Reported on the mailing list: https://groups.google.com/forum/#!topic/debezium/eXIbVGGLrDw.

The connector will go to FAILED state only after switching to log reading mode, change events from the snapshotting phase may get lost. The reason appears to be that we try to commit the latest offset with the database server, but the call to ReplicationStream#flushLsn() blocks indefinitely as long as we're not yet in streaming mode.