> Now, I will re-write my combinators and see which style better suits my needs.

Finally, impressed by the speed of the record solution, I have
re-written the whole library with records.

I have measured the performance of this implementation against the
Ref-based one on "real" code. Thus, I have simulated my Chord [1]
implementation with a simulator instrumented such that it drops the
processing time for each arrow processing.

The results can be found at [2].

In a nutshell : the record solution has a negligible overhead compared
to the Ref solution.

Thanks all,

[1]: http://pdos.csail.mit.edu/chord/
[2]: http://perso.eleves.bretagne.ens-cachan.fr/~dagand/projets/opis/processing_speed_records.pdf

-- 
Pierre-Evariste DAGAND