Choosing an I/O approach for Earth system models: lessons learned from a modular I/O server for MESSy

Modern Earth system models increasingly hit I/O limits—not only in performance, but also in reproducibility, maintainability, and developer productivity. As data volumes and workflows evolve, tightly coupled, file-centric I/O approaches can become hard to scale and hard to extend.

We present the design and lessons learned from introducing an asynchronous, modular I/O server concept in the Modular Earth Submodel System (MESSy). I/O operations were decoupled from the Fortran-based scientific core and implemented as separate Python services, where the communication between the two components was implemented using the Yet Another Coupler (YAC) library. This architecture was chosen to improve flexibility and long-term maintainability, while enabling heterogeneous workflows and evolving storage backends.

Using MESSy as a case study, we discuss practical decision criteria for selecting an I/O concept in large models (e.g., scaling behavior, accessibility for developers, testing and CI strategies, and reproducibility).  We conclude with lessons learned from bridging Fortran and Python communities and from lowering entry barriers for user-developers in a large modeling system.