I know it depends on the analysis, but I often am doing somewhat embarassingly parallel things. So just knowing GNU parallel for mid-scale things (and R/python basically parallelism, although shared memory is a bear), and how to temporarily scale across the cloud to like 500 core, is huge.