Replacement algorithm seems to be slow, and the g3 system is suboptimal
The g3 recsys seems to be very long when responding to a rec request (like 30 seconds).
We need to optimize it by precomputing recommendations.
And we also ensure the on-the-fly replacements (by a random picked baseline) works within 2 seconds to prevent the user to have a very long wait when refreshing.
Maybe we can disable the g3 system for now?