Currently, if log rotation fails (e.g. due to destination file already existing), then the server will always abort even though it could theoretically continue with the old log destination just fine.

Report the failure to rotate in the log, but do not invariant (unless the cause was actually something critical).