Environmental forecasting systems are increasingly migrated to cloud infrastructure to benefit from elastic scalability, global accessibility, and managed services. This shift, on the other hand, changes the way things work and the way they are built: object storage takes the place of traditional file systems, and distributed execution models take the place of tightly coupled environments. Because of this, there are new problems with data layout, request latency, locality effects, and geo-redundant deployment strategies. This thesis examines the impact of storage representation, execution models, and architectural design on the performance and resilience of cloud-native forecasting workloads. A streamlined performance model is created to define object-storage behavior and to extract design specifications for cloud-compatible array formats. Based on these standards, modern storage formats are looked at in terms of how well they amplify metadata, how well they let you access subsets, and how well they handle multiple requests at once. An experimental assessment utilizing GOES-19 satellite data on Amazon S3 compares multi-file NetCDF and chunked Zarr formats in both region-local and cross-region access scenarios. Additionally, various execution strategies are assessed for iterative workloads in both I/O-bound and CPU-bound scenarios. An end-to-end hurricane tracking pipeline exemplifies a case study to evaluate practical applicability and pinpoint real-world bottlenecks. The results demonstrate that storage layout significantly influences performance in object storage, that execution strategies must be determined based on workload characteristics, and that geo-redundancy changes the primary resilience threshold towards distributed metadata management. By integrating experimental evidence with expert validation, the thesis provides evidence-based design guidelines for scalable and resilient cloud-native forecasting systems.