Monday, November 9, 2026 Today's Paper · Subscribe · Sign In

The Daily Ledger

Independent since 1974

Technology · Platform Chaos

Facebook Accidentally Sends 2.4 Billion Friend Requests to the Wrong People

A single misconfigured variable in a recommendation service reconnected strangers, exes and at least one estranged godparent. Engineers say the queue will take eleven days to drain.

MENLO PARK, Calif. — A section of the recommendation cluster that misfired early Thursday. Engineers describe the fault as "one integer, one very long night." Illustration: The Daily Ledger

For roughly nine hours on Thursday, Facebook was a place where everybody knew nobody. A single misconfigured integer inside the company's "People You May Know" service caused the platform to dispatch an estimated 2.4 billion friend requests to accounts that had no business being introduced — strangers on other continents, dormant high-school acquaintances, and, in one documented case, a woman's estranged godparent and his current wife.

Meta confirmed the incident shortly after 6 p.m. Eastern, describing it as "a routing error of unusual scale" and stressing that no messages, passwords or payment details were exposed.

"A configuration push at 2:14 a.m. changed a neighbourhood-radius variable from a distance in metres to a distance in miles, then applied it to a cluster that hadn't been restarted since March," said Priya Raghunathan, an infrastructure director at the company. "By the time the alerting caught it, the queue had already gone out."

'It just kept going'

Internal dashboards reviewed by The Daily Ledger show the outbound request queue climbing past 100 million per hour before engineers throttled the service at 11:20 a.m. Requests continued to arrive on recipients' phones for several hours afterward as the backlog drained, producing a distinct and widely reported notification sound that users compared to a smoke alarm with a grudge.

The company says roughly 61 percent of the requests have since been withdrawn automatically. The remainder sit pending, waiting on the judgement of people who have not spoken to each other in years.

"I got a request from a man in Oshkosh who appears to be a competitive woodchopper. We have one mutual friend. I accepted. He has not messaged. I think about it constantly." — Marguerite Lyle, 34, of Asheville, N.C.

Raghunathan said the fault was traced to a stale branch of the recommendation model that had been deprecated in 2024 but never fully retired. Restarting the affected cluster, she said, would have "unwound three years of cached relationships," so engineers chose to let the queue finish on its own.

Eleven days of strangers

Meta expects the backlog to clear by Nov. 20. In the meantime the company has frozen all outbound friend-request traffic, which means new requests sent by actual humans will not arrive until the queue is empty.

Not everyone is treating it as a malfunction. Analytics firm Sensorfield recorded a 14 percent rise in daily direct messages on Friday, largely in conversations between accounts that had never previously interacted.

"The product has spent a decade guessing who you should know, and it has never once been this interesting," said Dr. Ellis Vann, a researcher who studies platform recommendation systems. "The wrong people turned out to be a better dataset than the right ones."

Meta shares were down 1.8 percent at the close.

Correction, Nov. 9, 12:04 p.m.: An earlier version of this article said the fault affected a database in the company's Oregon facility. It affected a recommendation cluster in Menlo Park.

Editor's note: This is a fictional article. Any resemblance to actual outages, living persons or competitive woodchoppers is a coincidence worth enjoying.

DP
Dana Petrossian

Dana covers platforms and infrastructure for The Daily Ledger. She has reported on three continent-scale outages and one very long night in a parking garage. Reach her at dpetrossian@example.com.

``` A few things you can swap if you want it more convincing: the masthead name, the byline, and the date. If you'd rather have it as a screenshot-looking image instead of code, say the word and I'll render a desktop and mobile version.