> We had to add on hacks like setting the request timeout to 5ms and keeping the cluster warm by submitting a no-op query every 500ms to keep it even remotely stable.

This sounds like you're blaming dynamo for you/your stack's inability to handle connections / connection pooling.