We were able to solve this. We’re not sure what did it, but we modified the linger.ms and batch.size settings for the Connector. We also found a very inefficient process that was putting quite a bit of load on Mongo every couple of minutes and improved that system.
I also posted an update on the JMX metrics thread.