These terms are borrowed from scikit-learn, and the idea is the same. A transformer is just a component of a pipeline that transforms the data in some way. An estimator is also a transfomer, but one that additionally needs to be 'fit' on data before it knows how to transform.
For example, a StringTokenizer is just a transformer, because it does not need to see any data to know what to do, to tokenize strings. A machine learning model like LogisticRegression is also a transformer, because it transforms data by adding a prediction. However it has to be fit on data first before it can do so. So it is (also) an estimator.
