And the fact that they're spending so much time on it and were afraid to rewrite it tells you all you need to know.

Rewriting it in C++, leveraging the platform-specific APIs for disk I/O (sync_file_range on Linux, not even wrapped by a Python module) would yield not only much better performance, but also much higher reliability.

As it is this daemon is a recipe for page thrashing. Not that I would expect a Python dev to actually know how a kernel works.