Avoid os.fork() prior to stats upload.

[kwlzn]

Problem

Presently, we do an os.fork() just before posting build stats in order to reduce the time-until-timeout for users who aren't connected to the VPN (and thus can't reach the stats server).

For users on OSX 10.11 (and potentially others), this seems to have been causing a silent segmentation fault in the forked process due to https://bugs.python.org/issue13829 .

With the addition of faulthandler support to pants, we now see a full traceback to stderr when this segmentation fault happens in the forked process due to retaining handles to the original processes stdio descriptors. At first this appeared to be a new issue encountered during release handling, but after investigating it seems like it's just faulthandler's improved logging revealing a long-standing masked segfault.

Solution

Remove the os.fork() call prior to stats upload and revisit.

Result

Can no longer reproduce the segmentation fault.

[stuhood]

Thanks Kris.

Maybe (re)open an issue about making stats posting async? As mentioned offline, doing something with a persistent queue or the sqlite db we already have, would be good.

[kwlzn]

sure, created #4920