In my previous blog post, I documented how to take a Java Spring Boot application and convert it into a serverless function, which can be run in AWS Lambda.
Anyone who's done this before knows that cold starts are a big downside - Java and Spring Boot are not known for their speedy startup times, and a typical full fat Spring Boot-converted lambda can take anything from 10 to 90 seconds depending on how much memory and CPU you allocate it. This may force you to over-provision them to compensate for the cold starts, but that's a pretty expensive sledgehammer. There is always provisioned concurrency, but that doesn't work out much cheaper either (and negates the responsive scalability of lambda as you have to anticipate how many you'll need in advance).