Show HN: List of serverless failure stories

In the first, there was no testing, aside from the trivial. Oddly enough, a lack of formal testing is described as being a symptom of not having "a local version to play with." In the second, the test environment was used in production ("So that we didn’t have to wait for the timer to fire “naturally” when we were testing, we were using RunOnStartup = true"). And in the third, there is no test environment.

So with AWS Lambda, humans have to decide which microservices should be put into lambdas, and how much they need to be split up (based on the AWS Lambda Limits, which increase over time).

Whereas in a language like Elixer/Erlang or Go, we write code at a high level and the compiler chops it up into threads and spreads it across as many processors as are available dynamically.

I feel, rather strongly actually, that the manual approach is a code smell. Which means that serverless leans towards being more of a trendy solution than a formal one built from first principles. Which means that given time (usually 1-3 years with web tech) we'll see something arrive that looks like a compiler and lamda-izes code somewhat automagically.

That said, there is a window of profit here where anyone who jumps on serverless and knows how to use it can generate cashflow.

Knowing all of this, I would look at previous evolutions of other web technologies for pitfalls to avoid. Having never used AWS Lambda, this is what I would do:

* Treat lambdas as single input/output executables like in the shell. Analogously to avoiding how AngularJS got mired in two-way data binding. This lets us use vast prior experience in data processing graphs rather than having to analyze complex webs of nondeterministic logic and side effects.

* Start from the beginning with development, staging and production environments. Just because it's lambda doesn't mean that we throw out everything we've learned about reliability just to run everything "serverless".

* Don't succumb to premature optimization. All logic can be reduced to functional programming, the highest expression of which is the 0th order functional programming of spreadsheets (which I saw making the rounds of HN the last few days). Write all code to not depend on global shared state. Even if it feels a little inefficient at first, try to do everything declaratively, as functional transformations on data rather than imperative steps storing temporary state (in class member variables, as an unfortunately common example). Then analyze the code near the end of the project and decide which functions might work better as lambdas. Maybe they are utility methods that don't justify having an entire container overhead around them. Maybe they are just glue code that swizzles one interface into another like GraphQL. That's something you'll have to decide.

I've seen problems with highly-partitioned code where versions of included services get out of sync, the friction of dealing with things at the service level outstrips any gains found by separating the code, and it becomes nearly impossible to trace the codepath across services. It becomes obvious that whoever wrote the code wasn't familiar with things like the Actor model. It's fine to break the rules once one is familiar with them, but serverless has a fair amount of brogramming thrown in and the move-fast-break-things mentality is just not compatible with reliable software engineering IMHO. I know, I know, Facebook and Amazon made billions with that approach, but guys like me have to clean up the mess and it's anathema to finding career zen.

Edit: P.S. I forgot to say why monolithic code that is parallelized/optimized internally is usually better (for both rapid-prototyping and long-term). It's because the closer that code is to fully abstracted and statically-analyzable, the more reliable, reusable and resalable it is. This is my rudder after many years of programming, but I look forward to hearing any other opinion someday that might replace it.