LSM thread pools – with N LSM trees, we currently end up with 3N threads by default. This won't scale well to 1000 tables. Have a pool of LSM threads that share the trees, use a calculation like the current throttle calculation to prioritize flushing and merging in different trees.

This could be with named groups like shared cache, but I would like a simple way to configure a pool of threads that operate on all LSM trees in a database.