Chef is a nice idea written by people who don't understand Unix. (evidence? check out popen in their source, try to figure out why subprocesses occasionally get SIGPIPE, and then grep to see where it's used and count the file handle leaks. I've had to reimplement that six or seven times and theirs was the one with the most subtle errors. Yes, I'll submit a patch when I have time, which at the current rate the day job is pushing me will happen shortly before hell freezes over) I like the design but the implementation scares me.

Puppet is the more mature technology of the two, but (so I'm told) there's an important limitation in that there is no way to automatically e.g. reload the web server if its config file changes. Both these systems have some issues with information flowing in only one direction.

Once you get rolling with Chef or Puppet, updating your servers feels great - no need to manually fiddle with shell scripts, or wonder "have I installed that package on all my machines?"

Getting to that point can take quite a bit of effort though. In my most recent attempt to install a Chef server on a fresh Ubuntu box, I quit the installation halfway through as I realised I wanted to change some of the values I'd entered into the setup script. After re-running the script I encountered errors, and all my attempts at scrubbing the aborted installation failed. It was easier to just start with a fresh image. I had these sorts of experiences all the time when using Chef and Puppet, they were really frustrating to setup.

I felt more comfortable writing my scripts in Chef because it didn't limit my ability to use Ruby, and the more relaxed approach to handling dependencies suited me. Puppet was easier to setup, had better documentation, and behaved more reliably for me. Neither seems like a clear winner, but I haven't spent a lot of time with them in about a year now, so things may have changed!

I recently (past couple of months) had to look at chef vs puppet. I went with Chef, since it seemed easier to setup, and I liked the methodology behind it better. That said, there are a few issues around its choice of dependencies (solr, how I hate you).

Basically, I liked how chef treated the terminals as "dumb" and the configuration could be stored in source control. All of my nodes/"cookbooks"/recipes/users/data bags are stored in version control, with a few simple shell scripts to batch load the data. Also, chef nodes ping back to the server, which could cause problems (if you had tens of thousands of nodes hitting frequently), but also solves issues like "what if a server is offline when an important update should be pushed"

I believe that puppet has far more tools for seeing what's going on with your servers, when they last updated, that kind of thing. Also, the recipes/cookbooks you can get from the chef repository can be a little hit and miss (I had to heavily modify a few to make them work in my environment)

Puppet has invented a new language with its own syntax.

Chef has invented a new object system and execution model within Ruby.

Neither's approach is particularly good in this respect, but the latter is less onerous.

They're both straightjackets. Chef is easier to cut your way out of and then use as a normal jacket

Chef seems to be gaining much more traction, at least from my memory of blog posts, tweets, conference talks, and comments in the past 6-8 months. Completely unscientific, but take from it what you will.

Ditto, but I switched to learning Puppet tonight and I find it much easier to pick up.

High level:

Capistrano is a tool for deploying your applications. You use it to push out new versions of your own codebase.

Chef and Puppet are tools for provisioning/configuring servers. You use it to install the right third-party libraries, programs, config files, etc.

You can just take a look at the diagram, which is hosted at Gliffy: http://www.gliffy.com/publish/2886461/