multiprocessing.Pool.imap() is supposed to be a lazy version of map. But it's not: it submits work to its workers eagerly. As a consequence, in a pipeline, all the work from earlier steps is queued, performed, and finished first, before starting later steps.

If you use python3's built-in map() -- aka the old itertools.imap() -- the operations are on-demand, so it surprised me that Pool.imap() doesn't. It's basically no better than using Pool.map(). Maybe it saves memory by not materializing large iterables in every worker process? But it still materializes the CPU time from the iterables even if unneeded.

This can be partially worked around by giving each step of the pipeline its own Pool -- then, at least the earlier steps of the pipeline don't block the later steps -- but the jobs are still done eagerly instead of waiting for their results to actually be requested.