Since serverless is a kind of a PaaS, it offers many of the same benefits. However, unlike Heroku, Cloud Foundry, OpenShift, and other traditional PaaSes focused on supporting long running applications and services, serverless frameworks offer a new kind of a platform for running short lived processes and functions, also called microflows.
The distinction between the long running processes and microflows is subtle but important. When started, long running processes wait for an input, execute some code when an input is received, and then continue waiting. In contrast, microflows are started once an input is received, and are terminated by the platform after the code in the microflow finishes executing. One way to describe microflows is to say that they are reactive, in the sense that they react to the incoming data.
They are not always terminated, in an effort to avoid cold boot startup times. In fact, they run much like any other service framework, where your code is behind a function lookup, and called when an appropriate request is received.
AWS Lambda in particular will spin up a new container running their framework, which will then stay alive for a period of time before being terminated.
With socket activation (using xinitd or systemd) and a minimal framework, you could implement much the same on your own VMs.
>They are not always terminated, in an effort to avoid cold boot startup times.
That's true. I should have pointed out that from the developer's point of view it appears as if the execution is terminated, in the sense that no state is carried across code executions. However, the issue of whether the code is cold booted or not is an implementation detail of a specific serverless platform.
Any code written in a common web framework (such as Django, Rails, or Play) could easily be written with the same constraints - no state left over from request to request. That we typically choose to retain state between requests as an optimization speaks of one of the weaknesses in serverless architecture.
If you can't maintain state inside the application, you need to maintain it elsewhere (DB, Redis, memcache), and that elsewhere has a cost associated with accessing it.
That's a part of Twelve Factor apps as well. No state in the app! And while there is a cost to keeping state in an external data store like Redis, there are costs to keeping state in the app as well. First, there is the question of complexity and testability. Your app becomes harder to reason about, harder to test, and much more likely to be dependent on complex integration to get to a testable state. Second, there's the need for external support for internal state in any horizontally scaled app. This can be very expensive.
Consider session state in a web app. If you keep state in memory and scale horizontally, you need session affinity, so stateless inbound http connections hit the same server instance. That means a session-aware router, delicate and extremely expensive F5s and other "smart" routers. But if the app is stateless and session state is in external storage, then all you need for load balancing is a simple round robin. A cheap router, that's more reliable and requires less administration, for a fraction of the cost.
There are other ways stateful apps are expensive, but that's a big one. I'd rather pay $5k than $100k for a router.
Diego, the underlying scheduler for Cloud Foundry, has a concept of pre-defined one-off tasks. It'll eventually show up in the rest of the platform.
However, it'll be in the middle of the range for startup times, as some agents are sideloaded. Latency is around a second. The initial target is tasks like database migrations and batch jobs, so the ~1s delay isn't an issue.
One of my colleagues in NYC spiked on using one-off tasks for a FaaS and concluded it's not a slam dunk fit.
If I was to shimmy FaaS-style stuff into Cloud Foundry, I think it'd be easier to have standing apps with an array of functions already running, then push the magic into route services. I suspect this is how AWS already does it, given the vast difference in spinup times between Node and JVM requests -- Node makes hot-patching a list of functions easy, the JVM is less hot-patch friendly and really expects you to provide all the JARs at launch time.
Disclaimer: I work for Pivotal, we donate the majority of engineering to Cloud Foundry.
How would you modify systemd unit files for each user of your public cloud? How would you manage the sockets once you have more users than sockets on a server?
The distinction between the long running processes and microflows is subtle but important. When started, long running processes wait for an input, execute some code when an input is received, and then continue waiting. In contrast, microflows are started once an input is received, and are terminated by the platform after the code in the microflow finishes executing. One way to describe microflows is to say that they are reactive, in the sense that they react to the incoming data.