> With this design, it’s possible to run native SQL selects on tables with hundreds of thousands to millions of columns, with predictable (sub-second) latency when accessing a subset of columns.

What is the design?