Guidelines for data science teams — a summary of Daniel Molnar’s talks
In my opinion, his talks may be summarized as applying the good practices of Agile teams and “lean startups” to data science. He recommends striving for fast and good enough. He also strongly advises against over-engineering and using fancy tools.
I like that he encourages people to care about data quality and to choose KPIs that are challenging. After all, what is the point of measuring something just for the sake of feeling good?
In the data engineering area, he advises that doing ETLs in batches and building a data warehouse is enough, because making the ETL real-time or using streams is not worth the effort.
He also recommends an iterative approach to machine learning. First, deploy fast a good enough model. Later, build the model offline and redeploy each quarter. It looks like a suggestion to avoid the hype and focus on solving real problems. If you have been reading my blog for some time, you know that I love this approach.
For me, the most essential idea from his talks is the “Friday 17:00 test”. In short, it tells you whether your data product (dashboard, report, recommendation, prediction, etc.) is actionable. Daniel Molnar suggests asking a question:
What can a person do with this result if he/she gets it at 5p.m. on Friday?
If the answer is “nothing, at least until Monday,” maybe you need to rethink that product.
You may also like