Clojure is a Lisp for the JVM, so in a sense they are using Java. Besides, Clojure has been used in several high-throughput data analysis applications (FlightCaster for one), so it's not exactly an "unproven second-class language".
Flight Caster or the other company which google bought use huge data but with that they also use huge amount of rules. They are developing rule systems and engines and not Not a data mining system.
The company in the post has a machine learning problem, what they require is fast matrix multiplication and great mathematical models and not some weird ass FP language ideals.
Go to mloss.org and look for projects that use FP languages.
in real life no one gives a fuck about lisp and other FP idiocacy, that why AI book by norvig has now code written in Python, thats why Google being one of the most leading AI companies uses C++ Python and Java.
Thats why you wont find any one in any data mining/ mahcine learning department in any US university MIT/CMU/Stanford using some shit FP lanuages.
Beacuse thats not their core business!
nor spelling correctly is mine.
Good developers (who they are trying to hire) do. This is a recruitment pitch. In that context, "language ideals" are important.
user: MADMEN
created: 38 minutes ago
a) Doing data mining research and exploratory programming in a small company is probably the best use case for a Lisp. When you're Google you can afford to write this in C++ or Java (with layers of libraries taking the boiler plate out) as well waste research/prototype time on compile/link cycle (rather being able to program interactively). When you're Google, you can also afford to only hire machine learning people who know how to write production quality C++ (very rare combination of skills). Clojure is a more acceptable language: despite of what fanboys say, FP is not a black art reserved for 99.9th percentile programmers; the more C++ I write, the more convinced I get that C++ is.
b) Clojure interop with Java makes it easy to use this sort of system with Hadoop, Mahout, Colt, etc... Sure there are more effective libraries available in C++, but they either proprietary to Google (MapReduce + GFS), not designed for fault tolerance (MPI) or more time consuming to use (LINPACK, LAPACK)
c) Python is going to be 10-20 times slower, (perhaps 5x when you're using C interop).