Yup, these results are pretty consistent with what I'd expect (& why I noted the impact of indices) cause even string data has a lot of superfluous information when expressed in the DDL ("INSERT INTO foo ...") - I would expect all of that to exceed any bookkeeping within the btree. And non-string values like blobs or numbers are going to be stored more efficiently than in the dump which is a text encoding (or even hex for blobs) which is going to blow things up further.