Hi Nicolas, the params you have are correct, the lowest hanging fruit is spatial/temporalImpl: cpp, and #columns not too high - both checked in your case. I remember sending a similar rant to ML with anomaly a while ago, I can't find the numbers now but i think I had something like: start 100Hz (100rec/sec) med 30Hz (still acceptable for me) and gradually getting cca 7Hz (? not 100% sure now)
Scott confirmed this trend, and that it will logarithmically converge (so hopefully not the 900yrs you mentioned)..but it probably wouldn't get faster either. Is it similar to the numbers you're getting? Mine was cca 100+k dataset. Cheers, Mark
