Reducing cross-region image pull costs

WHAT THIS TESTS This checks distributed-systems thinking about image distribution: locality, consistency, and cost trade-offs.

A GOOD ANSWER COVERS Two main approaches. A pull-through cache places a registry mirror in each region; the first pull fetches from the upstream registry and caches the layers locally, and subsequent pulls in that region are served locally, so most traffic stays in-region and egress drops. This is lazy and cheap but suffers cold-start misses and needs an eviction or retention policy. Geo-replication instead actively replicates a single logical registry to regional endpoints so content is pre-positioned; pulls are consistently fast but you pay storage in every region and tolerate replication lag where a freshly pushed tag may not be present everywhere immediately. Either way configure nodes to resolve the regional endpoint, handle authentication consistently, and monitor cache hit rates. Choose based on whether deploy spikes need pre-warmed content or whether lazy caching suffices.

COMMON WRONG ANSWERS Keeping all clusters pulling from one central registry. Ignoring replication lag and assuming instant global consistency. Forgetting cache eviction and storage cost. Overlooking auth propagation to mirrors.

LIKELY FOLLOW-UPS How do you handle a cache miss during a large rollout? What is the consistency model of replication? How do you size cache storage? How do nodes discover the regional mirror?

ONE CONCRETE EXAMPLE Clusters in three regions all pulled from a single registry, generating cross-region egress on every node scale-up. You deploy a pull-through cache per region; the first node to need an image warms the regional cache, and the rest pull locally, cutting egress and shaving pull latency. For predictable release spikes you add geo-replication so new tags are pre-positioned, accepting extra per-region storage and brief replication lag in exchange for fast, consistent rollouts.