[Friedduck]

We serialized some XML as a backup to the data we were extracting (and properly modeling), given that the vendor was prone to changing the schema without proper notification, and that there were some data elements we weren't using at that time.

We also built the necessary tools to extract/re-process records easily, and the architecture worked well for us. As our needs or the schema changed we could easily accommodate those changes without undue effort.

It doesn't directly address the question, and I'm not sure that I'd use the same solution if the volume were predicted to be significantly higher, but in our case it worked beautifully. (Happily our volume was predictably within a known range, for reasons I won't go into.)