Distributed System Is Slower Than a Laptop

Article note: I love how people periodically figure this out. In the "Big Data" boom era it was usually "Laptop with basic unix utility pipeline beats hadoop cluster" but the principle is the same: Distributed systems are fucking expensive, deeply fiddly methods of last resort when you can't do what you need any other way, and are willing to pay for massive amounts of hardware and engineering. Getting local parallelism right is hard enough. I've spent plenty of time helping wrangle problems on to clusters, and increasingly my impulse is "probably don't" for any problem that isn't inherently embarrassingly parallel (island GA, mote carlo trials, etc.)