During ALS model load, serving can slow down due to solver recomputation
srowen opened this issue · comments
As new data arrives, the AtA solver in ALS must in theory be recomputed regularly. This is an intensive operation and constantly recomputing it consumes a lot of CPU. One way to improve this is to prioritize building the solver with all cores when none is available, but, otherwise treat it as a low-priority task that just uses one core.