Two issues:

1. Unimported blockers may not be returned.
2. We may issue a BuildableQuery with an empty list of buildable PHIDs.

I can't actually test to see if this fix resolves an issue since it occurs at work, and I won't be back at work until next Wednesday (I work part time).

This scenario should only really occur if $parents returns an empty array, but under what scenario would this be an empty array? (unless it's the first commit in a repo, but I don't think this is the case)

Is it possible the result of 'diffusion.commitparentsquery' is impacted by the new caching code?

I suspect this is occurring because parents have not imported yet.

There is no guarantee that commits are parsed in order, so it's completely expected that a child may fully import, trigger builds, and arrive here before a parent is imported. For example, the parent may be much larger (take longer to import) or encounter a temporary or permanent error.

The parents array will be empty for the first commit in a repository, as well.

(Recent changes haven't caused us to fill the parents array out of cache yet, so behavior on that side should be unchanged.)

(I'll take a shot at at least running this code before I land it.)

I sanity checked this locally, at least.