`rand` throws `InexactError`

[matthieuzone]

rand throws an InexactError when called on multi-dimensional DomainSets of Ints

Examlple

rand(1..10) returns a Float64 between 1 and 10
rand((1..10) × (1..10)) throws InexactError: Int64(random float)

If 1..10 is supposed to be continuous, I guess (1..10) × (1..10) should too.

Note that
typeof(1..10) is IntervalSets.ClosedInterval{Int64}
and typeof((1..10) × (1..10)) is DomainSets.Rectangle{SVector{2, Int64}}

suggesting that these are intervals of integers, so rand would maybe give an integer ? Or else, 1..10 should automatically convert it's type to float ?

[daanhb]

Thanks for reporting. It's a valid point that if a float is returned for the univariate case, one would expect a vector of floats to be returned in the higher-dimensional case.

This is what happens if you replace 1..10 by 1.0 .. 10 in your example, because in that case the eltype of the rectangle is a vector of floats. If you do that, everything should work.

So this seems an eltype issue, which you already hint at. The error comes from the toexternalpoint function, which tries to map a vector of floats (the random element) to a vector of integers (the element type of the domain).

Unless it matters a great deal to you, I'd rather not fix this. In this case the fix would be a hack, which is worse than the bug. For now it is best to always construct intervals with floating point endpoints. We are moving ("some day....") to a version in which we are more strict with element types, in which case rand(1..10) would indeed return an int (or an error) and rand(1.0..10) would return a float. When that time comes, this bug will simply go away.