This pattern has been huge for us. It means we only have to do data munging / feature generation once - in Looker. We can write model outputs back into Redshift, and users can see model outputs right next to model inputs.

Any way to hack this so that if there’s a day with 0 orders - and therefore no row in the query results - the rolling 7 day average is still accurate?

+1 for this. Also, the ability to have a dimension vs. dimension scatter plot, and include a 3rd/4th/etc dimension in the tooltip.