Drones Have Shut Down Three Yandex Data Centres in Four Days. Two of Its Three Cloud Zones Went With Them

Yandex told its customers on Sunday morning to fall back on disaster recovery arrangements held somewhere other than Yandex. That instruction, posted to the company's own cloud status board, is the clearest measure of what four days of drone strikes have done to Russia's largest technology company: the cloud platform could no longer promise that anything inside it would stay up.

The immediate cause was an attack early on 11 October on Yandex's data centre on the outskirts of Vladimir, in the Energetik district. Yandex said infrastructure was damaged and operations at the site were "fully stopped." No one was injured. Reuters described it as the third major Yandex data centre knocked out of action in a few days.

Three sites in four days

Reuters attributes the Sasovo and Kaluga strikes to Ukrainian drones. For Vladimir, Yandex said only that it was a drone attack; the first reports came from the ASTRA and Crimean Wind Telegram channels before the company confirmed the shutdown, and Ukrainian responsibility has not been officially established. Russia's defence ministry claimed air defences intercepted 479 drones overnight across 16 regions — a figure that, like most such claims, is unverified.

Yandex operates five large data centres; three have now been hit. The company removed the locations of its data centres in four Russian regions from its own maps product, and its shares fell about 4% on the Moscow Exchange earlier in the week.

The platform went to emergency mode

The sequence on the cloud side is worth reading closely, because it is where the engineering lesson sits. Power disruption was recorded in one Yandex Cloud zone at around 04:00 Moscow time. Shortly afterwards, two of the platform's three availability zones — ru-central1-a and ru-central1-b — became unavailable, and Yandex Cloud moved to what it called emergency operation. The company described the remaining resource configuration as unstable, warned that its support channels might be unreachable and that the status board was updating with a delay, and advised customers to use alternative recovery plans at external sites.

More than 80 Yandex Cloud services were disrupted, spanning compute, networking, storage, managed databases including PostgreSQL and ClickHouse, Managed Kubernetes, developer and business tooling, and the AI layer — the YandexGPT API, SpeechKit and Vision OCR among them. On the consumer side the Alice voice assistant, Yandex Music, Telemost and Yandex Smart Home went down. Reporting from Kyiv Post put the knock-on effects in banking apps, ride-hailing platforms and the retailers Magnit and Fix Price.

Multi-zone redundancy answers a question nobody is asking

Availability zones exist to absorb the failure of one building: a transformer, a cooling loop, a fibre cut, a fire. They are designed on the assumption that the failures are independent. A drone campaign that works down a list of sites breaks that assumption directly, because the correlation is not technical but geographic and political. Zones placed a few hundred kilometres apart inside one country are independent with respect to hardware and entirely dependent with respect to being in range.

The Vladimir strike also showed the second-order path. The regional governor, Aleksandr Avdeev, said the attack damaged electrical substations, cutting power to around 5,000 residents; heat, water and about 80% of electricity were restored by Sunday morning. A data centre does not need to be destroyed to stop computing. It needs only to lose the grid, and the grid is a softer and larger target than the building.

Training compute does not fail over

Serving capacity can, in principle, be rebuilt elsewhere; it is replicated, stateless in parts, and replaceable with enough money. Training clusters are not. They are single installations of tightly coupled hardware, built once, in one hall, with an interconnect that only works at short distances. Reuters reports that two of the three supercomputers Yandex uses to develop its AI model sat in Sasovo — the site hit first, on 8 October. There is no availability zone for that. Whatever those machines were training stopped when the building did.

This is the part of AI infrastructure that the industry's own language obscures. Capability gets discussed as though it lives in weights and research teams. It lives in a small number of named buildings, and the number is smaller than most people assume.

The ledger runs both ways: Russian strikes have taken Ukrainian facilities offline too, with Omega Telecom's Kyiv data centre stopped after an attack and the dline.ua data centre destroyed. The lesson is not about either side's capabilities. It is that compute is now infrastructure in the ordinary sense — a physical, immovable, mappable asset with a power feed, and therefore a target. Every cloud provider's resilience documentation is written against hardware failure. None of it is written against this.

Sources: Reuters, via Central Oregon Daily · The Moscow Times · RBC Ukraine · Kyiv Post · iXBT, citing status.yandex.cloud · Ukrainska Pravda