Ruggero Tonelli - @ruggero.bsky.social

Everyone talks about multi-cloud and reliability. Here’s what it looked like in practice when I joined Adevinta five years ago. The company had just agreed to acquire the eBay Classifieds Group, its secondhand marketplaces division (Marktplaats, mobile.de, and others). On the table: private cloud, GCP, and AWS. Multi-cloud made it to the final spreadsheet, but only so everyone could see how crazy expensive it would be. In the past, AWS and other hyperscalers had single points of failure. As @GergelyOrosz mentions in the latest Pragmatic Engineer newsletter, there are still some components that are “global” and run on a single region (us-east-1). That’s not the norm, and the vast majority of services can be run in multiple areas. It’s up to the business to understand (a) what its critical paths are and (b) what its risk appetite is versus the cost of running multi-region or even multi-cloud/hybrid setups for specific scenarios. For some businesses, 24/7 uptime is more critical than for others. Take Glovo or Uber, for example: there’s a high per-minute cost during peak hours. Why? The cost of switching is low. Users also have Lyft, FREE NOW, Just Eat, and Uber Eats on their phones, so they can easily open another app to get from A to B or order a meal. If one of Adevinta’s portals (secondhand, real estate, or motors) is down at 2 a.m. on a Friday, that’s mostly OK. Yes, there are peak times, but there’s also a counterbalance thanks to (a) network effects — a lot of the inventory on @leboncoin isn’t listed anywhere else; everyone in France uses it — and (b) low time sensitivity. If you’re looking for a flat or a car, it can wait a bit. Multi-cloud (and sometimes even multi-region) sounds smart until you price it, model the risks, and realize: it’s usually not worth it. Reliability is an architecture problem, not a cloud problem.