Hey! I'm tickled to see this on HN. I'm the author. If you have any questions, just ask. I'll do my best to answer them here.
Thank you! The thing I find tricky here is choosing the weight. I think maybe one obvious way you would want to weight samples would be for recency. E.g. if I have a table of user login events then I care about seeing more of the ones that happened recently but still want to see some of the older ones. Would the algorithm still work if I converted a `created_at` timestamp to epoch time and used that? Or would I want to normalize it in some way?