Nope - delta store is not a row store.
Concerning the question: sure, it's fairly common to have to update data.
However, back posting 20 Mio records is quite a huge batch to go for in a single command.
One approach to deal with this would be to perform the update in several batches, so that the delta merge can actually work on the table in between.
This would likely be not as quick as running the update in one go, but it would allow to finish it given your memory restriction.
Another thing to consider: the big fact table is already a billion records in size.
You might consider partitioning to make it more manageable from a memory/delta store perspective (and also in terms of parallel insert/update performance).
- Lars