Reviews

What did 2025's public incidents agree on?

A grid, a cloud, and everything between them. The year-end read of the incidents that published enough detail to learn from.

By TIS Partners · · 2 min

Year-end lists are cheap. Incident reports are not; the public ones cost their authors real candor, and reading them together is the best free education the year offers. Two dominated 2025's syllabus, and they disagree about everything except the lesson.

Two-color illustration of several distinct structures all leaning on one hidden support

In April, the Iberian grid collapsed in seconds and was rebuilt in a day. In June, a Google Cloud failure took a swath of Cloudflare's services down with it, and their customers with them. One is electromagnetic physics, the other is distributed software, and structurally they are the same event: a dependency that was load-bearing turned out to be shared far more widely than its dependents had modeled, and the failure crossed a boundary everyone believed was an independence.

What would have to be true for your estate to be different? Two conditions, both checkable in a January afternoon.

First, the level-two dependencies are known. Not your vendors: your vendors' vendors, which is where both events actually propagated. We walked through the method in July; the short form is that status-page correlation on a bad day reveals more topology than any architecture document you will be shown.

Second, the restart order exists. The grid's recovery was the year's most underrated engineering performance, and it worked because the black-start sequence predated the blackout. The software translation, which services start from nothing and in what order, remains unwritten at most companies, and 2025 supplied no reason to believe it will stay untested.

The disagreement between the two events matters too. The grid had rehearsed; the peninsula was relit in hours. The software estates downstream of June 12 mostly had not; their recoveries took whatever the afternoon gave them. Same lesson from both directions: the dependency graph is a fact whether or not you draw it, and the recovery is a plan only if it has been run.

Next year's list will feature different logos and the same two findings. The reports are public, the tuition was theirs, and January is a fine month for homework.