In 2015, a single-threaded laptop beat 128-core clusters at graph processing. A decade later, the industry still celebrates scaling charts with no single-threaded baseline. From microservices to stream processing to AI infrastructure, the same pathology persists: systems parallelise overhead, then measure the parallelism.

CodeGood

The COST metric (Configuration that Outperforms a Single Thread), introduced in a 2015 paper, showed that single-threaded laptop implementations often outperform distributed systems running on 128+ cores. A decade later, the industry still evaluates systems by scalability charts rather than against single-machine baselines. The same pattern persists across microservices, stream processing, and AI infrastructure: teams introduce coordination overhead, parallelize it across more hardware, then measure the parallelism as progress. Key forces sustaining this include framework constraints that favor distributed algorithms, career incentives rewarding complexity, cloud vendor economics favoring horizontal scale, and premature scaling disguised as prudence. The practical takeaway: before adopting distributed architecture, ask what the simplest single-machine implementation could achieve. PostgreSQL often beats Spark for datasets under hundreds of gigabytes; a Go binary often outperforms Kafka+Flink pipelines. Distributed systems should earn their place by demonstrating genuine gains over simpler alternatives.

Your Distributed System Is Slower Than a Laptop. You Just Don't Know It Yet., CodeGood