How should a memory-intensive step be configured to run in parallel via Apache Beam on Google Dataflow?

Question

I\'ve got a dataflow pipeline that processes some satellite imagery. This pipeline works as expected with lots of small test images, and it also works when processing a sing