Closed tbekolay closed 5 years ago
Small update for posterity: to work-around this temporarily, we partially reverted the above commit (e9cd851b01b20b774d58c82a2ba775b531cf2bab) and switched to -n 1
py-test workers in #202. But we're still seeing a different related issue (os.chdir
not changing the directory back in the hardware interface's __exit__
, and then coverage.xml
gets put into the wrong place) if -n
is omitted altogether.
Still investigating why the directory is not being switched back. This happens consistently on the same test: nengo_loihi/hardware/tests/test_allocators.py::test_block_size
.
The above problem is solved in #218. The larger issue described in e9cd851b01b20b774d58c82a2ba775b531cf2bab is still at hand.
This seems to have been working OK since https://github.com/nengo/nengo-loihi/commit/ee0ca71fc341685865d26404cc5745bda070e89e#diff-bb7e722a615e3b3171b86e0b952a51fc (hard to say for sure with concurrency issues, but it's at least a pretty rare issue which can be resolved by re-running the build if it ever comes up again).
In 70ecb8a475985846c81bc5e8059056042f855a34 we started running hardware tests in parallel, and made a quick patch to ensure the C file for the snip could be access by two threads at once. It seems that this can still fail so we should try to reproduce that issue and fix it.