I only have experience with GCP’s Cloud Functions, but in that environment, “serverless” is only cheaper than virtual servers if your load can’t saturate the lowest end VM GCP has to offer.
Once you have enough load to justify going “unserverless” the prices drop to approximately 1/4 that of Cloud Functions for the same burst performance, and that’s before playing billing games like long-term commitments or using preemptible instances.
What’s truly scalable about “serverless compute” versus VMs is the line item on your bill. Sure, “they manage the auto-scaling” but for the per-unit price you rapidly hit a point where you might as well set up a Kubernetes load balancer and eat the setup costs. The pricing model only works out in your favor if you _don’t_ have load.
It would not surprise me to learn that the same is true of AWS prices.
I dunno... maintaining virtual servers is still a lot of work. You have to
* monitor lots of things like filesystem usage and CPU usage
* build new machine images (as well as possibly new container images!) to keep up with security updates
* tune autoscaling at multiple levels based on whichever server-based systems you're using
* maintain a secure way for production support folks to log into servers to see what is going wrong or perform emergency fixes
* build additional failover automation as well as what you'd already need for serverless
Serverless is a bit more expensive overall and less tunable, but there's just less to go wrong: much of the underlying server SRE work is handled by provider automation and engineers for whom that is their full-time job. So, I'm still pretty happy with it.
Once you have enough load to justify going “unserverless” the prices drop to approximately 1/4 that of Cloud Functions for the same burst performance, and that’s before playing billing games like long-term commitments or using preemptible instances.
What’s truly scalable about “serverless compute” versus VMs is the line item on your bill. Sure, “they manage the auto-scaling” but for the per-unit price you rapidly hit a point where you might as well set up a Kubernetes load balancer and eat the setup costs. The pricing model only works out in your favor if you _don’t_ have load.
It would not surprise me to learn that the same is true of AWS prices.