These counts were transiently correct if the client was slow enough, so the tests usually passed, but actually they were wrong.
These counts were transiently correct if the client was slow enough, so the tests usually passed, but actually they were wrong.