# Pre-Position the Artifact Rather Than Optimizing the Fetch ## Claim When a resource is slow to retrieve across a boundary, replicate it to where it will be consumed rather than making the retrieval faster, so the slow path stops existing instead of improving. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds where the artifact is stable enough that replication stays current and storage cost is acceptable — container images, package mirrors, CDN assets. - Breaks for rapidly-changing or user-specific data where replication cannot keep pace, and where regulatory constraints forbid the copy itself rather than merely the transit. - Strongest where the boundary being crossed is legal or geographic (cross-border data residency) rather than merely technical, since technical bottlenecks can sometimes be widened but legal ones cannot. ## Evidence Hyundai AutoEver's Harbor registry holds 70,000+ artifacts across 5,000 repositories, with topology driven by a legal prohibition on cross-border data movement. Rather than accepting cross-region pulls as a fixed cost, they resolved it through replication, proxy cache, and policy controls that keep artifacts close to where clusters consume them — routing over a private backbone rather than optimizing a slow cross-region fetch. ## Links - [[3. Resources/Permanent Notes/Artifact Distance Dominates Transfer Time at Large Sizes]] — because extends; the Subaru case shows the same mechanism at a single-cluster scale (in-cluster registry vs. external) rather than Hyundai's cross-region scale - [[3. Resources/Permanent Notes/Tooling Cannot Be Faster Than the API It Polls]] — because contrasts; that note is about a case where the slow layer cannot be bypassed (a gated public API), this one is about a case where it can be (an artifact you're allowed to copy) - [[3. Resources/MOCs/Open Source Governance MOC]] — because applies-in; registry and artifact-distribution architecture is CNCF-project infrastructure governance in practice ## Tests / falsifiers - Measure whether pull-time improvement from relocation scales with artifact size and boundary distance; if improvement is largely size- and distance-independent, the mechanism claimed (distance dominates) is wrong and something else explains the gain. - Identify a case where the artifact legally cannot be copied at all (not merely slow to move) — there the claim doesn't just weaken, it becomes inapplicable, which is a sharper falsifier than a gradual breakdown. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Full Document Contents/Articles/Out of the Box, at Multi-Region Scale How Hyundai Scales Its Platform - Jaewoo Choi, Hyundai]] — KubeCon Japan 2026 lightning talk on Argo CD and Harbor at scale.