Hi Juliano, Thank you for that elaborate analysis. Aside from functionality comparison, I find one thing missing that would be important if the MediaWiki project were to facilitate another tool for SCM: multi platform support. Can you add some information on that, too, because I know that we have developers on Linux, OS X, and Windows, and we wouldn't really like to exculde one of those groups just because of the used SCM tool. Having proper clients on all major platforms is one of the reasons I really like Subversion for... Cheers! Siebrand -----Oorspronkelijk bericht----- Van: wikitech-l-bounces(a)lists.wikimedia.org [mailto:wikitech-l-bounces@lists.wikimedia.org] Namens Juliano F. Ravasi Verzonden: zaterdag 15 november 2008 4:32 Aan: Wikimedia developers Onderwerp: Re: [Wikitech-l] SVN update on live sites to r43514, questions on branching Hello Brion, Brion Vibber wrote: That is probably not branching that you meant, but merge tracking. There have been reasonable merge tracking support available for a long time for Subversion, either through the svnmerge.{sh,py} utility (that is distributed along with Subversion itself, since 1.2) or through SVK (both existed since before Git)... and *nobody cared*! People either branched and merged anyways, like described in svnbook (until 1.5), or a few looked for these handier tools when they needed. [..]

Hello Siebrand, Siebrand Mazeland wrote: Yeah, that is something that Subversion still shines bright compared to all other tools. If you don't mind using the command line in Windows, Git, Mercurial and Bazaar all have good command line support, virtually the same between different OSes. But usually, Windows users don't like the command line, and here is where it makes a difference. Git is only officially supported in Windows in Cygwin, which makes it slow as a rock, and definitely no real Windows integration. Git itself was not designed to run on non-Posix platforms. Then there is the msysGit fork that strives to "fix" that, and have Git compile on MingW32. The problem, again with Git original design, is that msysGit end up depending on bash, perl, tcl, tk, vim and other GNU ports from MingW32, most of these are aliens in a pure-Windows environment. You mostly have to run msysgit in its own private bash shell, which redefines the meaning of some Windows commands, like find, in order to match those in Linux (that is expected from Git's porcelain). Git-svn is another pain. MsysGit also offers virtually no Windows integration. It comes with some standalone interface (git-gui, gitk), and there are plans of a Tortoise-like shell integration called git-cheetah, not ready for consumption. People argue that Git works "fine" in Windows through msysGit, but usually for people with good Linux experience, that fell comfortable in a bash shell. Most Windows developers don't match this profile. Mercurial provides *much* better Windows support, since it is designed to run over Python with whatever it provides and making few assumptions about the underlying operating system. Python is natively supported on Windows, so the end result is much better than Git. It had a few problems in the past (e.g. with filesystem case-insensitivity), but that is fixed by now. Two-way Svn communication is still under development. There is TortoiseHg, that provides Windows integration, and supports most operations the user usually needs, but it's interface is currently really ugly, in Gtk+ (not Windows GDI native), with buttons in strange places, etc. It shares libraries with TortoiseSVN 1.5, and works really well. It includes kdiff3, which is a great tool for merging. Bazaar shares the same advantages of being written in Python with Mercurial. Bazaar properly handles Unicode filenames in different locales (something that both Git and Mercurial fail), and is the most user-friendly of the three. TortoiseBzr is included with the default Windows installer, has better i18n support (more friendly for non-English speakers), and is really well polished compared to TortoiseHg; but it is still experimental, and a lot of functionality is either missing or broken. It has somewhat good two-way Svn communication. And, in case you want to keep your central Subversion repository, there is SVK. It works great in Windows, but is just a command-line tool. There is no GUI, but you have the advantage that all your current Subversion tools (specially, TortoiseSVN and Eclipse) are completely usable in a distributed manner, without modifications. In this case you use the SVK command-line to mirror, pull, push and merge, and Subversion for the local branch development. Not really user-friendly, but works well most of the time. I have some big interest (as in scientific research interest) in version control systems. I give presentations and short courses about VCSes to teams of developers, most of them PHP, Java and C# developers, most of them in Windows environment. The first thing that I notice clearly is that Windows developers, who don't have a lot of contact with Linux or open-source, don't care a lot about distributed version control. For them, having a user-friendly, feature-complete, fully-integrated UI like TortoiseSVN and Subeclipse is way more important than being capable of sharing commits in a distributed manner. They recognize the benefits of DVCSes, but they are more objective: "Do I really need these specific features? Will them give me a raise? No? So thanks." Those who are more enthusiastic and want to jump into distributed version control, they find Mercurial and Bazaar much easier (and saner) to use than Git. I was once told that there is no point in spending too much time dealing and learning about a single version control tool, even though it offers more features (referring to Git over Hg), than doing actual work. That sounded sensible, and made me think about the fact that most DVCS fuss came after Torvalds presentation (even though they weren't any news for quite a time already), and mostly around Git, and mostly moved by open-source enthusiasts. It really looks like the "Linus said so" syndrome. Open-source enthusiasts tend to regard Torvalds as a deity, and give him too much credibility, even when his mouth is just expelling a lot of bullsh*t. -- Juliano F. Ravasi ·· http://juliano.info/ 5105 46CC B2B7 F0CD 5F47 E740 72CA 54F4 DF37 9E96 "A candle loses nothing by lighting another candle." -- Erin Majors * NOTE: Don't try to reach me through this address, use "contact@" instead.

Platonides wrote: What SVK does is very similar to svnsync, and that is the point. The svnsync is *slow* over the network. With a dumpfile, SVK allows a very fast initial mirroring, using the svnsync protocol only for the later syncs. -- Juliano F. Ravasi ·· http://juliano.info/ 5105 46CC B2B7 F0CD 5F47 E740 72CA 54F4 DF37 9E96 "A candle loses nothing by lighting another candle." -- Erin Majors * NOTE: Don't try to reach me through this address, use "contact@" instead.

On Sun, Nov 16, 2008 at 12:38 AM, Aryeh Gregor <Simetrical+wikilist(a)gmail.com> wrote: Git AFAIR even has a native commandline client on Windows. Which has some problems: a) speed, it is slower than on *nix b) the commandconsole itself, cmd.exe is bitrotten from down to top, look at command history for excellent examples. not everyone has the SFU installed. c) AFAIR(!) it doesn't support all commands of *nix git Cygwin of course is the other possibility, but here applies the problem of speed and the work amount needed to install it. It can really suck to get a working cygwin environment... Marco

Aryeh Gregor wrote: Branching is a more general term. An svn checkout, for example, is a branch. Not the kind of named branch you create in the repository, but a small, one-level branch, nevertheless. Most branches created in Subversion are release branches and tags (create-only, locked branches), that either don't require merge tracking, or share only single commits between them. Also, Subversion supports the idea of inner branches, that no other version control (AFAIK) supports. Like: cd skins/ svn cp Monobook.php MyOwnSkin.php svn ci -m "branched Monobook skin" /* make changes in MyOwnSkin.php */ svn ci -m "changes to my skin" /* time passes */ svn update /* changes to Monobook are received */ svn merge $repo/skins/Monobook.php MyOwnSkin.php svn ci -m "merged changes from Monobook to MyOwnSkin" Not really. To explain, I'd have to dig too much into the inner workings of both Git and Svn. But in short, they both boil down to very similar procedures. The difference is in how this information about the merge is stored and interpreted. Also, since Svn was born as centralized, the information registered and immediately available on svn log doesn't match what a Git user would expect. It is possible (but still cumbersome) to check the diffs in svn:mergeinfo properties, and extract the same information very similar to what Git stores. Also, every commit in Git has a unique SHA1 identifier, that is perceived in different branches. In Svn, every commit is referenced by where it occurred first (in the branch where it happens with the lowest revision ID), then this information (branch path + revision id) is copied into other branches svn:mergeinfo properties to indicate that that commit was transplanted. Note also that Svn merge tracking is more complex (in an algorithmic point-of-view) than Git's and Mercurial's, since Svn supports detachable directories and partial checkouts. This same extra complexity is what enables the support for merge tracking of inner branchs I mentioned above. I'll take a look at it later. But I agree that using other DVCS with less restrictions than SVK is better. But if you know how to use SVK correctly, it plays nicer with the upstream Svn repository than the git-svn bridge. That is from my own experience. I'm suggesting SVK just because AFAIK MediaWiki is not yet planning to move to distributed version control. When that day comes, then surely there are much better tools than SVK. Directory version isn't just about empty directories. Directory versioning also plays a big role for move/rename history tracking, merge tracking and version control metadata (Svn properties). Simple example: you rename a directory with a thousand versioned files inside. You don't make any changes to the files, just the directory name. In Svn, it is recorded as a single directory move operation, all files register *no* change, since they actually weren't touched. In Git, it is recorded as a thousand files removed, plus a thousand files added, which is a *wrong* description of what happened. Also, since Git doesn't record rename history upon commit (that happens only once), this information has to be figured out later (that may happen many times) using an exponential-complexity algorithm. It doesn't make any sense when you think about it. But this problem of copy/move information being ignored/discarded have already caused so much pain to Git that I still wonder why they insist in calling that a feature. Demanding the user to pass -M, -C and --find-copies-harder options to find and use the correct information, causing an exponential-complexity algorithm to be performed many times over the same information that could be simply stored in the commit manifest. It sounds completely illogical. Sure. For many reasons (technical and ethical reasons). Kind off-topic in this thread and this list. No, you are right. SVK is almost always forgotten. That means smaller community, less documentation, less support, slower development, etc. I use it, with MediaWiki's repository. Using it, I can sync between MW's repo, my local mirror, and my own public MW repo using it, something that would be extremely difficult with git-svn. Rewriting history is something that SVK doesn't support by design. To rewrite history means to lose information, and that is considered wrong (even if you think that you won't need the overwritten history anymore). That doesn't mean you can't *recreate* a new history as you like. You rebranch from the point you want to recreate, do whatever changes you want until the head of the former branch, then delete the former branch. You have the same effect, but information is never lost forever. If you find that you screwed up something trying to recreate the history, the old history is still there to aid you. Not exactly true. It is true that SVK is optimized to work with upstream repositories, but it is not true that you are unable to share local branches without the *main* upstream. You can have *multiple upstream repositories* in SVK. That may blow your mind if you think that git-svn is the best bridge between your local repository and the upstream Subversion repo. SVK doesn't share branches directly with other SVK, but nothing blocks you from setting up a temporary repository between you and your friend that is on this cruise, replicate your local branch to this temporary repository, and give him access to this temporary repository. It is just a different way of developer-to-developer push that other DVCSes support. It may be even simpler if you give him access directly to your internal SVK depot, but you usually don't want him to have direct access to your depot. You have to know the tool you use. The problem is that very few people know how to use SVK, very few people put the same effort they do into learning Git, and take the easy path of talking bad about it. SVK doesn't even create *any* bookkeeping directory, unless you order it to do so. SVK checkouts are *completely* clean by default. So this is even less of an issue for SVK. Because he didn't read the documentation. It was a design decision to keep checkouts completely clean and have bookkeeping information stored in a different private area. SVK provides commands to move checkouts, and to relocate checkout paths if you move a working copy without using SVK. And now (since 2.0 I think) SVK already supports floating working copies, if you want. Best regards, Juliano. -- Juliano F. Ravasi ·· http://juliano.info/ 5105 46CC B2B7 F0CD 5F47 E740 72CA 54F4 DF37 9E96 "A candle loses nothing by lighting another candle." -- Erin Majors * NOTE: Don't try to reach me through this address, use "contact@" instead.