Looking through the available config options, I don’t see how we can limit the LSM footprint in RAM. As an example, running a server of 50GB of RAM, we may want to reserve at most 20GB to the key space (LSM tree - all levels). MaxTableSize only caps the base level.
Loosely related to that question, I assume stalling is a “stop-the-world” event, meaning no gets/puts can succeed? We are trying to understand how the NumLevelZeroTables threshold would be reached which would create the stall, and more importantly understanding roughly how we can determine how long the stall lasts.