For DESC, I have built an "online coaddition" class that is a replacement for afw.statisticsStack written in pure python/numpy that can do mean and clipped mean stacks using a running sum. This results in significant memory savings vs statisticsStack and does not require reading sub-regions. Tests on ci_hsc show it is 15-20% faster than statisticsStack for mean coaddition (due to reduced overhead), but it is slower for clipped means (in a way that scales with number of images) due to the increased i/o.

This code currently lives at https://github.com/lsst/pipe_tasks/tree/u/erykoff/onlinecoadd . The AccumulatorMeanStack class will be put in meas_algorithms, a more appropriate location. Changing to the online coaddition code in AssembleCoaddTask will be configured with a doOnlineCoaddition config option that will only apply to "MEAN" and "MEANCLIP" coadds.