[k__]

Their main idea was that B-Trees are slow and LSMs are fast.

This was a partially right assumption, but only for writes.

If you write something in a DB you check some constraints and those checks are reads.

So most DB writes come with a bunch of reads.

The reads were slower with the LSMs, so the B-Trees performed better in "real world" writes (which come with reads) and LSMs only performed better in "artificial" writes (without reads).