Autoscaling scale in policy working

Some things to consider:

Since you have configured target tracking on CPU utilization, check the Cloudwatch Alarm TargetTracking-xxx-AlarmLow-<random_string> (or something similar) for the Condition defined. The alarm is not editable and the condition for me is defined as being < some threshold for 15 datapoints over 15 minutes.

Normally the LB deregistration delay is 300 seconds and is defined in target group attributes.

And often there is a slight delay for datapoints / metrics to be reflected in cloudwatch; somewhere around 2-3 minutes.

If you add add this up, it takes around 22-23 minutes for the instance to start shutting down. If the above values match your infra also, the 25 minutes you have reported is sort of normal. In this condition the only thing you can do to reduce the time to inititiate shutdown is to decrease the deregistration delay. But that might affect your inflight connections in the LB.

--Syd