Complete separation means that every signed training margin is strictly positive for some affine score. In unpenalized logistic regression it causes coefficient norms to diverge and prevents a finite maximum-likelihood estimate.
New to topics? Read the docs here!