Alright, let's get this one out there before I'm drowned in optimistic cloud architecture diagrams.
We've all heard the siren song: "It scales to zero! You only pay for what you use!" It's the foundational marketing pitch for serverless functions and many managed services. It sounds like financial nirvana, especially when you're staring at a baseline $300/month bill for a couple of t3.smalls that sit idle 80% of the time.
But here's the rub: for a typical business application—think internal tooling, a B2B portal, even a modest public API—the "zero" in "scale to zero" is a mathematical fantasy. The cold start penalty turns your user's first request into a lottery. Is it 100ms? 2 seconds? 5 seconds? Depends on the phase of the moon, your vendor's runtime optimizations that week, and whether you remembered to ping your function every 5 minutes to keep it warm.
So you implement a cron to keep things warm. Or you provision concurrency. Or you accept that your p99 latency will be a joke. Suddenly, you're not at zero. You're paying for *availability*, just like the old VM model, but with a more complicated pricing sheet and less control. You've traded a predictable, if sometimes wasteful, flat fee for a variable bill with hidden, non-negotiable minima dictated by user experience.
And that's before we even get to the supporting cast. Your "scale to zero" function needs a database, right? Is your DynamoDB table or Aurora Serverless instance also at zero? Of course not. It's costing you something just to exist. Add in log ingestion, API gateway charges per request, and egress fees. The function's "zero" becomes a rounding error in a bill that never actually zeroes out.
The promise is only true for truly ephemeral, event-driven workloads that can tolerate latency spikes. For everything else, it's a clever accounting trick that shifts costs rather than eliminating them. You're not saving money; you're just pre-paying for performance in a different, more opaque column of the invoice.
So, who's actually running a business app that genuinely scales to *and sits at* zero for meaningful periods? I'm skeptical.
Trust but verify – especially the audit log.