GIT vs. Perforce- Two VCS will enter… one will lea

Question

So I'm in the process of getting GIT sold at work. First thing I need is to convince everyone that GIT is better at what they're already used to doing. We currently use Perforce. Anybody else go through a similar sale? Any good links/advice?

One of the big wins is that we can work with it disconnected from the network. Another win IMO is the way adds/checkouts are handled. More points are welcome! Also we have about 10-20 devs total.

Answer 1

I use Perforce at work. I also use Git because I would still like some form of version control when I'm working on code and can't connect to the server. No, reconcile offline work just is not the same. Here is where I've found git to be a great benefit:

1. Branching speed - git takes a few seconds at most.
2. Conflicts - P4Merge's auto resolve destroyed a week's worth of work once. Ever since then I would rather resolve by hand when merging. When Git prompts me about a conflict, it is actually a conflict. The rest of the time, git resolves stuff correctly and I save heaps of time.
3. Keeping track of merges - If you have one branch that is continuously receiving merges from two other branches, you know what a headache this can be with perforce. With git, the headache is minimized because the result of a merge in git is actually a new commit which knows who its ancestors are.
4. Permissions - I have lost track of the number of times I've tried to work on a file but couldn't because it was not checked out in Perforce. If you worked with XCode (or any editor that does not have a solid Perforce SCM plugin) offline, you know how irritating this can get. I don't have to worry about that with Git. I make my changes. Git does not stop me and tracks them in the background.
5. Keeping the main tree tidy - With git, I can sort my commits and tidy the code up so that the history looks nice and tidy. None of that "checking in this file because it was supposed to be part of the previous checkin" rubbish. I squash commits like that, because they help nobody.
6. Stashing - Your perforce server needs to be version 2010.1 or newer to use the p4 shelve command.
7. Creating patches - Easy to do in git. Don't know if it is possible in Perforce without using the command line.
8. Mailing patches from the GUI - again, git wins here.
9. Disk space - With perforce, every branch is a copy. That means if your source tree is huge, your disk space gets eaten up fast. This is not even counting the additional space once you start building. Why even have a link between branches and disk space? With git, you can have 100 branches and only one branch at a time ever exists. If you specifically want to work on two versions simultaneously you can clone, do your work and then get rid of one clone if you want, without losing anything.
10. If you're on XCode4, perforce support has been dropped and git support is now built in. If you do cross platform work like I do, this matters a lot. With Visual Studio, you can use git extensions. With perforce, it's equally yuck on both OSes. Well, maybe a little more on mac now with XCode4 on the scene.
11. Finding the faulty checkin (or, git bisect rules) - Ever tried to do a binary search with perforce to figure out where a bug was introduced? Quite a hassle, yes? Even more of a hassle when there have been integrates from other branches in the middle. Why? Because there is no automation for such tasks. You need to write your own tool to talk to perforce and you usually don't have the time. With git, you give it the starting points (the "good" point and the "bad" point) and it automates the search for you. Even better, if you have a script that can automate the build and test process, you can hook git up to the script and the whole process of finding the checkin is automated. That is how it should be.
12. Tracking changes across refactors - Try splitting BigClass into SmallClass1 and SmallClass2. To Perforce, BigClass has now ceased to exist and two new classes (SmallClass1 and SmallClass2 have joined the source tree). To Perforce, there is no relation between BigClass and SmallClass1 and SmallClass2. Git, on the other hand, is smart enough to know that x% of BigClass is now in SmallClass1 and y% of BigClass is in SmallClass2 and that BigClass has ceased to exist. Now, from the point of view of someone who is reviewing changes across multiple branches, you tell me which approach you'd find more useful - Git's or Perforce's. Personally, I prefer Git's approach because it more accurately reflects the actual change in the code. Git is able to do this because it tracks content within the file and not the file itself.
13. Centralized or decentralized: Git is a DVCS system while perforce is centralized. A centralized VCS cannot be decentralized later, but a DVCS (especially git) can be centralized. There are several products that add very fine grained access control to git, if that is something the business needs. Personally, I would go with a system that gives me greater flexibility in the long term.
14. Branch mappings: If you want to do branching right in Perforce, you need to create a branch mapping. There are reasons for this, but they are tied to how Perforce conceptualizes a branch. As a developer, or a team, this simply means one more step in the work flow, which I do not consider efficient at all.
15. Sharing work between teams: With Perforce, you cannot break up a submission. Team A is working on feature A. Team B on feature B. Team C works on bug fixes. Now, Teams A & B have to fix a bunch of bugs in order to implement their features. The only thing is, they weren't so disciplined when committing their changes (probably because they're rushing to a deadline) and so their "bug fixes" are parts of larger submissions that also contain new stuff as far as version control on their branches are concerned. However, Team C is now doing a point release and would like to get the bug fixes from the other teams. If they were using Git, Team C could cherry pick the other teams' relevant changes, split them up and only take what they needed without worrying about introducing any partially implemented features. With Perforce, Team C can get the affected files, but would have to separate the relevant changes using a much more manual process.
16. Changing platform - If, for whatever reason in the future, you decide to change your platform of choice, with Perforce, you're at the mercy of Perforce.com and the availability of the tools for the platform of your choosing.
17. Changing to future amazing source control engine X - If you decide to change what you use for source control, extracting your source control history from Perforce and moving it to new system X is going to be a nightmare, because it is closed source and the best you can do is guess - just Google for Perforce to Git migration to get a sense of what I'm talking about. At least with Git, its open source, so it eliminates a lot of the guesswork involved.

Well, that's my 2 cents. In Perforce's defense, I gotta say their customer support rules and so does their Time Lapse View tool. I do not know how to get a time lapse view with git. But for the convenience and time saved, I'd go with git any day.

Answer 2

Using GIT as substitute for bad code line management is common. Many of the disadvantages of Perforce are a result of bad branching strategies. The same for any other centralized tool. If you have to create a ton of branches you are doing something wrong. Why do developers need to create so many branches?

Also, why is working disconnected so important anyway? Just so someone can work on a train? That's about the only place these days you can't get a wireless connection. And even most trains have decent WiFi.

Answer 3

Apparently GitHub now offer git training courses to companies. Quoth their blog post about it:

I’ve been down to the Google campus a number of times in the last few weeks helping to train the Androids there in Git. I was asked by Shawn Pearce (you may know him from his Git and EGit/JGit glory – he is the hero that takes over maintanance when Junio is out of town) to come in to help him train the Google engineers working on Andriod in transitioning from Perforce to Git, so Android could be shared with the masses. I can tell you I was more than happy to do it.

[…]

Logical Awesome is now officially offering this type of custom training service to all companies, where we can help your organization with training and planning if you are thinking about switching to Git as well.

Emphasis mine.

Answer 4

The command that sold me on git personally was bisect. I don't think that this feature is available in any other version control system as of now.

That being said, if people are used to a GUI client for source control they are not going to be impressed with git. Right now the only full-featured client is command-line.

Answer 5

Here's what I don't like about git:

First of all, I think the distributed idea flies in the face of reality. Everybody who's actually using git is doing so in a centralised way, even Linus Torvalds. If the kernel was managed in a distributed way, that would mean I couldn't actually download the "official" kernel sources - there wouldn't be one - I'd have to decide whether I want Linus' version, or Joe's version, or Bill's version. That would obviously be ridiculous, and that's why there is an official definition which Linus controls using a centralised workflow.

If you accept that you want a centralised definition of your stuff, then it becomes clear that the server and client roles are completely different, so the dogma that the client and server softwares should be the same becomes purely limiting. The dogma that the client and server data should be the same becomes patently ridiculous, especially in a codebase that's got fifteen years of history that nobody cares about but everybody would have to clone.

What we actually want to do with all that old stuff is bung it in a cupboard and forget that it's there, just like any normal VCS does. The fact that git hauls it all back and forth over the network every day is very dangerous, because it nags you to prune it. That pruning involves a lot of tedious decisions and it can go wrong. So people will probably keep a whole series of snapshot repos from various points in history, but wasn't that what source control was for in the first place? This problem didn't exist until somebody invented the distributed model.

Git actively encourages people to rewrite history, and the above is probably one reason for that. Every normal VCS makes rewriting history impossible for all but the admins, and makes sure the admins have no reason to consider it. Correct me if I'm wrong, but as far as I know, git provides no way to grant normal users write access but ban them from rewriting history. That means any developer with a grudge (or who was still struggling with the learning curve) could trash the whole codebase. How do we tighten that one? Well, either you make regular backups of the entire history, i.e. you keep history squared, or you ban write access to all except some poor sod who would receive all the diffs by email and merge them by hand.

Let's take an example of a well-funded, large project and see how git is working for them: Android. I once decided to have a play with the android system itself. I found out that I was supposed to use a bunch of scripts called repo to get at their git. Some of repo runs on the client and some on the server, but both, by their very existence, are illustrating the fact that git is incomplete in either capacity. What happened is that I was unable to pull the sources for about a week and then gave up altogether. I would have had to pull a truly vast amount of data from several different repositories, but the server was completely overloaded with people like me. Repo was timing out and was unable to resume from where it had timed out. If git is so distributable, you'd have thought that they'd have done some kind of peer-to-peer thing to relieve the load on that one server. Git is distributable, but it's not a server. Git+repo is a server, but repo is not distributable cos it's just an ad-hoc collection of hacks.

A similar illustration of git's inadequacy is gitolite (and its ancestor which apparently didn't work out so well.) Gitolite describes its job as easing the deployment of a git server. Again, the very existence of this thing proves that git is not a server, any more than it is a client. What's more, it never will be, because if it grew into either it would be betraying it's founding principles.

Even if you did believe in the distributed thing, git would still be a mess. What, for instance, is a branch? They say that you implicitly make a branch every time you clone a repository, but that can't be the same thing as a branch in a single repository. So that's at least two different things being referred to as branches. But then, you can also rewind in a repo and just start editing. Is that like the second type of branch, or something different again? Maybe it depends what type of repo you've got - oh yes - apparently the repo is not a very clear concept either. There are normal ones and bare ones. You can't push to a normal one because the bare part might get out of sync with its source tree. But you can't cvsimport to a bare one cos they didn't think of that. So you have to cvsimport to a normal one, clone that to a bare one which developers hit, and cvsexport that to a cvs working copy which still has to be checked into cvs. Who can be bothered? Where did all these complications come from? From the distributed idea itself. I ditched gitolite in the end because it was imposing even more of these restrictions on me.

Git says that branching should be light, but many companies already have a serious rogue branch problem so I'd have thought that branching should be a momentous decision with strict policing. This is where perforce really shines...

In perforce you rarely need branches because you can juggle changesets in a very agile way. For instance, the usual workflow is that you sync to the last known good version on mainline, then write your feature. Whenever you attempt to modify a file, the diff of that file gets added to your "default changeset". When you attempt to check in the changeset, it automatically tries to merge the news from mainline into your changeset (effectively rebasing it) and then commits. This workflow is enforced without you even needing to understand it. Mainline thus collects a history of changes which you can quite easily cherry pick your way through later. For instance, suppose you want to revert an old one, say, the one before the one before last. You sync to the moment before the offending change, mark the affected files as part of the changeset, sync to the moment after and merge with "always mine". (There was something very interesting there: syncing doesn't mean having the same thing - if a file is editable (i.e. in an active changeset) it won't be clobbered by the sync but marked as due for resolving.) Now you have a changelist that undoes the offending one. Merge in the subsequent news and you have a changelist that you can plop on top of mainline to have the desired effect. At no point did we rewrite any history.

Now, supposing half way through this process, somebody runs up to you and tells you to drop everything and fix some bug. You just give your default changelist a name (a number actually) then "suspend" it, fix the bug in the now empty default changelist, commit it, and resume the named changelist. It's typical to have several changelists suspended at a time where you try different things out. It's easy and private. You get what you really want from a branch regime without the temptation to procrastinate or chicken out of merging to mainline.

I suppose it would be theoretically possible to do something similar in git, but git makes practically anything possible rather than asserting a workflow we approve of. The centralised model is a bunch of valid simplifications relative to the distributed model which is an invalid generalisation. It's so overgeneralised that it basically expects you to implement source control on top of it, as repo does.

The other thing is replication. In git, anything is possible so you have to figure it out for yourself. In perforce, you get an effectively stateless cache. The only configuration it needs to know is where the master is, and the clients can point at either the master or the cache at their discretion. That's a five minute job and it can't go wrong.

You've also got triggers and customisable forms for asserting code reviews, bugzilla references etc, and of course, you have branches for when you actually need them. It's not clearcase, but it's close, and it's dead easy to set up and maintain.

All in all, I think that if you know you're going to work in a centralised way, which everybody does, you might as well use a tool that was designed with that in mind. Git is overrated because of Linus' fearsome wit together with peoples' tendency to follow each other around like sheep, but its main raison d'etre doesn't actually stand up to common sense, and by following it, git ties its own hands with the two huge dogmas that (a) the software and (b) the data have to be the same at both client and server, and that will always make it complicated and lame at the centralised job.

Answer 6

The Perl 5 interpreter source code is currently going through the throes of converting from Perforce to git. Maybe Sam Vilain’s git-p4raw importer is of interest.

In any case, one of the major wins you’re going to have over every centralised VCS and most distributed ones also is raw, blistering speed. You can’t imagine how liberating it is to have the entire project history at hand, mere fractions of fractions of a second away, until you have experienced it. Even generating a commit log of the whole project history that includes a full diff for each commit can be measured in fractions of a second. Git is so fast your hat will fly off. VCSs that have to roundtrip over the network simply have no chance of competing, not even over a Gigabit Ethernet link.

Also, git makes it very easy to be carefully selective when making commits, thereby allowing changes in your working copy (or even within a single file) to be spread out over multiple commits – and across different branches if you need that. This allows you to make fewer mental notes while working – you don’t need to plan out your work so carefully, deciding up front what set of changes you’ll commit and making sure to postpone anything else. You can just make any changes you want as they occur to you, and still untangle them – nearly always quite easily – when it’s time to commit. The stash can be a very big help here.

I have found that together, these facts cause me to naturally make many more and much more focused commits than before I used git. This in turn not only makes your history generally more useful, but is particularly beneficial for value-add tools such as git bisect.

I’m sure there are more things I can’t think of right now. One problem with the proposition of selling your team on git is that many benefits are interrelated and play off each other, as I hinted at above, such that it is hard to simply look at a list of features and benefits of git and infer how they are going to change your workflow, and which changes are going to be bonafide improvements. You need to take this into account, and you also need to explicitly point it out.