@Gary_Tang I have noticed same when triggered spark streaming app using 10.3 connector. Also found corresponding jira where others reported same. Below links if helps, unfortunately its not documented on mongodb docs. For existing data i can do a first run as batch mode which is not a concern, but wondering how one is able to perform maintenance activity on spark application using this connector while events are still coming into mongodb without stopping spark app.
https://jira.mongodb.org/projects/SPARK/issues/SPARK-303?filter=allopenissues