Calculating Rolling Sum But Can't Remove Duplicates

// First of all, you should not store all data
// in one table. That's not only Bad Practice...
// it's also dangerous and leads to subtle bugs
// you'll not be able to spot. Please ALWAYS
// use correct star- or snowflake-schemas.
// Second, you are counting same users many times
// because you're iterating dates and for each
// such date you are doing a distinct count
// of the users... but on the currently iterated
// day. What you should do instead is you should 
// select the whole period at once and then
// do a distinct count of users.
// The measure should most likely be only shown
// when only one day is visible in the current
// context, therefore it must check for the
// number of days visible.
// The most important table in all models is
// a date table (Calendar). So, please make sure
// you're doing it RIGHT.
// If you have a CORRECT model, the calculation
// proceeds as follows:

[7D Rolling Count] =
var __oneDayVisible = hasonevalue( Calendar[Date] )
var __lastVisibleDay = max( Calendar[Date] )
var __periodToCountOver =
	datesinperiod(
		Calendar[Date],
		__lastVisibleDay,
		-7, day
	)
var __result =
	calculate(
		distinctcount( FactTable[UserId] ),
		__periodToCountOver
	)
return
	if( __oneDayVisible, __result )

If you want to learn a bit about CORRECT MODELS, you can try these:

https://www.youtube.com/watch?v=78d6mwR8GtA

https://www.youtube.com/watch?v=_quTwyvDfG0

Creating DAX on INCORRECT or MESSY models is not only difficult. It's also error-prone. Good model = simple, fast DAX. Bad model = complex, slow DAX. Easy as that.

Best

D