Well, these machines, that I got for free, are getting a little long in the tooth, so I’m starting to come to terms with that–the single CPU performance still seems better than a DO droplet, though. If solving the problem with money in the short term were an option, then I’d likely not have this particular client who wants enterprise performance at business prices. 
But I also see that I had the number of unicorns hard-coded somewhere else in the chain, so I’m still running only 3.
Sadly, my current configuration, only talks to docker on the one host. I should spend a bit more time and see if I can stick a couple unicorns on the other machine too, though. Probably about time to look at HAproxy again, but I’ve got another project that I really want to get launched first.
Thanks very much for your insight.
EDIT: And when I finally moved to 5 unicorns rather than 3, the performance graphs look about the same (but maybe a tiny bit slower?), but the 429 errors dropped significantly. It looks like once the image processing is done, this is going to work just fine. 