How many samples do we need to cover all bugs?

We have used 125K+ configurations and observed that some lead to build failures/bugs... 
It's an incredible amount of resources: retrospectively, can't we use a smaller sample and still find the failures/bugs? can we actively learn to prevent failures [1]? 
can we find failures that would justify to gather more data (builds)? 

[1] we did that last summer by putting some constraints we learned   