I have very large table groupedby_x_y (2B rows) with the columns:

date,
x,
y,
z_count

This is the result of counting this z value from a more raw table and grouped by columns x and y. Turns out we are also computing from the underlying raw table another table groupedby_x (500M rows) which has the columns:

date,
x,
z_count

The way I see it groupedby_x should just be groupedby_x_y with z_count summed over y:

recover_groupedby_x = 
SELECT date, x, sum(z_count) as z_count from groupedby_x_y
group by x

But when I do this, the two tables groupedby_x and recover_groupedby_x are nearly identical but not quite – they differ by about 5M and I cannot figure out why. The groupedby_.. tables are perfectly and fully rectangular with no nulls or funny malformed data etc.

Is my idea just wrong or might I be missing something a little subtle here ?