A few years ago I picked up a bit of OCaml in order to put together a small, fast-launching GUI program. I had a lot of fun with that specific program, but soon afterward I dropped OCaml for general use because it was very awkward to pull in third-party code from outside the standard library. A month or two ago, a co-worker told me he was learning OCaml and mentioned that things had come a long way in the past few years. I was pleasantly surprised that not only had a creative-commons licensed book been written, but the library situation was vastly improved with the introduction of OPAM, a fully respectable package manager.
In my mind OCaml is an excellent complement to Clojure by filling many of Clojure's weak places. Startup time, distributable/runtime size, type inference, and C integration are all fantastic in OCaml but often unsatisfactory in Clojure, while true concurrency, library availability, flexible/clear syntax, and size of the community are places Clojure has the edge and OCaml is very weak in.
When we've polled Leiningen users in the past, the top complaint has always been startup time. Most people just keep a running repl in their editor, but this makes certain workflows awkward. My latest project, called Grenchman, addresses that problem by offering a fast-launching executable that connects directly to a running Clojure process and sends it code over the nREPL protocol.
I was able to implement the first release of Grenchman in a few weeks with about 400 lines of OCaml while basically re-learning the language from scratch with the excellent Real World OCaml book. My experience with the language has been very positive. Obviously the type system is the primary thing for which OCaml is notable, and with good reason. It's unobtrusive for the most part, and in my experience nearly every time the compiler complained, it was because of something stupid I had done. If you've used a static type system without inference, or even one that works on locals only, don't let bad experiences there put you off—the seamlessness of full Hindley-Milner inference is totally different.
The widespread use of the Option type (often called Maybe in other languages) stands out as a particularly helpful feature of the type system. Rather than allowing nil values to propagate through your program until someone tries to do something with them, any operation that could fail to return a sensible result requires you to deal with it explicitly before it will compile. It's occasionally a little more verbose, but it prevents all kinds of shortcuts in less strict languages (static and dynamic) that can obscure the root cause of subtle bugs. The certainty you get from consistent use of Option is similar to the easier-to-reason-about properties of referential transparency—there's a dimension of error-prone guesswork around failure semantics which simply evaporates.
That said, the type system isn't without its trade-offs. Any program that does communication with outside systems typically has "edges" where the compiler can't infer much about the data which originates outside those boundaries. In the case of Grenchman there is very little explicit mention of types, but acting on the Bencode-formatted messages sent from the Clojure server requires some explicit typing. Wrapper types are introduced in order for Strings and Lists and so on to be able to partake in the Bencode type, and these must be removed before the underlying values can be consumed. In more common serialization formats like JSON and XML, the wrapping and unwrapping code can be auto-generated from schemas, but Bencode is a more obscure format lacking in such luxuries.
The other place types must be made explicit is when interfacing with C functions. I used the Ctypes library in order to make calls to GNU Readline, and I greatly appreciated the ability to invoke that functionality directly instead of using a port. That said, I had a difficult time getting it to work, partly due to the slim documentation and partly due to my own unfamiliarity with the calling conventions of C libraries.
One of the things which drew me to OCaml initially was the ability to compile small, easily-distributable native executables. This ended up turning out a little differently from my expectations for a few reasons. Firstly the use of Ctypes meant that I couldn't statically link everything—Ctypes is built upon libffi and dynamically loads libreadline, making cross-distribution compatibility much more complicated. But most OCaml programs won't run into that problem unless they need access to C code.
However, Grenchman is built on the Core and Async libraries from Jane Street, one of the largest industrial users of OCaml. Async allows for monadic faux-concurrency that avoids a lot of the callback headaches of other event-driven tools, but it is fairly monolithic. This affects the size of the binaries emitted from OCaml's native compiler; even after running strip(1) on them they were still between 8.5 and 11MB. The book I was using cheats a bit and treats Jane Street Core as OCaml's standard library, which is nice because it results in code that's a lot more consistent and clean than it would be with OCaml's actual standard library, but the associated size trade-offs are unfortunate, and I wish they had been stated up-front.
While OPAM is fairly impressive and a huge improvement over the state of things two years ago, a couple things about it still bother me. One annoyance is that it's entirely source-based, so installing new packages can take a very long time. One impressive feature of OPAM is that it can handle installation and compilation of the OCaml compiler itself (and switching between separate versions), but this means pulling in a full dev stack for a given project can take as much as an hour. This isn't a huge deal though, and if I had to pick one I'd rather have a source-only package manager than the other way around. Creating a system that reliably deals in both source and binaries is very difficult. More worrying is the fact that it seems to encourage a "yeah, just get me the latest version of that" approach to versioning rather than defaulting to declaring explicit dependency versions. (It's possible to pin to specific versions, but this was only mentioned in passing, tucked away on the "Advanced" section of the docs, so I assume its use is rare.) This pattern is unrealistically optimistic in practice and caused lots of headaches in Rubygems. It was eventually be abandoned in favour of Bundler, but only after a very long and awkward transition phase. Hopefully the OCaml community can make this shift more gracefully than the Rubyists did by learning from their mistakes.
I can't say I've written lots of code in OCaml yet, but I've really enjoyed it so far. The extra help the compiler gives you allowed me to make some fairly major changes in Grenchman with a high degree of confidence, and being able to read type signatures in the libraries I was using usually made up for a general lack of documentation. If you've been thinking about picking up a new language, now is a great time to start with OCaml given the book and package manager.