In the changestream case, we can continue to advance the resume token even when no documents have been read from the stream.

However in the kafka case, if we're not reading documents, there's no point in writing another checkpoint, the offsets have not advanced at all. So the main goal of this work is to avoid writing new checkpoints for Kafka $source when they are not useful (when we have not read any documents since the last checkpoint). 

There are some other things to consider below, but we should probably leave these for a follow up ticket https://jira.mongodb.org/browse/SERVER-82440:

and/or sizable amount of input has been processed

and/or oldest seen input message is quite old