Tomi Ollila <tomi.oll...@iki.fi> writes: > On Wed, Dec 09 2020, David Bremner wrote: > >> Under a sufficiently high level of parallelism [1] there seems to be a >> a race condition that allows sphinx-build to start running before the >> docstrings are extracted. This change moves the docstring stamp from >> the phony targets sphinx-html and sphinx-info to the file targets that >> they depend on. I'm not sure why this makes things better, but I am >> fairly confident it does not make things worse, and experimentally it >> seems to eliminate the race condition. > > Good enough for me if this helps. I also don't see reason why that would > make things better (and probably not things worse), just that I got > headache reading that Makefile ;) (and, for example, these particular > targets mentioned would not need to be marked .PHONY...) > > I'd suggest to monitor the behaviour for a while and if things are > consistently better then merge -- such a things when we don't know > enough experience may be enough (or someone(tm) could try to parse make > debug logs ;/) -- or someone(tm) may tell us why that heleps :D
I'll try a test upload to Debian and see if that fixes the problem there. I did 2000 builds (sortof by mistake) at -j32 on a 16 thread / 8 core machine, with no crashes. Much less parallelism than the original report, but lots of repetition. d
