tests do pass, the marked failure was a transient failure in a grpc test. in general i think that running using xdist will put pressure on our test suites -- tests which fail only under weird circumstances will fail more often. this is both good (in that edge cases that might appear under production loads may be more likely to surface) and bad (in that we may spend cycles refining test harnesses to solve issues that won't arise in prod)
speeding up just the core tests doesn't effect BK total time - do you run tox locally? I always just pytest directly.
Honestly I will take this diff just for updating snapshots and deps - but will gives some time for others to chime in on xdist use itself
nit: extra spaces after API
just confirming understanding here - we established that this only can actually happen in tests that use CliRunner.invoke right? (i.e. it's a test-specific issue - users shouldn't be synchronously executing command-line options in Python)
Generally it should be safe outside of tests to assume that a CLI command is on the main thread when it starts
The changes to the migration test setup here should make them safe to run concurrently (in python_modules/dagster/dagster/utils/test/__init__.py and the back_compat tests). In general, this would probably be more fragile than the single-threaded running, but like I said above, that's good and bad.