Thanks for the detailed response, looking forward to the VLDB paper when it happens! Your vision sounds cool. I wonder if you could get most of the way there by exposing the workload (across different pipeline stages) to a materialized view recommender.

In the class project mentioned elsewhere, I found normalizing queries to be pretty slow in practice (naive standardized formatting + query templatization, tried various Python libraries, settled on pglast). I didn't think about trying "skip if fingerprint matches", which may help considerably. Fast normalization is nice! :)