Not entirely clear what the volume of the monitoring system messages will be.
But the architecture requirements described sounds to me as follows:
- The agents on the EC2 instances are (constantly?) polling some centralized service, which is a poll based architecture
- The messages being sent are to a specific predetermined EC2 instance, which is a push based architecture.
To support both options without significant filtering of the messages I suggest you try using an intermediate PubSub system such Kafka, which can be managed on AWS by MSK.
Then to differentiate between the instances, create a Kafka topic named by the EC2 instance ID.
This should give you a unique topic that the instance will easily know to access messages for itself on a topic denoted by it's own instance ID.
You can also send/push Producer messages to a specific EC2 instance by sending messages to the topic in the cluster named by it's EC2 instance ID.
Since there are many EC2 instances coming and going you will end up with many topics. To handle the volume of topics, you can trigger and notify CloudWatch on each EC2 termination event and check CloudWatch to see which EC2 instances were terminated and consequently their topic needs deleting.
Alternatively, you can trigger a Lambda directly on the EC2 termination event event and log it by creating a file denoted by the instance ID to an S3 Bucket, which you can watch using an additional Lambda that will delete old EC2 instance topics from the Kafka cluster when their instance ID's appear there.