Instead of setting a predictable seed, randomise the data to write,
as well as the offsets to which to write. A failed run can then
be replicated by reusing the seed mentioned in the error message.