On 20 October 2026, between 06:00 and 12:10 UTC, Google Cloud's London region (europe-west2) had a bad morning. One of its three zones, europe-west2-b, lost power completely. Google's incident page lists close to 40 affected products, from Compute Engine and GKE to Cloud SQL, Cloud Storage, Pub/Sub and Cloud Load Balancing, and gives the duration as 6 hours and 10 minutes.
The interesting part is not that a data centre lost power. It is that one dark zone degraded the whole region, including the two zones that never lost power. For anyone running a website or SaaS on any cloud, that is the lesson.
What happened
According to Google's preliminary incident report, published on 22 October, the facility hosting europe-west2-b lost its grid supply at 05:53 UTC. The automatic switch to UPS batteries worked as designed. What did not work was the next step: both the primary and backup PLC automation systems (the controllers that start the emergency generators and transfer the load) failed to do so. The batteries drained in about seven minutes. At 06:00 UTC the zone's servers, network switching and local storage went offline together.
Google's monitoring saw elevated API errors at 06:08 UTC. The first public status update came at 08:35 UTC, and the first workaround advice (migrate to another zone or region if feasible) at 09:24 UTC. Utility power returned within about an hour, but the zone came back in stages: generator controllers reset and tested, hosts restarted, storage integrity checked, traffic re-admitted gradually. Google reported the power event mitigated at 11:12 UTC, new VM creation working at 11:46 UTC and over 99 percent of VMs healthy by 12:10 UTC. The 14:03 UTC closure notice said most customers were recovered, with a very small number still affected. A final root cause report is still pending.
Why the healthy zones wobbled too
Zones a and c never lost power, yet customers there saw admission queues, slow database failovers and intermittent HTTP 500 and 503 errors. Google names two contributing factors. First, a demand surge: regional failover plus client retries pushed the dead zone's load onto the other two zones at once, and compute there saturated. Second, synchronous control-plane dependencies: API calls in the healthy zones still validated something in europe-west2-b and waited until they timed out.
In plain terms, the failover mechanisms did their job and that was the problem: every retry and every database promotion landed on infrastructure that was already full. Managed services with built-in failover, such as Cloud SQL, AlloyDB, Memorystore for Redis and Bigtable, did move to other zones. Anything that lived only in europe-west2-b, including plain VMs and zonal persistent disks, was simply unreachable until the power came back.
Who was affected
UK media reported many consumer-facing disruptions that morning. Coverage summarised by dev.to citing the BBC, and ABC News, mentioned Revolut (more than 1,500 Downdetector reports shortly after 08:00 London time), HSBC, Halifax and other Lloyds Banking Group brands, the HM Courts and Tribunals Service video platform used for remote hearings, DPD, the budgeting app Emma and the e-bike service Forest. Worth stating clearly: most of these companies described a third-party or general disruption and did not officially name Google Cloud. Treat the link as reported, not confirmed.
The pattern is instructive either way. A business does not need to be a Google Cloud customer to be hit. One dependency, a payment provider or an identity service, sitting in one London zone is enough.
What to do now
None of this needs a multi-cloud strategy. It needs you to know your single points of failure before an outage finds them for you.
- Find your zonal resources. On Google Cloud, VMs, zonal persistent disks, zonal GKE clusters and single-zone Cloud SQL instances live in exactly one zone. Each can disappear for six hours. Prefer regional disks, regional GKE control planes and Cloud SQL with high availability.
- Test losing a zone, on purpose. Drain one zone in staging and watch. If your service needs every zone up, you found the problem cheaply.
- Budget capacity for failover. Two zones at 80 percent cannot absorb a third. Leave headroom, or autoscaling that can actually obtain capacity.
- Make retries polite. Exponential backoff with jitter, capped retries and circuit breakers keep your clients out of the retry storm.
- Watch the status page, not just your own dashboards. The first public update came two and a half hours after the failure. Subscribe to provider status feeds, but alert on your own error rates too; you will usually know first.
- Check what your customers see. Does a cached or static page still load? Does DNS keep resolving? Is your own status page hosted somewhere else than the product it reports on?
How Scantinel helps
Scantinel scans your websites and public endpoints from the outside, as an attacker sees them: exposed software versions, known-vulnerable components, misconfigured headers and files that should not be public. It does not see inside your cloud account or which zone your database lives in, so the zonal versus regional homework above is yours.
Where it does help is the inventory step. Its scans map the hostnames, services and third-party dependencies exposed on your domains, a practical starting point for the question this outage raised: which of these would go dark if one provider, region or zone did? Knowing that list on a quiet day is cheaper than discovering it at 06:08 UTC.
Comments
No comments yet.