Enjoyed the read. Looks like you’re up to some very interesting stuff.

I’m surprised you got better performance out of pandas than SQL. Do you think this was a technology limitation with SQL, or was your team just more comfortable doing the processing in Python?

Did you end up doing a columnar store for your DB? I’ve read good things about Redshift.

Have you thought about splitting your database into 1 row-store RDBMS for user data, etc., and 1 column-store RDBMS for your analytics service?