The downside is that projecting a row requires random I/O across a larger number of pages, which also means more evictions from the in-memory buffer and worse c

The downside is that projecting a row requires random I/O across a larger number of pages, which also means more evictions from the in-memory buffer and worse cache efficiency. Apache Arrow, Parquet, Redshift, Bigtable/Spanner, Snowflake are all hybrid columnar, for example. You get good row data locality while still being able to exploit SIMD/vectorized ops.