In MIT's Mathematics for Computer Scientists course (6.042), I remember learning from Tom Leighton that in certain cases randomly distributing jobs actually IS the optimal solution to load balancing. This technique was somehow associated with Akamai, Professor Leighton's company.
I think the matter is if random distribution is what is documented, no one will raise an eyebrows. I believe the Heroku claimed intelligent routing based on server loads to route traffic to idle servers. That threw people off.
I dont know, but I would assume it has to deal with the fact you can't predict how long a job will take to complete, and a random selection would have more resources-more free-more often, verses FIFO where a group of really large tasks could halt everything.
And what if they turned the whole thing into a request queue and workers get the "messages" as they become available? To ensure there are no timeouts, they could define a max time a request may stay in the queue, and then take it out and handle it with emergency workers or kill it?
Maybe I am misremembering...