Multi-Region Replication for Feature Pipelines with Latency-Aware Consistency and Conflict-Free Resolution

Abstract

Modern feature pipelines increasingly operate across multiple cloud regions to reduce tail latency for online inference, satisfy data residency constraints, and improve availability under regional faults. Replicating feature state across regions is straightforward when features are immutable or when a single writer exists, but production feature stores routinely ingest concurrent updates from streaming joins, late-arriving events, backfills, and on-device signals that can originate in different regions. These realities create a tension between low-latency reads, bounded staleness, and deterministic conflict handling when updates race or arrive out of order. This paper studies multi-region replication for feature pipelines under latency-aware consistency objectives and conflict-free resolution requirements. The central idea is to treat feature values as mergeable state with explicit semantics and to couple replication policies to end-to-end inference service-level objectives rather than to a single global consistency level. We define a system model that connects feature freshness to model error through sensitivity estimates, enabling the system to allocate consistency and bandwidth budgets where they matter most. A latency-aware controller selects replication and read strategies per feature and per workload slice, while conflict-free resolution uses merge operators that preserve feature meaning across counters, sets, time-series aggregates, and embedding-like vectors. The design integrates approximate sketches and compression for cost control and exposes reproducible evaluation methodology based on workload traces and fault injection. The result is an architecture that can deliver predictable tail latency while keeping inconsistency-induced feature drift measurable, bounded, and operationally auditable.