My main fear is maintainability. At some point in future, you may want to do some serious changes in the way data is structured - possibly even changing the relationships/class hierarchy etc. With relational backend, you are already separated a bit (because of explicit mapping to OO structures), but, what is most important, you can as well write some migration procedures for data. With serialization, it can be a major pain.
I have used serialization for long term storage in the commercial project and while it has speed up the development probably twofold (and allowed us to finish the project in required time with given resources) we started to be hit by it in major way already few months after going into production. We ended up with classes which had like 3 alternative slots for the same data, expecting nulls in every possible place even if it was set to reasonable value in every possible constructor, etc, etc. All this within 1 year after going to production, I'm worried what tricks they have to do currently to maintain it... As a side note, we had a major problems with customer data security department - after project was put to production, they have requested they should be able to read every single value in their SQL tools...
For data mining, I'm not really thinking about anything in particular - I have never run MMORPG, but I know that with old time mud, I used to grep through player files for amount of gold/level
I suppose that any reasonable MMORPG has to run a lot of statistics to get the feedback for balancing process - class/race distribution, amount of quests done per day, wealth levels for players, most often equipped items, etc. While some of the events can be probably logged separately to dedicated file/database, persistent data like equipment and players have to be queried from live data.