Can you use feature datasets as logical folders to store feature classes?
There is an overhead associated with feature datasets, as actions that are taken on one feature class in the feature dataset are taken on all feature classes in a feature dataset. For example, if you want to edit one of the feature classes in the feature dataset, you must register the entire feature dataset as versioned. This means that every feature class will have an A and D table created. Any time data is registered as versioned the queries become much more complex. When you begin an edit session for a feature class, all of the feature classes in the feature dataset will be queried. This results in more queries being sent to the database and subsequently more processing occurring on the server.
It is not advisable to use feature datasets for logical groupings of feature classes. Feature datasets should only be used to group feature classes that have a topologic relationship. If you use feature datasets as logical containers, limit the number of feature classes you place in the feature dataset. Generally, grouping 20 or fewer feature classes will not cause a major performance impact.