Why Big Outages Arrive in Clusters, Even When They Are Unrelated
Several large services failing inside a few days looks like coordination. Shared plumbing and a shared calendar explain most of it, and that explanation is less comfortable than a conspiracy, not more.
A week in which a mail service, a bank, a games network and a government system all fail feels like it must mean something. The honest answer is that it usually means something, just not the thing that first comes to mind.
Start with the arithmetic. There are a few hundred services large enough that an outage becomes news. Each has a bad day a few times a year. Weeks with several newsworthy failures are not a departure from the pattern; they are what the pattern produces. We notice them because the second outage makes the first one newsworthy again, and because a cluster is a better story than a coincidence.
Then there is the plumbing, which is the part worth being uncomfortable about. The apps on your phone come from thousands of companies and run on a short list of foundations: a few cloud providers, a few content delivery networks, a few managed DNS operators, a few single sign-on platforms, a handful of certificate authorities, and the global routing table underneath all of it. A fault in any one of those surfaces as a dozen unconnected brands breaking in the same hour. The companies are genuinely unrelated. The failure is genuinely shared. Both statements survive, which is how a spokesperson can say the outages are not connected and be telling the truth about their layer while being useless about yours.
Timing stacks on top of that. The industry patches on roughly the same days, because the vendors publish on roughly the same days. It freezes changes before the same holidays, which means the riskiest work gets pushed out in the window just before the freeze rather than delayed past it. Quarters end simultaneously. Certificates and credentials issued together expire together, years later, on an afternoon nobody has thought about since. None of that is coordination. It is a shared calendar acting on a shared dependency, which produces something that looks a great deal like coordination.
The last ingredient is sameness. The same orchestration tools, the same configuration languages, the same handful of libraries, the same default settings. A monoculture is efficient and it fails in correlated ways: a bad release of one widely used component can take out organisations that have no relationship with each other and no idea they had anything in common.
So what should you actually do with a cluster of outages? Mostly, suspend judgement for a few days. The only document worth updating your beliefs on is the postmortem, and postmortems arrive late, written in flat language, naming a config push and a missing guardrail. They are boring on purpose and they are the real story. Claims made in the first forty-eight hours - by officials, executives or anyone else - are hypotheses competing for attention, and the people making them generally know less at that moment than they will next week.
What a cluster is genuinely good evidence for is concentration. If the same week can take out your mail, your bank and your authenticator, that is worth knowing about your own setup, whatever the root cause turns out to have been.
- Most consumer services rest on a handful of clouds, DNS providers, content networks and identity platforms, so one fault surfaces in dozens of apps at once.
- Risky changes cluster by calendar - the industry patches on the same days and freezes before the same holidays, which pushes deploys into the same windows.
- "Unrelated" and "same underlying cause" can both be true: unrelated companies, one shared dependency.