Git --distributed-is-the-new-centralized

Smart HTTP Transport

When I was done writing Pro Git, the only transfer protocols that existed were the git://, ssh:// and basic http:// transports. I wrote about the basic strengths and weaknesses of each in Chapter 4. At the time, one of the big differences between Git and most other VCS's was that HTTP was not a mainly used protocol - that's because it was read-only and very inefficient. Git would simply use the webserver to ask for individual objects and packfiles that it needed. It would even ask for big packfiles even if it only needed one object from it.

As of the release of version 1.6.6 at the end of last year, however, Git can now use the HTTP protocol just about as efficiently as the git or ssh versions (thanks to the amazing work by Shawn Pearce, who also happened to have been the technical editor of Pro Git). Amusingly, it has been given very little fanfare - the release notes for 1.6.6 state only this:

* "git fetch" over http learned a new mode that is different from the
  traditional "dumb commit walker".

Which is a huge understatement, given that I think this will become the standard Git protocol in the very near future. I believe this because it's both efficient and can be run either secure and authenticated (https) or open and unauthenticated (http). It also has the huge advantage that most firewalls have those ports (80 and 443) open already and normal users don't have to deal with ssh-keygen and the like. Once most clients have updated to at least v1.6.6, http will have a big place in the Git world.

What is "Smart" HTTP?

Before version 1.6.6, Git clients, when you clone or fetch over HTTP would basically just do a series of GETs to grab individual objects and packfiles on the server from bare Git repositories, since it knows the layout of the repo. This functionality is documented fairly completely in Chapter 9. Conversations over this protocol used to look like this:

$ git clone
Initialized empty Git repository in /private/tmp/simplegit-progit/.git/
got ca82a6dff817ec66f44342007202690a93763949
walk ca82a6dff817ec66f44342007202690a93763949
got 085bb3bcb608e1e8451d4b2432f8ecbe6306e7e7
Getting alternates list for
Getting pack list for
Getting index for pack 816a9b2334da9953e530f27bcac22082a9f5b835
Getting pack 816a9b2334da9953e530f27bcac22082a9f5b835
 which contains cfda3bf379e4f8dba8717dee55aab78aef7f4daf
walk 085bb3bcb608e1e8451d4b2432f8ecbe6306e7e7
walk a11bef06a3f659402fe7563abf99ad00de2209e6

It is a completly passive server, and if the client needs one object in a packfile of thousands, the server cannot pull the single object out, the client is forced to request the entire packfile.

In contrast, the smarter protocols (git and ssh) would instead have a conversation with the git upload-pack process on the server which would determine the exact set of objects the client needs and build a custom packfile with just those objects and stream it over.

The new clients will now send a request with an extra GET parameter that older servers will simply ignore, but servers running the smart CGI will recognize and switch modes to a multi-POST mode that is similar to the conversation that happens over the git protocol. Once this series of POSTs is complete, the server knows what objects the client needs and can build a custom packfile and stream it back.

Furthermore, in the olden days if you wanted to push over http, you had to setup a DAV-based server, which was rather difficult and also pretty inefficient compared to the smarter protocols. Now you can push over this CGI, which again is very similar to the push mechanisms for the git and ssh protocols. You simply have to authenticate via an HTTP-based method, like basic auth or the like (assuming you don't want your repository to be world-writable).

The rest of this article will explain setting up a server with the "smart"-http protocol, so you can test out this cool new feature. This feature is referred to as "smart" HTTP vs "dumb" HTTP because it requires having the Git binary installed on the server, where the previous incantation of HTTP transfer required only a simple webserver. It has a real conversation with the client, rather than just dumbly pushing out data.

Setting up Smart HTTP

So, Smart-HTTP is basically just enabling the new CGI script that is provided with Git called git-http-backend on the server. This CGI will read the path and headers sent by the revamped git fetch and git push binaries who have learned to communicate in a specific way with a smart server. If the CGI sees that the client is smart, it will communicate smartly with it, otherwise it will simply fall back to the dumb behavior (so it is backward compatible for reads with older clients).

To set it up, it's best to walk through the instructions on the git-http-backend documentation page. Basically, you have to install Git v1.6.6 or higher on a server with an Apache 2.x webserver (it has to be Apache, currently - other CGI servers don't work, last I checked). Then you add something similar to this to your http.conf file:

SetEnv GIT_PROJECT_ROOT /var/www/git
ScriptAlias /git/ /usr/libexec/git-core/git-http-backend/

Then you'll want to make writes be authenticated somehow, possibly with an Auth block like this:

<LocationMatch "^/git/.*/git-receive-pack$">
        AuthType Basic
        AuthName "Git Access"
        Require group committers

That is all that is really required to get this running. Now you have a smart http-based Git server that can do anonymous reads and authenticated writes with clients that have upgraded to 1.6.6 and above. How awesome is that? The documentation goes over more complex examples, like making it work with GitWeb and accelerating the dumb fallback reads, if you're interested.

Rack-based Git Server

If you're not a fan of Apache or you're running some other web server, you may want to take a look at an app that I wrote called Grack, which is a Rack-based application for Smart-HTTP Git. Rack is a generic webserver interface for Ruby (similar to WSGI for Python) that has adapters for a ton of web servers. It basically replaces git http-backend for non-Apache servers that can't run it.

This means that I can write the web handler independent of the web server and it will work with any web server that has a Rack handler. This currently means any FCGI server, Mongrel (and EventedMongrel and SwiftipliedMongrel), WEBrick, SCGI, LiteSpeed, Thin, Ebb, Phusion Passenger and Unicorn. Even cooler, using Warbler and JRuby, you can generate a WAR file that is deployable in any Java web application server (Tomcat, Glassfish, Websphere, JBoss, etc).

So, if you don't use Apache and you are interested in a Smart-HTTP Git server, you may want to check out Grack. At GitHub, this is the adapter we're using to eventually implement Smart-HTTP support for all the GitHub repositories. (It's currently a tad bit behind, but I'll be starting up on it again soon as I get it into production at GitHub - send pull requests if you find any issues)

Grack is about half as fast as the Apache version for simple ref-listing stuff, but we're talking 10ths of a second. For most clones and pushes, the data transfer will be the main time-sink, so the load time of the app should be negligible.

In Conclusion

I think HTTP based Git will be a huge part of the future of Git, so if you're running your own Git server, you should really check it out. If you're not, GitHub and I'm sure other hosts will soon be supporting it - upgrade your Git client to 1.7ish soon so you can take advantage of it when it happens.