If what you actually want to do is relatively simple, eg, read a JSON message from an AWS service, like a Cloudwatch Alert, an autoscaling action, or an API call logged in Cloudtrail, you can easily wire those events up to Lambda functions for which you only need to code the bit that parses the JSON and does something in response. So then if the event happens once a day, once a month, or 50 times a second, you don't have to actually worry about calculating scale, except for cost purposes. But if it's a rare event, then Lambda comes in way cheaper than a $20/month server you have to maintain somehow.
Of course, you can build apps directly on Lambda, but there's a lot of tools you'll miss when you get around to going full scale production. But for simple glue to wire up actions to events, it's actually a pretty slick service.
Need to keep a security group in sync with the IPs some 3rd party service publishes? cloudwatch cron + lambda function.
Need to do some weird scattering in response to s3 object created events? lambda function.
want to roll some of your own automation to manage and respond to health events with services you host? some event stream + stepfunctions + lambda functions.
I use it for creating image derivatives on a project which doesn’t have dedicated servers for that task. Every S3 upload triggers a conversion task which saves the result to another bucket. The long-term average number of servers is close to zero since the content is released in batches but when a new batch arrives it’s reliably done in 5 minutes no matter how large it is, and I don’t have to manage autoscaling policies, server builds, etc. – just write a couple dozen lines of Python.
For a sufficiently large amount of work the cost tradeoff might go the other way but in my experience there are a LOT of tasks which are below the threshold where the savings in server time is greater than the savings in sysadmin time.
If you’ve got 100s of services or handle huge load spikes, why spend extra rev cycles on concerns like “do I have enough machines” or “are these things patched?” or even “how do I make sure my service stays up?”
Where I think Lambda is interesting is in areas of high variability.
The sort of canonical example is image resizing. While you certainly can do that on the same machine as your web server it makes a lot more sense to put the images to a S3 bucket (keeps them off the server, out of versioning, etc.)
You can tie a Lambda to a S3 create file event. So when the file hits the bucket it automatically gets resized into all the appropriate sizes needed elsewhere on the site. This setup sidesteps a ton of potential issues, makes your app much more robust and is pretty easy to get going.
Thanks to Lambda's generous free tier, I've paid literally nothing for these services. I could handle 100x the current load and still pay nothing.
At the free tier your options include, on a monthly basis:
* a single EC2 t2.nano box with 0.5gb memory and 1vcpu
* 1,000,000 requests (or 40,000 gb-seconds worth of requests) to your lambdas
Assuming you're taking in less than 1m requests monthly the lambdas will handle traffic spikes much more gracefully than the small VM will; as well, since it's not an either-or proposition this also means you still have a free VM you can do something with.
One obvious concrete benefit of serverless is not managing infrastructure.