I rewrote my experimental triplestore using the Mikio Hirabayashi's software. Right now, I've got the Subjects, the Predicats, and the Literals stored in a Tokyo Cabinet Hash, the Triplets in a Tokyo Cabinet Table, and my abstract Node object (which is basically a set of Triplet) in Memcached.
I tried Tokyo Cabinet mostly because of the benchmarks, and indeed, I confirm, it's fast !
But also because I don't need all the features of a traditional RDBMS, and it really feels good to get rid of this complexity.
Nice Sunday afternoon hacking.