Give some ideas and descriptions about an NLP machine learning pipeline that will process the reviews of online shopping platform, first create and maintain a $($random$)$ sample of the up$-$to$-$date large amount of reviews, then find and maintain the most popular topics, then able to classify individual reviews to the topics we found in the previous step. What are possible related sampling methods, clustering and classifying algorithm, which ones may be better than the others. How to know when we need to refresh the most popular topics list.