Generally we will prefer stratification. The error of the estimators is always much smaller.
In any case we must know exactly, the probability of inclusion of the sampling unit in the sample.
In all but the first case, weights must be computed for the estimates to represent the total population, since the units have different probability of inclusion.