I'd like to call out that SAP Hana is a HTAP database that does store the data in a columnar format only(!!!) and still achieves ACID guarantees and high single row performance.

This is done by focusing on memory structures.

The entire row vs column storage is related to block storage. If one row sits in one block, it is read much faster than having stored the data per column, where one block contains multiple rows of the same column and for one row with 100 columns, 100 blocks must be read hence. If a OLAP query is aggregating a single column, having data of a single column in one block speeds up that.

The entire argument does not apply if the query is answered from memory. For RAM it is equally fast to read data from a nearby memory address or a far away. It is the same thing.

Furthermore, for OLTP queries you usally use the hot data, hence keeping the relevant data in memory is an option.

Think about a spreadsheet. What can be read faster, one row or one column? If the rows are organized horizontally or vertically? We are used to reading left to right, hence we have a cultural preference, but purely from a technical point, it does not make a difference.

Also note that Oracle does the same thing as Snowflake. They have a row format on disk and if activated, store the data a second time in columnar format to speed up OLTP queries.