interesting to having recommendation function as part of SQL query. Not sure how much latency/delay is added executing recommendation query vs standard query against a big dataset.
I guess the latency depends on how much the recommendation data fits in main memory. However, based on my understanding RecDB implements the recommendation algorithms as a physiccal operator which expected to be much more efficient than implementing it as a stored procedure or a user defined function.
A relational database system can scale to few terabytes of data, which is much more than currently available recommendation datasets. However, based on my understanding, if we are talking about Petabyte scale or more, I don't think RecDB is the right option !