(I haven't tested this yet, though)

As the Python 3.8 version of this, I'm not sure how well this will play with 2.7 and 3.4, at the least.

Yep, unfortunately, the fix is python +3.7. I'd say we start recommending switching to python 3.7 if the workflow is memory-hungry.

Given the recentnumpy bug, I'm hesitant to suggest moving scientific applications to 3.7 until upstream projects have had a bit more time to find and fix critical bugs.

How hard would this be to port to 3.6? I suspect if we get it working there, 3.5 and 3.4 will fall into place quickly.

I'm okay with abandoning 2.7, as long as we don't break current usage further.

Some thoughts:

• As per your link, anaconda fixed the bug in numpy 1.15.2, and it is fixed with 1.15.3 withPypi. I guess we should update the pinned versions if python == 3.7.
• I'm not proposing to discontinue python <3.7, I would just make this fix available for python +3.7.
• If nothing of the above works, then I'd say it should not be too hard to backport to 3.6 or even 3.5 (prior to 3.5 it would be harder). Although I'm not sure it is worthwhile.

Codecov Report

Merging#2773 intomaster willdecrease coverage by3.4%.
The diff coverage is100%.

@@            Coverage Diff            @@##           master   #2773      +/-   ##=========================================- Coverage    67.5%   64.1%   -3.41%=========================================  Files         340     338       -2       Lines       43271   43223      -48       Branches     5364    5362       -2     =========================================- Hits        29210   27707    -1503- Misses      13360   14446    +1086- Partials      701    1070     +369

Continue to review full report at Codecov.

Legend -Click here to learn more
Δ = absolute <relative> (impact),ø = not affected,? = missing data
Powered byCodecov. Last update199d9a6...b220c86. Read thecomment docs.