In reality both diagrams will turn out to be the same: Metadata + Caching = ETL + DWH.

The ETL logic is complex because the transformations to turn the raw data into KPIs are complex. When removing ETL from the picture, the transformation logic does not go away, it just happens at a different time (e.g. at query time) and in a different engine (e.g. SQL).

Removing the database as engine for fast joins, quick searches, complex aggregations requires another engine to step in. Simply applying brute force (bigger clusters) helps somewhat but the difference in user expectations (sub second query response) and the limitations of distributed computing (n-way joins that need reshuffle) is unsolved.

And yes, it is the end of DWH. Only structured data, daily refreshes, aggregated data,... that is no longer sufficient.