Related
I am about to start participating in the development of a medium-sized project (~50k lines) that was until now written by a single person, and not versioned; as a result folders are cluttered with different versions of the same file (named file1, file2, file3, etc.).
I proposed to start using a VCS for it (a priori Mercurial, which is the only one I've ever used -- for my personal projects --, but I'm open to suggestions), so I'm taking any good ideas as to how to "start" the repository. E.g., should I make an initial commit with all the existing files, and immediately make a new commit with the unused files removed? Or something else?
(constructive remarks on mercurial vs bazaar vs git vs whatever are also welcome.)
Thanks for your tips.
E.g., should I make an initial commit with all the existing files, and immediately make a new commit with the unused files removed?
If the size of the repository is not a concern, then yes, that is a good starting point. Otherwise you can just commit what's actually used, and go from there.
As for which system, all DVCSes stick to the same core principles. Which one you pick is entirely subjective — the only way to truly know which one you like is to try each one.
I would say use what you are the most comfortable with and meets your needs. As far as where to start, I personally would seed the repo with the current source as is, that way you can verify that everything builds and runs as expected. you can make this initial seed a branch. That way you can always go back to your starting point before refactoring.
My approach to this was:
create a Mercurial repository in the existing project folder ("existing")
commit all project files to "existing"
create an empty repository in what a different location ("new")
As files are tested and QA'd (this was necessary because there was so much dross in "existing") pull them from "everything" to "new".
Once files had been pulled into "new"; delete the corresponding files from "existing". If access is needed to these files while the migration is under way, push them back from "new" to "existing".
This gave me the advantage of putting everything under some sort of control for recovery purposes, control over introducing the project to the DVCS. Eventually the existing project folder became completely tested and approved for the project moving forward. At this point the "everything" directory could be deleted or changed into a working folder; and "new" became the actual project folder.
I think Mercurial is a good choice. Lightweight, fast, very simple to use and well-integrated with Windows (if that's the platform you're dealing with).
I would probably get rid of all the clutter before the first commit. Delete everything you don't care about, run all the necessary tests and only then do the commit.
Yes, I'm dead set against the 0-day cluttering of repos.
Granted, a 50K SLOC project isn't very big, but if you commit files you already know you won't need, they will make your repo slightly bigger.
Also, remember to check that the tree doesn't contain large binary files. If it does, get rid of them if at all possible.
I have been following the tutorial http://www.burlingtontelecom.net/~ashawley/rcs/tutorial.html on how to work with files using RCS. This works well but only with one file. Is there a way to create an RCS file with directories as well?
I have a project folder called myproject, and in this directory I have all my files for that project. I want to create a revision control system for the myproject folder and all its files that are inside.
As William's comment says, RCS only works with single files. (It also doesn't seem to be particularly suitable for multiple-user stuff.)
Of course, nothing stops you from putting each (source) file in a directory under RCS control; in fact, this is essentially what CVS does (though in recent versions it handles the RCS data itself, rather than invoking RCS to do it as it used to do). Unfortunately, this fragments the change history rather badly; a commit affecting many files ends up as separate commits to each file, which just happen to have the same commit message (and timestamp?), and in general every file will have a different revision in what the user might like to think of as the "same" revision. (This makes tags quite essential.) CVS also has issues with the atomicity of commits: you could end up with commit A and commit B getting tangled up, such that in file foo commit A precedes commit B, but in file bar commit B precedes commit A!
SVN (Subversion) is an attempt to rectify some of the problems in CVS, though it also brings some new limitations, and keeps many of the existing ones; it is probably wiser (as William implies) to just use a distributed version control system (DVCS) for your multi-file projects. There are many choices:
Darcs uses a unique patch-based model: a repository is treated as a sequence of patches, which can be applied to an empty tree to build the current revision; patches can often be reordered by "commuting" pairs of patches, and cherry-picking patches from other repositories is quite easy. The downside is that the change history is a bit less clear than in most DVCSes. See http://wiki.darcs.net/Using/Model, http://en.wikibooks.org/wiki/Understanding_Darcs/Patch_theory.
Directed-acyclic-graph (DAG) based DVCSes model a repository as a directed acyclic graph of revisions, where each revision can have one parent, two parents, or perhaps more. Each revision has an associated file tree state; sometimes renames are also tracked somehow.
Git, as already mentioned. Has a very simple model, but a very complicated interface: there are many commands, some of which are not really intended for humans to use (owing to many parts of it having been prototyped in shell script, probably), so it can be hard to find the ones you want. Also, its model might be a bit too simple: it doesn't track renames at all.
Bazaar (a.k.a. bzr) has a more complicated model, including support for file/directory renames. It's difficult to say how much more complicated, though, because whatever documentation may exist is not nearly as accessible as Git's. It does, however, have a rather simpler user interface, and there are a number of useful plugins, including a distributed-development-friendly SVN plugin: committing from a branch back to SVN need not interfere with the validity of others' branches of your branches, and bzr metadata is even committed back to SVN. Can make things much less painful if you want to start hacking on an SVN-based project without having commit access, but hope to get your changes committed eventually. Bazaar is my personal favorite DAG-based DVCS.
Mercurial (a.k.a. hg) seems fairly similar to Bazaar, though I think it tracks renames only for individual files, not for directories. It also supports plugins, though its SVN plugin isn't as nice as Bazaar's: it doesn't support lossless commits, so branching from other peoples' branches is unwise. I don't have much experience with it, so I can't really evaluate it in-depth.
As the comments already mention, if you are starting out with version control, you would be well advised to choose a newer system than RCS (git, mercurial, fossil, subversion, ...). That said, RCS still works fine for a single developer working primarily on a single machine - I still use it for my own code because I've not yet OK worked out how to get the (20+ years of) history I want into git in the way I want it.
Anyway, to use RCS, make sure you have an RCS sub-directory in each directory where you have working source code under RCS management. The RCS files will be placed in the sub-directory automatically, and retrieved automatically. If your version of make is not already aware of RCS, then you can train it so that it is - or get a version of make that does (GNU Make, for example).
TL:DR - Look into DCVS for an alternative of RCS. It uses CVS, which uses RCS, but it's more modular for working in a repository that is distributed, as well as having a hierarchy of directories.
I'm currently going through a similar issue, and may have found something worthy of note, especially for people who are being forced to use a light, command-line based revision control systems with multiple team members.
My manager will not get off this idea of using RCS as our version control. But for the specifications, he wants developers to be able to create and edit on their own repository on a localized server within our company. Two issues with this:
RCS does not create, nor hold any sort of 'repository'. It is software that keeps track of file edits, on a Per File Basis. Meaning that the 'repository' is nothing more than another directory with RCS checked-in files. This is sub-par for team-geared projects, to say the least.
On a large project with multiple directories and tens of individual working files, even the prospect of creating a top-level RCS directory with a symbolic link in the working directories gives rise to complications such as naming conventions, as well as forgetting which file came from which bottom-level / working directory.
With what SamB posted, even CVS gives additional problems with RCS that we now have to account for, but gives us a slight ability for some additional hierarchy. But one suggestion he forgot was DCVS.
It's nothing more than an extension of CVS, CVSup, and:
contains functionality to distribute CVS repositories with local lines of development and automatically handles synchronization of the distributed repositories in the background.
I work on a large project where all the source files are stored in a version control except the project files. This was the lead developer's decision. His reasoning was:
Its to time consuming to reconcile the differences among developers' working directories.
It allows developers to work independently until their changes are stable
Instead, a developer initially gets a copy of a fellow developer's project files. Then when new files are added each developer notifies all the rest about the change. This strikes me as far more time consuming in the long run.
In my opinion the supposed benefits of not tracking changes to the project files are outweighed by the danger. In addition to references to its needed source files each project file has configuration settings that would be very time consuming and error prone to reproduce if it became corrupted or there was a hardware failure. Some of them have source code embedded in them that would be nearly impossible to recover.
I tried to convince the lead that both of his reasons can be accomplished by:
Agreeing on a standard folder structure
Using relative paths in the project files
Using the version control system more effectively
But so far he's unwilling to heed my suggestions. I checked the svn log and discovered that each major version's history begins with an Add. I have a feeling he doesn't know how to use the branching feature at all.
Am I worrying about nothing or are my concerns valid?
Your concerns are valid. There's no good reason to exclude project files from the repository. They should absolutely be under version control. You'll need to standardize on a directory structure for automated builds as well, so your lead is just postponing the inevitable.
Here are some reasons to check project (*.*proj) files into version control:
Avoid unnecessary build breaks. Relying on individual developers to notify the rest of the team every time the add, remove or rename a source file is not a sustainable practice. There will be mistakes and you will end up with broken builds and your team will waste valuable time trying to determine why the build broke.
Maintain an authoritative source configuration. If there are no project files in the repository, you don't have enough information there to reliably build the solution. Is your team planning to deliver a build from one of your developer's machines? If so, which one? The whole point of having a source control repository is to maintain an authoritative source configuration from which you build and deliver releases.
Simplify management of your projects. Having each team member independently updating their individual copies of your various project files gets more complicated when you introduce project types that not everyone is familiar with. What happens if you need to introduce a WiX project to generate an MSI package or a Database project?
I'd also argue that the two points made in defense of this strategy of not checking in project files are easily refuted. Let's take a look at each:
Its to time consuming to reconcile the differences among developers' working directories.
Source configurations should always be setup with relative paths. If you have hard coded paths in your source configuration (project files, resource files, etc.) then you're doing it wrong. Choosing to ignore the problem is not going to make it go away.
It allows developers to work independently until their changes are stable
No, using version control lets developers work in isolation until their changes are stable. If you each continue to maintain your own separate copies of the project files, as soon as someone checks in a change that references a class in a new source file, you've broken everyone on the team until they stop what they're doing and carefully update their project files. Compare that experience with just "getting latest" from source control.
Generally, a project checked out of SVN should be working, or there should be tools included to make it work (e.g. autogen.sh). If the project file is missing or you need knowledge about which files should be in the project, there is something missing.
Automatically generated files should not be in SVN, as it is pointless to track the changes to these.
Project files with relative path belong under source control.
Files that don't: For example in .Net, I would not put the .suo (user options) web.config (or app.config under source control. You may have developers using different connection strings, etc.
In the case of web.config, I like to put a web.config.example in. That way you copy the file to web.config upon initial checkout and tweak what settings you'd like. If you add something that needs to be added to all web.config, you merge those lines into the .example version and notify the team to merge that into their local version.
I think it depends on the IDE and configuration of the project. Some IDEs have hard-coded absolute paths and that's a real problem with multiple developers working on the same code with different local copies and configurations. Avoid absolute path references to libraries, for example, if you can.
In Eclipse (and Java), it's fine to commit .project and .classpath files (so long as the classpath doesn't have absolute references). However, you may find that using tools like Maven can help having some independence from the IDE and individual settings (in which case you wouldn't need to commit .project, .settings and .classpath in Eclipse since m2eclipse would re-create them for you automatically). This might not apply as well to other languages/environments.
In addition, if I need to reference something really specific to my machine (either configuration or file location), it tend to have my own local branch in Git which I rebase when necessary, committing only the common parts to the remote repository. Git diff/rebase works well: it tends to be able to work out the diffs even if the local changes affect files that have been modified remotely, except when those changes conflict, in which case you get the opportunity to merge the changes manually.
That's just retarded. With a set up like that, I can have a perfectly working project containing files that are subtly different from everyone else. Imagine the havoc this would cause if someone accidentally propagates this mess into QA and everyone is trying to figure out what's going on. Imagine the catastrophe that would ensue if it ever got released to the production environment...!
Is it a best practice to have my local source tree mirror the server source tree? It seems to me that it should, however my department at work does not do it that way and I find it very confusing. If not, what are scenarios where it makes sense to deviate from the server source tree?
EDIT: To clarify what I mean - say the source directory we want to map on the local machine is here on the server:
\\TeamServer\Project\Releases\2008
On our local machine, that directory would be mapped like this:
D:\2008_Releases
instead of:
D:\Project\Releases\2008
I don't know if it's a best practice but I prefer mirroring the source tree too. It's one less "gotcha" in terms of getting a new developer up and running. Not mirroring the source tree can eventually come back to bite you when it comes to relative paths.
Someone probably made a mistake when they set things up originally and it never got corrected. IMHO it's a minor annoyance; just one of the side effects of not living in a perfect world ;)
If your local tree needs to be at the same path as it is on the server, then you can't have multiple copies checked out. At the last two places I've worked, it was common for me to have several copies of (parts of) the tree checked out at any given time, depending on how many different bugs or features I was working on, and how many branches of the product the bugs or features were in.
Personally, I have no idea where the source trees were stored on the servers, and I didn't need to. I just ran cvs co or svn co to get a copy of the tree in my working directory. As long as I ran make or ant somewhere in the source tree, everything below it would compile.
I prefer to have individual project/release pairs in their own directories as close to root as reasonably possible. It's a horrible PITA to go clickety-click on directory tree or having to type the "c:\project\releases\2008" -part for the umpteenth time.
I also think checking out sources to different path tends to flush out bugged assumptions about project locations (we have postbuild events that do some nasty things).
There is one reason not to do this: When you have access to the production machine. In this case, I prefer to have different paths. This makes it more likely that I notice my "rm -rf" is on the wrong box...
i think mirrowing is a good practise and i like to do it so too, however i have an extra temp folder for the case i need to checkout a fresh version for e.g. to test something with the current version or to fix a bug which has a high priority then what im doing atm.
The most important thing is to understand why the choice was made and use the team to determine if this is something that you would want to change. Team buying is important for these kind of decisions - and the team is more than just the developers. Source control trees contain things such as documentation, tests, resources, etc., so there is a legitimate chance that the structure was determined by multiple parties to find a common ground.
We use TFS where I work, as well. We do have a common root folder called C:\Source Control that is the main folder that all Team Projects live under. This choice was made because of the extreme dislike, by all parties, that the drive would get cluttered with various folders.
With TFS, you have the option to map multiple workspaces, so what you do on your local machine is not dictated by the structure on the server. My personal preference, and my team's preference, as well, is to use a single workspace mapping to the source control root. Given the shelving functionality in TFS, there is not a concern about having multiple copies checked out, since they can be shelved if something else needs to be worked on.
The build server has the same mapping, as well. This in no way, however, matches the deployment structure. Builds are dropped based upon the criteria of the project. The only time this presents a problem is when absolute paths are used, in configuration files, for example. Since we are not using absolute paths (by definition in our developer guidelines) this is a non-issue.
People have different opinions on how to organize their hard disks, why should you enforce a particular layout?
One should be able to check out multiple working copies of the same project in the repository - impossible if you insist using the same hierarchy.
In one large project I worked on, literally thousands of files assumed that the working copy was in a fixed absolute hard-coded path (\projectname). To work with several branches of the project, we had to resort to using different drive letters (on Windows), meaning that we had to divide our hard disks into many partitons (6 or more were common). This was very inconvenient.
Eventually we had to integrate the project as a sub-project of a larger project. This meant changing all those absolute paths, a tedious and time-consuming task.
Using relative paths gives you much more flexibility; however, if everybody has the project root at the exact same location, you won't notice if someone adds an absolute path somewhere by accident.
I need to keep under version some large files (some Gigs).
I don't need, and I can't keep under version all the version of the files.
I want to be able to remove from my VCS large files version at some moment.
The files that I want to keep under version control are big .zip files or ISO images.
These files may contains executable software or data (seismic data, SAR images, GNSS data) and they are provided by the software supplier of my company.
What control version system could I use?
In CVS you can do that by removing the files from the repo. Subversion allows that by dumping the content of the repo and filter it to remove the files (that is a bit cumbersome). Perforce has an obliterate command for that. Many of the newer distributed VCS make it rather difficult by their usage of hashes all over the places and the fact that your repo may have been replicated elsewhere also complicate things. Hg has a strip command (part of the Mq extension), Git can also do that I think.
I don’t think there’s any version control system that allows you do that regularly because that goes against everything version control systems stand for.
Perforce generally allows files to be put in two way, as head revision only (so, you'd only every have one copy) or all revisions. Perforce does have the admin level obliterate command that can be used to delete revisions. Its up to you to query for a list of files, possibly by date or number of revisions, and to specify the revisions to the obliterate command. As the name suggests obliterate deletes the revisions permanently from the database, so, I always generate scripts to do this and review them before running them. If the obliterate command is NOT run with the -Y flag, it will generate a list of what would be obliterated, also very useful.
Somehow I get the impression that you should not use a version control system at all. As said before, what you're trying to do goes against everything you would need a version control system for in the first place.
I suggest you create a file system directory structure that makes sense for what you're trying to accomplish and so that you can structure your data. And just make backup's of those files.
TFS has a destroy command that you can use to permanently delete files or revisions as you see fit.
There is more information at this MSDN article.
Many version control systems allow you to configure them in a way so that they store only the differences between several versions of a file and save space through that.
For example if you have a 1Gig file committed, change a part of it and commit it again, only the changed part will be stored in the version control system.
There won't be 2Gigs used (initial and new file) but only 1Gig+sizeOfChanges.
There's just one downside:if you're storing files which change their whole content from revision to revision this can also be counter-productive as the changes take almost the same space as the original version. Archive files are a example for such files where only a small change in the (real) content can lead to a completely changed content of the archive file.
I'd suggest to test several version control systems on your own and with your specific needs and environment and monitor each one at the server-side how the storage requirements for each system changes.
Some distributed version control systems allow to create "checkpoints" that allow you to use this version as kind of a base revision and safe you from pulling all the history before the checkpoint on every checkout. So you can remove the big files, create a checkpoint, and checkout/clone the repository from that checkpoint to a new directory. Then you have there a new, small repository, but without the history before the checkpoint. It you don't need that history you can burn the old repository on CD and use the new, partial one from now on.
I've only tested it in darcs, and there it works, but YMMV depending on version control system and use cases.
It sounds to me like you need an intelligent backup system, rather than version control.
I use SyncBackSE; it allows you to keep a number of previous versions, and can also do things like "ignore all files changed more than 30 days ago".
It's one of the few bits of paid-for software I use. I think it's worth checking out.
I think you're talking about something like "AlienBrain" "bucket" system, aren't you? The ability to remove some revisions from version control.
If you want to destroy an item, it's normally called "obliterate" and it's supported by a number of systems out there.
Buckets, AFAIK are supported by:
AlienBrain
Accurev
PlasticSCM
I would save such files under a unique name (datestamped, perhaps), and perhaps additionally make a textual reference to the external file in the version control system.
Fossil allows you to do this via the "shun" mechanism. Fossil being a distributed SCM, however, means that this does not affect all repositories (for obvious reasons).