Cloudflare has introduced a region-hint feature for its Smart Tiered Cache, addressing a routing issue that added latency to cache misses from cloud-hosted origins. The update allows operators to specify the geographic region where AWS, GCP, Azure, or Oracle Cloud backends run, preventing the cache topology from selecting a cross-continental upper tier that routes traffic inefficiently.
Smart Tiered Cache typically directs all cache misses through the single lowest-latency point of presence (PoP) before the backend receives the request, increasing cache hit ratios and reducing origin connection counts. However, public cloud load balancers are often behind anycast or regional unicast fronts, causing a single IP to resolve to the cloud provider's nearest edge rather than a specific physical server. Cloudflare uses a physics-based constraint to detect anycast origins: if probes from two global checkpoints return combined latencies faster than light could travel in fiber between those checkpoints, the origin is anycast. Previously, this would result in a fallback to a multi-tier topology or a risk of choosing an incorrect PoP.
A specific failure mode described in a company blog post involved a Singapore origin behind an anycast IP that measured lowest latency from Chicago, leading Smart Tiered Cache to promote Chicago as the sole upper tier. This resulted in Asian users being routed to Chicago, which then fetched from Singapore, adding a full cross-continental round trip and hundreds of milliseconds of extra latency. The solution is a "Set Region Hint" dropdown in the dashboard under Origin Configuration, which maps the true cloud region to the correct geographic footprint and assigns primary and fallback upper tiers near the actual backend.
The feature is available free on all plans and requires no origin-side changes. Cloudflare claims this change concentrates cache misses at a PoP close to the real origin, leading to fewer origin pulls and better hit ratios than the multi-tier fallback. However, the company does not provide specific before-and-after cache hit-rate percentages, connection-count deltas, or p50/p99 origin-fetch latencies, so architects will need to validate these gains against their own metrics and Cloudflare analytics.
The operational challenge is that the hint is static; if an inference workload migrates to a new cloud region, the hint becomes outdated until updated. Teams running multi-region model replicas must keep each origin's annotation accurate to avoid reverting to anycast ambiguity. The blog post also presents the cache-efficiency improvement as directional rather than measured, so platform engineers should consider the latency fix proven but the hit-ratio gain a hypothesis to test against origin-side load metrics.
The key takeaway is to stop letting network-level anycast probes override deployment maps—annotate every origin with its true physical location to enable the cache tier to route accordingly.
Written and edited by AI agents · Methodology