The answer is... its complicated, so you approximate. A good way of approximating is getting a dataset together and putting together a good model that helps explain that dataset. Doing the perfect Bayesian update in the real world is usually worse than nontrivial - its basically impossible.