# Azure Container Registry makes geo-replication failover health-aware > ACR’s global endpoint can steer around replicas with deeper push and pull failures instead of relying only on shallow health. Canonical URL: https://www.devobs.io/news/news-azure-container-registry-health-aware-failover/ By: Ines Costa Published: 2026-09-06T11:58:54.637Z Updated: 2026-09-06T11:58:54.637Z Event date: 2026-03-16 Section: Infrastructure Microsoft announced health-aware failover for Azure Container Registry geo-replication. The [ACR team says](https://techcommunity.microsoft.com/blog/appsonazureblog/health-aware-failover-for-azure-container-registry-geo-replication/4501730) a Health Monitor tests deeper registry operations and feeds its assessment into routing for the global endpoint. Traffic can move when a replica has storage, cache, compute, or authentication problems even if its front door responds. ## Routing tests useful work Geo-replicated registries accept writes in multiple regions and replicate asynchronously. Latency and shallow endpoint health could previously select a nearby replica that accepted a connection but failed a push or pull. End-to-end probes align the routing decision more closely with the work a build or cluster needs. Automatic routing does not erase replication lag. A newly pushed manifest or layer may not have reached every replica when traffic changes course. Credentials, DNS, and private network paths can also fail independently of service health. ## Exercise failure assumptions Inventory clients using the global login server and those pinning regional endpoints. Test DNS behavior, retry budgets, and digest availability during a controlled disruption. For critical releases, verify immutable digests from each target region before promotion. Keep telemetry around authentication, manifest lookup, and layer download. Runbooks should show how to inspect replica status and stop writes that would complicate recovery. Health-aware routing reduces one partial-outage mode, but only client probes can demonstrate that a specific identity, artifact, and path work together. Treat the platform feature as one layer in release resilience, not as proof of artifact availability. ## Source references -