Typical LSM key-value databases like goleveldb are unpredictable in throughput. They can write a big batch of data into L0 but then get 'stuck' until all so called 'compactions' are finished. In my application, a database access is shared between several users, and I'd like to balance the write throughput between them. But some 'greedy' user may 'hang' the database for all the others, I'd like to avoid it, but don't know a simple way to do so. Of course, the simplest way is a hand-tuned limiter constant, but I'd like my application not depend on particular storage setup. Goleveldb seems to provide per-level statistics that in theory can be used for some sophisticated math filter that can derives that limiter constant, still it does not seem to be a trivial task to make such a filter. But in BadgerDB, which looks like a good replacement for goleveldb, there seems to be even no way even to get that statistics.
How other people usually solve such a problem with database 'hanging' when predictale responsivity is needed?