I'd put the extra columns in their own data frame, and save that to disk
(use dates/times/process ids or some other unique identifier in the
filenames to distinguish them). When you need access to a mixture of
columns, load (or source, depending how you did the save) the columns you
need, and cbind them together into one big data frame.
If you are concerned about memory requirements when producing the pieces,
watch out that you don't write out so much data that you'll never have
enough memory to load all you need at once.
Duncan Murdoch