I've seen this test fail a bunch lately - mainly when the tests are all run (i.e. not individually) but not in parallel, it seems. If you dump out the sources when it fails, you see:
- The sources are numbered in the hundreds, so it's very unlikely that this actually gets the first one.
- The sources are numbered twice. There was logic to have the first one not be numbered, but that got messed up and now all sources are numbered twice, but the first one is only number once.
We could just remove the bad assertion, but then we're only testing the registry and not anything else about the way metrics flow all the way through the whole system. Worth it to fix the failing test, I think - knowing the source gets registered doesn't add a whole lot of value toward end-to-end metrics testing.