Your Distributed System Is Slower Than a Laptop

This piece argues that distributed stream-processing architectures are the default in mid-market SaaS companies not because they outperform simpler alternatives, but because the comparison is never run. It draws on the 2015 COST paper ("Scalability! But at what COST?") by McSherry, Isard, and Murray, which showed a single laptop thread beating multi-hundred-core graph-processing clusters on real workloads — including a connected-components task on Twitter's 1.5-billion-edge follower graph where a 1970s union-find algorithm on a laptop finished 16× faster than GraphLab on 128 cores.

The article constructs a representative 55-engineer company processing 2 billion events/day and prices the conventional distributed stack — Kafka, Flink, 128 vCPUs — at roughly $1.38M/year, with platform engineering salaries ($875K) dwarfing compute ($262K). The single-server alternative: two $10K machines and a fraction of one engineer, about $57K/year. The distributed system is slower, not just more expensive, because most cores burn on coordination overhead.

Four structural reasons keep the distributed-by-default reflex alive: frameworks benchmark only against other distributed systems, never a well-written single thread; careers reward operating complexity, not eliminating it; cloud vendors sell node-hours, incentivising sprawl; and premature scaling is mistaken for prudence. The article offers four honest justifications for distribution — data truly won't fit, availability requires replicas, latency demands geography, or the organisation genuinely needs partitioning — but treats each as a burden to prove with measurement, not a starting assumption.

The core remedy: build a single-machine baseline first. Run the real workload on one competent server, make that number the threshold a proposed architecture must beat, and treat the comparison as an ordinary financial control. The author cites Amazon Prime Video cutting costs 90% by consolidating cloud functions, Segment reversing its microservices migration, and Stack Overflow serving a billion monthly page views from nine web servers at single-digit CPU utilisation — all evidence that the laptop is still winning while the industry keeps not keeping score.