I imagine their systems have changed a lot, but 10+ years ago I heard the key reason for not implementing this at the time was due to how heavily denormalized and append-heavy their data model was. (Though they always had a mechanism for deletion, so..)