AbstractUnderstanding the ecological processes that maintain community function in systems experiencing species loss, and how these processes change over time, is key to understanding the relationship between community structure and function and predicting how communities may respond to perturbations in the Anthropocene. Using a 30-year experiment on desert rodents, we show that the impact of species loss on community-level energy use has changed dramatically over time, due to changes in both species composition and in the degree of functional redundancy among the same set of species. Although strong compensation, initially driven by the dispersal of functionally redundant species to the local community, occurred in this system from 1996-2010, since 2010, compensation has broken down due to decreasing functional overlap within the same set of species. Simultaneously, long-term changes in sitewide community composition due to niche complementarity have decoupled the dynamics of compensation from the overall impact of species loss on community-level energy use. These results highlight the importance of explicitly long-term, metacommunity, and eco-evolutionary perspectives on compensatory dynamics, zero-sum constraints, and the link between species-level fluctuations and community function in a changing world.Original submissionThis submission analyzes long-term data on rodent community abundance and energy use from the Portal Project. Sections of this timeseries have been analyzed in numerous other publications, but this is the first to analyze data from 2007-2020 on compensation on experimental and control plots.No prior publicationThis submission is posted as a preprint on bioRxiv at [bioRxiv].Animal welfareRodent censuses were conducted with IACUC approval, most recently under protocol 201808839_01 at the University of Florida.Open researchAll data and code to reproduce these analyses are archived on Zenodo at https://doi.org/10.5281/zenodo.5544362 and https://doi.org/10.5281/zenodo.5539881.Analytic methodsAll analyses were conducted in R version 4.0.3.