# Artifact Distance Dominates Transfer Time at Large Sizes ## Claim Once an artifact is large enough, where it is stored relative to where it is consumed matters more than any transfer optimization, so relocating the source beats tuning the path. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds above some size threshold where network path cost dominates fixed overheads — 30GB container images clearly qualify, small images clearly do not, and nobody has published where the crossover sits. - Breaks where the artifact cannot be relocated for licensing, provenance, or storage-cost reasons, and where relocation creates a failure domain worse than the latency it removes. - The measured result here bundles four simultaneous changes (in-cluster registry, host-network proxy, node co-location, load-balancer), so the specific 60× figure is an upper bound on what relocation alone would produce, not an isolated measurement. ## Evidence Subaru's ML images (30GB+) took roughly three hours to pull when the registry sat outside the cluster; moving Harbor in-cluster — alongside host-network Envoy, node co-location, and MetalLB — cut pull time to about three minutes. The reported cause is that image traffic previously crossed a slow network path entirely avoidable by storing the artifact where it's consumed. ## Links - [[3. Resources/Permanent Notes/Pre-Position the Artifact Rather Than Optimizing the Fetch]] — because extends; the Hyundai case shows the same mechanism applied across regions under a legal data-residency constraint, this note shows it applied within a single cluster under a purely technical one - [[3. Resources/Permanent Notes/Risk Concentrates at the Boundaries Between Correct Decisions]] — because contrasts; relocating an artifact to remove a network boundary can itself create a new single-point-of-failure boundary (the co-located node), which is the unnamed cost in this specific case - [[3. Resources/MOCs/Open Source Governance MOC]] — because applies-in; registry topology is core CNCF-ecosystem platform-engineering practice ## Tests / falsifiers - Plot pull-time improvement against image size for in-cluster versus external registries across multiple deployments. If improvement is size-independent, the mechanism is something other than distance. - Ablate the four bundled changes individually (in-cluster placement alone, versus with host-network Envoy, versus with node co-location); if in-cluster placement alone produces little gain, the claim misattributes the cause to distance when the real driver is elsewhere in the stack. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Full Document Contents/Articles/How Subaru Accelerated AI Model Development for Next-Generation EyeSight...- Ryoji Kobayashi, Subaru]] — KubeCon Japan 2026 case study on EyeSight ML platform bottlenecks.