One thing I noticed in looking at our current coverage report is that due to #2751 we are now missing coverage of a bunch of logging paths because we don't evaluate the log unless we are at a sufficient logging level.

This is good for performance, but it's kind of bad from a test perspective because we might now have crashing issues that we don't know about until someone turns on more verbose logging.

It's not completely clear to me how to fix this. We might need to think about how we do logging in tests. For example, maybe we always "log" at trace level but don't actually print logs unless the user has asked to see them at a certain level. This needs some thinking.

@jmarantz given your recent work on log testing, are you interested in tackling this?

cc @cpakulski