How the Economics of Multitenancy Work

https://news.ycombinator.com/rss Hits: 12
Summary

In the early days of Blacksmith, back when we were just a scrappy YC startup building a serverless cloud platform for CI workloads, we ran simulations to model our margins. We figured that with enough customers, the math would work out, and we crossed our fingers — but we didn’t have any real-world data to back up our predictions.About six months after we launched, I came across a blog post by Marc Brooker on the economics of multitenant systems. It captured what we were trying to do much more elegantly than the half-formed ideas in our heads. This post was heavily inspired by Marc’s, and reading it was a real moment of, "Oh! someone else has thought about this, and it makes sense."Now, we’re running thousands of jobs every minute and millions every month, and it’s been exciting to actually see this play out at scale and watch the math work in real life. Yet, people still often ask in disbelief how we actually make money from our setup — are we just burning sweet VC dollars with no return for them in sight? So, for the non-believers and anyone who is just a little bit curious, let’s take a peek behind the curtain and dive into how the economics of multitenancy work — using ourselves as a case study.CI Isn’t Like Production (And That Matters).Unlike production workloads, CI workloads tend to be very spiky. Below, we’ve plotted CPU utilization for one of our customers over a 24-hour window. It spikes when someone pushes code, then chills out in between.This customer runs 35 jobs on 16 vCPU machines for every git push, meaning they need over 500 vCPUs every time they run CI. See where the chart flatlines like a patient on the George Clooney classic “ER”? Since their team is split across the US and EU, there’s an 8-hour stretch in the middle with zero usage. And when a few engineers push at once — say, five people — they suddenly need 2,500+ vCPUs instantly. And all these CI jobs are short-lived. Most CI jobs finish fast (relatively speaking), somewhere between 5 and 40...

First seen: 2025-05-14 13:35

Last seen: 2025-05-15 00:36