Serverless: slower and more expensive

If your best choice was Rails or Django 10 years ago you probably don't have a viable startup today. Why? Because it's 10 years later. Technology moves on and market niches get filled. There are orders of magnitudes more people with the skill to setup a basic CRUD webapp, and about 15 years for the markets that these can serve to have been filled.

As a side note, I've learned that the existence of a large number of viable ways to accomplish a task is a pretty big anti-signal for the desirability of accomplishing that task in the first place. When I started my career in 2000, there was a huge debate over whether the "right" way to develop a desktop application was MFC or .NET or Java or Visual Basic or Qt or WxWindows. The real answer was "don't develop desktop apps, because the web's about to take over". When the big web 2.0 businesses were founded from 2005-2011, there were basically two viable options for building a webapp: Rails or Django. Now that everyone's arguing about microservices vs. Docker vs. Kubernetes vs. serverless vs. Beanstalk vs. Heroku, it's likely that the real answer is "develop a blockchain app instead".

Oh wow you're right. That's an impressive write up

> Is your API so large and/or unstable that optimizing for time needed to add an endpoint is your lowest hanging fruit?

Startup, I am adding endpoints on a very regular basis as I build out the app. I probably add a couple a month.

As an example, I am writing one right now. It'll take me ~30 minutes to write+deploy it to test.

My long running and CPU intensive stuff is running on a VM of course.

I'm sure if I had spent several days learning yet another tech stack that I could create a system that let me deploy endpoints to a VM that fast, but I have Better Things To Do(tm).

I could also shove everything behind a single endpoint that has a bunch of options, but that brings in a separate set of problems.

The best application for serverless that I've found is supplemental, not wholesale.

Have a server, but offload resource-intensive tasks to cloud functions. Trying to shove everything into that paradigm wastes unnecessary time.

I too use a mixture of server and serverless solutions on AWS and you hit the nail on the head. Serverless is a really great solution for certain things that don't need to run all the time. If you are hosting a web server that needs to run all the time then just set up a traditional server.

We have a few lambda functions we use that send notifications or perform very specific tasks that are sort of on their own. Running on serverless is perfect for these. These functions are only invoked occasionally (a few times a month). So it is great that the Lambda only boots up the few times it is needed and I don't pay when it sits idle (most of the time). It costs me 10 cents a month instead of say $10 a month with a small EC2 server that might otherwise do the same thing.

Another great use case is bursting or inconsistent traffic/invocation patterns. For example, we have a use case where users can text/SMS in and receive a promotion code. We have this running on serverless because it runs very iradically. A nation-wide commercial will run and ask users to text a code, which triggers hundreds of thousands of Lambda instances within a few seconds. But then it might not run again for a month. On a traditional server platform, we would need a load-balanced server solution with several servers set up for redundancy on a huge influx of traffic, only for those servers to sit idle for who knows how long. You can't shut it off because people might text it at other times too, but you need to have a system that is available for "worst-case scenario" at all times (aka when a national commercial runs). You also don't know exactly when the commercial will run, so you have to always be ready. This setup might costs thousands of dollars a month on a traditional server system, but on Lambda it could cost $100 because we only pay during that huge influx of traffic. What I also like is that you can just trust that the Lamdas will boot up under nearly any traffic load. So you don't have to worry about making sure your server is "ready" or "the right size". If we get 1 hit, then we pay a fraction of a penny. If we get ten million hits then we pay $20 (really - it costs $0.20 /Million). Try to get the same reliability with a $20 server that can do that... spoiler: you can't!.

So I don't look at it as a slower solution that is more expensive. If you are running data analytics or a web server that is always on then this is surely the case. In which place I would say to use a traditional server.

But serverless has a role and it is great for infrequent usage, or for burst traffic. We love that we know we can get nearly infinite traffic and just pay per request without worrying about the infrastructure being ready. It is infinitely scalable, so that we don't need to worry about it.

It is always a pro and a con. I love serverless because it opens up new options for developers and system admins. It is another tool in our toolbelt. It is not the perfect solution for everything. We still run many vanilla EC2 instances for internal CRMs and webservers that run constantly. Or for booting up long computational processes. These are cases where servers are still far better than serverless. If we all learned to understand it as another tool available to us and not a "one or the other" requirement then there is a wonderful world where both tools can coexist.

Note: I use the word "infinite" very loosely in this post. I mean "infinitely scalable" in that it is scalable up to any reasonable amount, beyond what I would ever have to worry about. If the day comes that my servers are getting more traffic than Amazon is able to handle in their data servers then I will be far richer than anyone else here and I will come up with a new solution at that point.