Thomas Munro <thomas(dot)munro(at)enterprisedb(dot)com> writes:
> I found 3 examples of this failing with an ERROR (though not turning
> the BF red, so nobody noticed) before the PANIC patch went in:
Yeah, I suspected that had happened before with less-obvious consequences.
Now that we know where the problem is, you could probably make it highly
reproducible by inserting a sleep of a few msec between the rename and the
second fsync.
regards, tom lane