The Serverless Revolution Has Stalled

[1] https://lwn.net/Articles/674770/

Instead of maintaining software compatible with the right operating system, you're maintaining software compatible with the right flavor of serverless by Cloud Provider. Now we're back to square one on at least one front.

On the control aspect, the bias against giving up control is not an unwarranted one. Maintaining control of critical infrastructure is extremely important. And in fact outsourcing your critical infrastructure is an existential one, and not just in an academic sense. When you give up control you give up the ability to prevent your infrastructure being hostile, but incompetent. In these cases it reduces the quality of your product for your customers.

I won't even go into the anti-competitive tactics Amazon themselves get into that make them not a good choice for your infrastructure. Instead I'll draw upon a recent experience that illustrates why outsourcing infrastructure, even at a higher level, is a bad idea.

My girlfriend recently was taking her NLN exam remotely. They weren't allowed to use calculators of their own, they had to use a virtual calculator provided by the company administering the test. Like most of these companies they are doing remote proctoring of the exams. During her exam this virtual calculator flat out wasn't available. The proctor told her to simply click through the exam and that once she submitted it she'd be able to call into customer service to get the exam re-scheduled due to the technical difficulties.

Well, that wasn't the case. After doing some deep digging for her, here is what I found. The testmaker NLN contracted out the test administration to a third party, Questionmark. Questionmark in turn contracted out yet another 3rd party, Examity to handle the proctoring. Examity proctors don't have access to Questionmark's systems. Questionmark doesn't have access to NLN's systems, etc. So how did we get this resolved? I had to track down the CEO of Questionmark, the CEO of Examity, and the head of testing for NLN. I had to reach out through LinkedIn inmail to get this on their radar. And then it was handled(quickly and efficiently I might add!). However, frontline support for each of these companies could do nothing. They just had to offload blame onto the support staff of each other. Another aspect of this is that each handoff between 3rd parties creates a communication barrier. In this case the communication barrier seems to have kept Questionmark from configuring this specific test correctly. I wouldn't blame any of these companies for this specific failure mode because it's just the nature of what happens when you offload your work to 3rd parties.

When you say, "Oh it's great we don't have to worry about X because Y can do it." The implication is that you lose all of the power of a vertically integrated company by essentially spinning off tons of subsidiaries and creating a communication overhead both before and when problems DO arise.

What is the future of software development going to look like when it reaches consumers and you have to say "Sorry, we can't fix that issue because Cloud Provider has to get back to me, and then in the background Cloud Provider has to say sorry we can't get back to you because we have to wait for our spunoff hardware division to get back to us?"

Maybe this type of business is fine for fun apps, but it's not fine for a lot of businesses. Even SLAs and disclaiming responsibility in your own contracts won't save your reputation. All it does is protect you financially!

The problem with PAAS and serverless until recently is that they still required a lot of devops activity. Most of that is pure drudgery: setting up networking, dealing with vendor specific weird shit (e.g. amazon's IAM makes everything harder than necessary), micromanaging instance types and trading off overpriced vcpus vs small chunks of memory, etc. Lots of things that you can do wrong, lots of subtlety, lots of poorly documented gotchas, lots of potential bugs, etc. And all of that is needed so you can say "go run this over there" where this is your packaged up docker application and there is months of work by some poor devops person to turn a large amount of poorly integrated tools into a vaguely coherent deployment experience. It's never simple. It's never cheap

I had a decent experience with Cloud Run recently. I was not in a mood or position to reserve 3 months out of my schedule to terraform myself a new server environment (which I actually know how to do). I already had the software and I wanted it running ASAP. So, I felt slightly dirty when I did this but got the job done in a few mouse clicks in their UI. This was shockingly easy after basically spending months piecing together arcane crap in AWS in my previous project.

This got me a cloud run deployment, a build pipeline against my github repository for continuous deployment, and a running service. I tuned the build file afterwards to do more than just docker build & deploy but that was relatively easy (similar to how most CI systems work these days).

They charge for cpu/memory used by requests. The whole thing has been running for a few months now. Less than 4 hours of work to figure it all out from me knowing absolutely nothing about Cloud Run to me having a service up and running. I reconfigured the deployment via the UI a couple of times to fiddle with memory and CPU settings.

I suppose I could sit down and terraform this at some point but I don't feel the need to do that right now. I have more interesting things to do. And technically me spending a day to do that would cost more than running the whole thing has cost so far. That's the point: the cost of devops is out of wack with the running cost for a lot of this stuff.