for performance. If a technique is fundamentally superior then it's just the way you should do things, and few things in tech are as unilaterally superior as git over svn.
At least in the Ruby world, if you are on SVN, say goodbye to any competent devs. We can't work under those conditions.
That's really ok with me (although I also use git).
I would totally push git (I pushed for git in some places if you wonder) if I thought it would be a good fit for them, but it's not.
What, having to use "git svn dcommit" to do pushes is a huge deal?
If you're going to be a religious git adherent, learn to use ALL of git.
It seems like this is the main reason why you switched from SVN to Git -- hype.
Now remote git repositories (akin to svn repos) do not have a good locking large binary file solution. But other than large binaries, git is very nice in comparison (You could do hg or bzr if you prefer and gain the same, although their git/svn bridges are not known to me).
When you are pushing your changes back to trunk, how safe is that process?
In a normal SVN workflow, when you're ready to commit your changes to the central repository to which a team member has also committed changes, you'll end up running 'svn up' followed by 'svn commit.' If you have substantial code changes and you screw up the merge during 'svn up,' there's little record of your pre-update code. In the case that you accidentally select the trunk version of a changed file for the merge, your own recent code may be overwritten.
With Git/Hg/etc, you are constantly committing your changes to the VCS. And you commit before you merge your working copy with any changes pulled from the "central" repository. If you screw up the merge, the DVCS has a record of your local work-in-progress code changes.
You could argue with SVN that any sufficiently large feature transaction should be done in a branch. Then the merge back to trunk will be safe with the branch's commit history. Pragmatically speaking, however, many people - especially small teams - end up doing all of their SVN-based development simply using 'svn up' and 'svn commit.' I know I have.
The safety that comes with a detailed commit history, especially for day-to-day file merges, is one big advantage of DCVSs.
The founders originally wanted to upgrade from CVS to SVN. I pointed-out that not only would they have transactional commits in Git, but that they could recover from anyone's copy of the repository, since everyone has a local clone. That, plus the first check-in (which ran blazingly fast) ended the argument.