Drone strikes knock out three Yandex data centers, forcing emergency cloud migration

Between October 8 and October 11, physical strikes by unmanned aerial vehicles disrupted three primary data centers operated by Yandex across central Russia: Sasovo in the Ryazan region, Kaluga, and Vladimir. The cascade of hardware failures forced Yandex Cloud, which served approximately 64,000 corporate clients in the first half of 2026, into an official emergency operating regime without a recovery timeline.

The disruption began on October 8 when the Sasovo facility caught fire following a strike, entirely shutting down zone ru-central1-b and causing outages across Yandex search, maps, mail, and taxi services, as well as third-party services like T-Bank and Avito. Yandex has not confirmed whether any hardware at Sasovo remains salvageable. By October 9, multiple modules at the 63 MW Kaluga site were taken offline. The crisis culminated on the night of October 11, when the 50 MW Vladimir facility went dark entirely, destabilizing zone ru-central1-a.

Platform telemetry on October 11 indicated that while IAM and APIs stayed online, Compute capacity was only partially functional. Many high-availability databases dropped into read-only mode or failed completely, Object Storage accepted reads while rejecting writes, and the platform froze all requests for additional resource quotas. In response, Yandex updated its user agreement on October 10 to classify drone and missile attacks explicitly as force majeure events.

With its remaining architecture labeled unstable, Yandex Cloud advised customers to evacuate workloads to rival providers, arranging expedited onboarding agreements with Selectel, K2Cloud, and VK Cloud. The company has not disclosed the fate of its three flagship supercomputers—Chervonenkis, Lyapunov, and Galushkin—which were housed inside the Sasovo and Vladimir sites. Industry analysts cited by habr estimate standard server replenishment will take two to three months, while rebuilding destroyed supercomputers could take up to two and a half years.

Source: habr

Primary source: Yandex Cloud