My azure job manager I mentioned last time worked rather nicely. I found myself running into various usage limits, which the people on the service desk happily removed for me, until I railed up against the public IP limit. Turns out I'd been misunderstanding how to structure my queue. I have now rewritten the code so that instead of launching 60 virtual machines with 60 unique IPs, I now have a 'head node' of sorts, with a public IP, and a virtual network of virtual machines hidden behind that one. I just submitted a big run using this new queuing system. We'll see how it pans out.
Our department was also visited by Iain Murray of Edinburgh this week. I had seen him speak a couple of years ago at the intractable likelihood conference in Bristol, and then more recently I bumped into his PhD student at a workshop in Warwick (who, as it turns out, also works on emulation of physical simulations), so I suggested him as a speaker. He gave an excellent talk about density estimation and neural networks in astronomy contexts, and we had a lot of interesting talks about my own work, and he provided some useful suggestions regarding Gaussian processes.
I also started reacquainting myself with the Gaussian process work flows I had developed before the COMPAS writing workload piled up, and I'll spend the next week working on these, as well as nursing simulations and data generation.