Lightning strikes Google Compute Engine data center in Belgium; services disrupted and 0.000001% of data lost
Yevgeniy Sverdlik / Data Center Knowledge :
Context & Ripple Effects
The 2015 lightning strike at Google's Belgium facility reads, in hindsight, as the opening data point in a decade-long pattern of physical-infrastructure failures at Google Cloud. Later entries in the coverage include the 2019 incident where Google lost a third of its cloud traffic, the UK heatwave cooling failures that took down both Google and Oracle servers in 2022, and the June 2025 outage that cascaded into Cloudflare, Spotify, Discord, and Snapchat.
What makes the Belgian strike notable is its shape: even with Google's replication machinery intact, a small fraction of persisted data was unrecoverable. That admission set the template for how Google has since explained every subsequent failure — root-cause posts, quantified loss, and reassurance that the blast radius was contained.
First-order effects
- Compute Engine customers running workloads in the Belgium region faced immediate service disruption, and some learned their replicated data had a nonzero irrecoverable floor — 0.000001% — rather than a guarantee of zero loss.
Second-order effects
- Customers hosting downstream services on Google infrastructure — the dependency pattern later visible when Snapchat and YouTube were caught in the 2019 traffic-loss postmortem and when Cloudflare, Spotify, Discord, and Snap were hit by the June 2025 outage — absorbed the blast radius, pushing them toward designing for regional failure rather than trusting single-region durability.
Third-order effects
- With environmental causes recurring across a decade — lightning here, record heatwave cooling failures in 2022, 'rare issues' behind the week-long UniSuper private-cloud outage in 2024 — hyperscale cloud reliability is shifting toward being judged on documented worst-case behavior and customer-side multi-region architecture, not on the assumption of zero-loss durability.
The trend: Google Cloud's decade of physically rooted failures is turning durability guarantees into an architectural conversation between hyperscaler and customer rather than a vendor promise.