Compute#Cloud#Backup

Yandex confirms 2.84% of cloud customer data is gone for good after drone strikes

After October 8–11 drone strikes on three data centers, Yandex restored 97.16% of affected customer data; 2.84% is unrecoverable. New VMs default to backups on.

A data center aisle of racks lit in blue

Yandex Cloud, the Russian cloud provider run by the country’s largest search company, confirmed in a November 10 blog post that its recovery effort after the early-October Ukrainian drone strikes on its data centers is effectively closed: 97.16% of affected customer data was restored, and the remaining 2.84% is permanently lost — the storage hardware was destroyed and the customers had not enabled backups on the platform. TechRadar picked the confirmation up on November 14 and German outlet heise reported it on November 12. To our knowledge this is the first time a cloud provider has publicly confirmed large-scale, irreversible customer data loss caused by a physical wartime attack. Yandex’s response is two changes: backups on new virtual machines are now enabled by default (you can still switch them off in the console), and a new availability zone, ru-central1-f, went live on November 11 in a separate physical location.

Key points

  • The split: 97.16% of affected customer data restored, 2.84% unrecoverable, per Yandex Cloud’s own blog (November 10). The base is customer data held in the struck data centers — not all Yandex Cloud data — and the company did not disclose how many customers were affected or how the loss was distributed.
  • The timeline: drones struck three data centers on October 8, 9 and 11; the recovery announcement came November 10, roughly a month later. TechRadar’s roundup is dated November 14.
  • The fixes: backups default-on for newly created VMs (opt-out in the management console), and a new availability zone ru-central1-f in a separate physical location, live since November 11 with resource quotas capped at launch.

Background

Yandex’s cloud carries a lot of Russian consumer and business load: search, maps, mail, ride-hailing and payments all felt outages after the strikes. TechRadar’s summary of the attack list says a October 8 strike hit the facility in Sasovo, Ryazan region, home to two of Yandex’s three supercomputers — Chervonenkis (29.42 PFLOPS peak) and Lyapunov (20.03 PFLOPS); a October 9 strike hit the Grabtsevo industrial park in Kaluga region, one of Yandex’s largest computing sites; and a October 11 strike hit a facility in Vladimir housing the third machine, Galushkin (20.64 PFLOPS). heise adds that the supercomputers themselves were reportedly undamaged and have been relocated, and are still used to train language models. About five data centers belonging to Yandex, Sberbank and VK were hit across Russia in October.

The failure mode is what makes this story different from the Microsoft 365 outage we covered on November 14, where a botched network capacity upgrade took Exchange Online and Teams down for roughly two days. That was a software-layer single point of failure: painful, but recoverable, because nothing was destroyed. Here the hardware and the data on it went together — which is why 2.84% of one cloud’s affected data is simply not coming back. Chinese coverage of the confirmation (NetEase reposts) has focused on the same default that made the difference: Yandex Cloud’s backup service, disk snapshots and managed-database backups were paid options, off unless the customer turned them on.

The facts

The numbers each source confirms, checked on November 15:

ItemValueSource
Recovery rate, affected customer data97.16%Yandex Cloud blog (Nov 10)
Permanently lost2.84% — hardware damage plus no customer backupsSame
Strike datesOct 8 (Sasovo), Oct 9 (Grabtsevo), Oct 11 (Vladimir)TechRadar (Nov 14); heise only confirms strikes happened about a month before the announcement
Recovery announcementNovember 10Yandex Cloud blog
New zone ru-central1-fLive Nov 11, separate physical location, quotas capped at launchYandex Cloud blog, corroborated by heise
Supercomputers at the hit sitesChervonenkis 29.42 PFLOPS, Lyapunov 20.03 PFLOPS, Galushkin 20.64 PFLOPSTechRadar

A few caveats on the record. Yandex’s own blog post never mentions drones or Ukraine — it says only that infrastructure was “damaged” in October; the attack attribution comes from heise, TechRadar and other outlets, so the officially confirmed part of the story stops at the percentages and the fixes. Customers who had used Yandex Cloud’s backup tools (Object Storage, Cloud Backup, disk snapshots, managed database backups) lost nothing — the blog states this explicitly. And the Russian government’s response is already layered on top: heise reports companies may now use foreign data center capacity only with approval from a government subcommittee and the FSB, while private domestic data center operators must reserve 10% of their capacity for state use.

What others say

Russian business daily Kommersant interviewed ITGlobal.com’s vice president of cloud services, who said interest in Kazakh data centers rose after the incident; TechRadar adds that industry sources claim some companies have begun moving operations abroad, though no specific company, data volume or permanent relocation is confirmed. Chinese commentary has converged on one line: a paid, off-by-default backup setting transfers disaster-recovery cost to the user’s own diligence — and this time 2.84% of the data paid for that default. Note on sourcing: we found no independent English-language wire coverage (Reuters, The Register and the like) of the confirmation itself; TechRadar’s strike details trace back to United24 Media, a Ukrainian state-affiliated outlet, so the three sites and the supercomputer figures cannot be fully verified from independent reporting.

Our take

The interesting part is not Yandex — it’s the sample. A cloud provider’s availability nines have never covered data survival: SLA exclusions almost always list “customer did not enable backups,” and almost nobody expects to hit that clause. It took a physical strike plus a paid toggle defaulting to off to produce 2.84% permanent loss; on a cloud where backups are on by default, the same attack plausibly reads as downtime instead of data loss. Three things worth taking away if you run anything on someone else’s machines: replication across a cloud’s own copies of your data is not a backup, because the copies can die with the hardware; an independent backup outside the affected region — or outside the cloud — is the only insurance against a physical event, which is the oldest 3-2-1 rule doing its job; and defaults beat diligence, which is exactly what Yandex admitted by flipping backups on for new VMs. Still unknown: how many customers and how many terabytes sit behind the 2.84%, where ru-central1-f physically is, and how the loss split between consumer and business customers.