What is the difference between implementation of 'drift_comparator' and 'skew_comparator' available in Tensorflow Data Validation?

Question

Need help some in understanding the difference between the underlying implementation of skew and drift comparator in TFDV.

Answer 1

Same thing, the only differences are:

1. If we set a drift threshold then a previous_statistics should be defined in stats validation. Whereas if we set a skew comparator we need to define a serving_statistics.

2. If we have an anomaly, with skew comparator we will get an anomaly message saying that there is a difference between training and serving datasets while with drift comparator the message will say that there is a distance between the two datasets.