Cloud-Native Scalable Localization: A Serverless CI/CD Framework for Multilingual Software Delivery
Global-scale software delivery increasingly relies on distributed localization pipelines to manage multilingual content across diverse deployment environments. While Continuous Integration and Delivery (CI/CD) practices have matured, existing approaches struggle to deliver scalable and reproducible orchestration under unpredictable, high-volume workloads. VM-based infrastructures often suffer from limited elasticity and high operational overhead, especially when integrating ML components such as Neural Machine Translation (NMT), which increases complexity in resource allocation and coordination of concurrent tasks. This paper introduces a serverless CI/CD framework that reimagines localization orchestration through cloud-native, event-driven design. The framework's core innovation lies in its fine-grained function scheduling, enabling parallel execution of translation workflows with dynamic scaling and adaptive resource utilization. Stateless serverless functions eliminate idletime overhead and support deterministic reproducibility through modular encapsulation and predictable invocation. A hybrid translation model integrates NMT with Translation Memory (TM), balancing adaptive fluency with domain-specific precision while maximizing concurrency across both orchestration and translation layers. Preliminary evaluation on Microsoft Azure demonstrates the framework's operational advantages. It achieves sub-second cold-start latency and scales to $\mathbf{1, 0 0 0}$ parallel instances within 15 seconds under specific consumption plans. The framework processes a $\mathbf{1 2 0}$-word batch $\mathbf{6. 4} \times$ faster (0.25 s vs. 1.6 s) and increases per-instance throughput by 72% (43 vs. 25 strings/min), supporting a peak capacity of 43,000 strings/min. These results validate the framework's ability to meet demanding localization workloads efficiently and at scale. By shifting focus from model-level optimization to orchestration-layer efficiency, this work contributes a scalable and extensible paradigm for ML-enabled localization. It demonstrates practical applicability by delivering a resilient, cloud-native solution that enhances scalability and efficiency in real-world multilingual software localization workflows.