I'm currently working on a new streaming system for a big data real-time-ish application (which is why we can't use existing libraries). We've come up with a design that is very similar to that described in the paper, as I understand it on a quick skim, but we don't use MSP and hence can't perform deforestation. We did actually consider using LMS but decided we weren't justified in doing so without getting some benchmarks first. I think we'll hit our targets without going down the staging route, but it's good to have prior work to build on if we need to!

Linear algebra is the other obvious-to-me application of this. Linear algebra libraries typically either create a lot of unnecessary intermediate arrays or provide ugly imperative APIs (hi BLAS!) and force that busy work onto the programmer. Deforestation is an excellent way to regain performance while still maintaining ease of use. Applications include deep learning, and this is broadly the approach taken by TensorFlow and other libraries.