Thanks for you comment Josh. In production I actually implemented PG version and it's still in use today. I thought it will scale better and last longer. I didn

Thanks for you comment Josh. In production I actually implemented PG version and it's still in use today. I thought it will scale better and last longer. I didn't do benchmarking or anything like that - it think it would really depend on size of you dataset. For something serious - compute recommendation in background and store them in database. I believe that is how "big boys" do it. :)