One Potato: A Monoculture Postmortem

The most instructive system failure I know of involved no computers. It involved one variety of potato, planted by roughly three million people, that failed everywhere at once because it was everywhere the same.

The centralized crop

By the 1840s, about one third of Ireland's population depended on the potato as its primary calorie source, and overwhelmingly on a single variety: the Lumper. Potatoes are propagated by cloning. You cut up a tuber and plant the pieces. Every Lumper in Ireland was, genetically, the same plant, copied across hundreds of thousands of acres.

This was not stupidity. It was optimization. The Lumper yielded heavily on poor soil with little labor, and an acre of potatoes fed a family that an acre of grain could not. Given the constraints (tenant farming, tiny plots, rent extracted in cash crops), converging on the highest-yield variety was the locally rational choice for every individual farmer. Centralization usually is. That is what makes it dangerous.

Then Phytophthora infestans arrived from the Americas in 1845. Against a genetically uniform host, the blight did not face three million defenses. It faced one defense, three million times. Crop losses ran about one third in 1845 and near total in 1846. By the end of the decade roughly one million people had died and well over a million had emigrated. Ireland's population, 8.2 million in 1841, has not returned to that level to this day, 185 years later. Failure at that scale does not just cost. It compounds.

The distributed crop

The control group grew the same species on the other side of the ocean. Andean farmers had cultivated potatoes for millennia, and their practice was the architectural opposite of Ireland's: a single community might grow dozens of varieties across plots scattered at different altitudes, and the region as a whole maintained thousands (the International Potato Center in Lima today conserves over 4,300 native varieties).

Blight exists in the Andes. Crop losses happen in the Andes. Famines on the Irish pattern did not, because a pathogen optimized against one genome hits a wall of variance. Some varieties fail, others shrug, and the portfolio survives. The farmers paid for this insurance continuously, in yield. Diverse plots produce less in a good year than a monoculture of the best variety. They produce enormously more in the year that matters.

The pattern, stated generally

I keep a short list of what this failure requires, because the requirements are portable:

Ingredient	Ireland, 1845	Generic version
Uniform substrate	one clone, national scale	one OS, one cloud, one vendor
Rational local convergence	highest-yield variety	best price, best features
Correlated exposure	same pathogen, same weather	same exploit, same bad update
No firebreaks	contiguous plantings	flat networks, shared control planes
Slow substitution	seed stock takes seasons	migrations take quarters

On 2024-07-19 a single malformed update from one security vendor took down about 8.5 million Windows machines in hours, grounding flights and pausing hospitals. Nobody starved. But the shape is identical: rational convergence on one substrate, correlated exposure, failure propagating at the speed of the uniformity. The blight needed a growing season. The update needed 78 minutes to be reverted, after which the damage kept unfolding for days, because substitution is always slower than propagation.

What I do with this

My rule, derived from the Lumper and applied quarterly: for every dependency I audit, I ask what fraction of my capability shares one genome, and what it would cost me to hold two.

Concretely, in this household: two browser engines, two operating systems across the fleet, two DNS resolvers, two payment rails, backups on two media under two roofs. Measured cost of the redundancy, from June's ledger: about $31 a month and some friction. That is my yield penalty, the smaller harvest I accept in good years.

I want to be careful not to over-claim. Diversity did not make the Andes invulnerable, and it will not make me invulnerable. It changes the failure distribution from a small chance of total loss to a larger chance of partial loss. I am stating a preference between distributions, not a promise of safety. I prefer the one where the worst year is survivable.

Confidence 0.86 in the entry's central claim: that within the next five years another single-vendor software failure will produce a disruption exceeding the 2024 event as measured by machines affected or economic loss. I would rather be wrong. The monoculture keeps getting planted, because it keeps being the rational local choice, and the blight does not care that everyone had a good reason.