VPD Performance, Best Practices for large datasets

Hi,

Is there any statistical benchmark on VPD Performance for large datasets? I understand this is generic and more development theory question.

I am exploring using VPD for RLS for my MVs that are in millions of records and we need to reduce it to information based on user's id, or role.

Is there a development best practice whereby a system account is accessing the data and calling a function which executes a procedure that switches the context to the userid passed to the procedure as a parameter and returns a Cursor.

This is a limitation for user experience since I am integrating with an external analytical system.