Not just in polling, but on anything that uses random samples, such as machine learning.
The idea is to segment the population into groups (aka strata) and then get the sample based on that proportion.
Researchers rely on the same properties for randomness in each subgroup as they would in the entire sample. In fact, this approach results in greater statistical precision as long as the thing you're grouping by has some correlation with the outcome variable.
IIRC this can also done after the fact: researchers will use different weights to make sure the proportions roughly match the population distributions for a certain characteristic.