> 3. A RDBMS, with databases for both the raw and normalized data
I've found the filesystem (local or network, depending on scale) works well for the raw data. A normalized file name with a timestamp and job identifier in a hashed directory structure of some sort (I generally use $jobtype/%Y-%m-%d/%H/ as a start) works well, and reading and writing gzip is trivial (and often you can just output the raw content of gzip encoded payloads). The filesystem is an often overlooked database. If you end up needing more transactional support, or to easily identify what's been processed or not, look at how Maildir works.
After normalization, the database is ideal though.
That said, I was doing a few gigabytes a day, not a dew terabytes, so you might have run into some scale issues I didn't. I was able to keep it to mostly one box for crawling and parsing, but crawlers ended up being complex and job-queue driven enough that expanding to multiple systems wouldn't have been all that much extra work (an assessment I feel confident in, having done similar things before).