To me, the biggest advantage of MongoDB is autosharding. If you read the friendfeed article, you'll see that they also have to deal with 'eventual consistency' on the indexes.
I'd rather partition the data based on function onto distinct clusters, you know like eBay do.
You can update the indexes in a transaction too, so that's not necessarily an issue. MySQL has problems with this due to locking but the MVCC implementation in PostgreSQL allows much better concurrency.