>That's 3 ms per second = 0.3% CPU time wasted for every waiting thread.

I suspect that's actually "per process, per database (usually 1)", and not based on number of threads or tables. `data_version` semantics mean there's no need for more than one connection polling it, and it's being used as a relatively lightweight "DB has changed, check queues" check (that's pretty much its whole purpose).

Also I believe this is mostly intended for multi-process use, e.g. out-of-process workers, so an in-process dirty tracker (e.g. just check after insert/update/delete) isn't sufficient.

So I do think it's somewhat crazy, but it is at least very simple. fsnotify-like monitoring seems like a fairly obvious improvement tho, not sure why that isn't part of it. Maybe it's slower? I haven't tried to do anything actually-performant-or-reliable with fs notifications, dunno what dragons lie in wait.