Currently, each user model is retrained every time a new observation is made for that user. It turns out that doing a lot of concurrent user-model training kills throughput. Doing some basic scheduling for user-model retraining may be useful, and I suspect that retraining on every new observation may not be necessary.
Currently, each user model is retrained every time a new observation is made for that user. It turns out that doing a lot of concurrent user-model training kills throughput. Doing some basic scheduling for user-model retraining may be useful, and I suspect that retraining on every new observation may not be necessary.