> In this case, a blob would be a lot faster.

hm, it's a trade:)

> I'd guess you were getting very good compression on your utf8
> varchars.

i might be totally wrong here, but let me try to work on your hint.

so the sort table writes the big varchars in the sort table,
duplicated with each resulted record because then it reads them at
linear access speed (~20-40M/s), which makes sense whenever
writing/reading linearly is faster than seeking for the big varchars
after sorting (millions of records maybe).

but in my case, since i get good compression on the big varchars, then
the duplication overhead goes bananas up to gigabytes and screws up
the linear vs random read cost assumption.