So you want a LSM for inserts and the DNN for reads? Seems OK. You still have to update/retrain the DNN after an insert into a larger layer, which will be expensive. So you’d probably get high latency at the 99% (or some high number).
There are no inserts into larger layers, only merges. Which are long (usually processed in background by separate thread) and that longness justifies training a new net in parallel to merge process.
