What Happens to Running Full-Load Tasks During a DMS Instance Scale-Up/Down.

In scale-up/scale-down solution for DMS Replication Instances, there are multiple instances, and each instance is monitored via CloudWatch Alarms. When an alarm is triggered, a message is sent to a topic that in turn calls a lambda to scale the instance appropriately. Eventually, the alarm sources DMS instance is either scaled to a smaller or larger instance.

We are attempting to determine the effects of the scale-up/down occurring while an attached task is in the middle of a "Full-load". The results appear to be inconsistent. Here is what has been observed.

1. Restart a full load task attached to a target instance.
2. Trip the alarm for that target instance via set_alarm_state.
3. Very quickly, the target DMS instance state changes to "modifying" and the links in the UI are greyed out.
4. While the DMS instance appears, in the UI, to be in a "Modifying" state, the Full-Load Task continues for another 5 minutes and completes.

Is this just happenstance?

Can anyone confirm that a state change from "Available" to "Modifying" for a DMS Instance is smart enough to wait for active full-loads to complete before rebooting or is this just some sort of latency between state changes and the UI?

It would be nice to have an explanation of what happens when a 30-minute full load task starts and the associated DMS instance scales up while that task is still running.