The paper introduces the concept of *eddies*, a mechanism for continuously adaptive query processing in large, federated, and shared-nothing databases. Eddies continuously reorder operators in a query plan as the query runs, allowing for dynamic adaptation to changes in resource availability, data characteristics, and user preferences. The authors characterize *moments of symmetry* where pipelined joins can be easily reordered and *synchronization barriers* that require coordination between different sources. By combining eddies with appropriate join algorithms, such as Ripple Joins, the system can merge the optimization and execution phases, enabling flexible ordering of query operators. The initial implementation demonstrates promising results, showing that eddies perform nearly as well as static optimizers in static scenarios and provide significant improvements in dynamic environments. The paper also discusses the architectural assumptions, the properties of query processing algorithms that allow frequent reordering, and experimental results illustrating the effectiveness of eddies in both static and dynamic environments.The paper introduces the concept of *eddies*, a mechanism for continuously adaptive query processing in large, federated, and shared-nothing databases. Eddies continuously reorder operators in a query plan as the query runs, allowing for dynamic adaptation to changes in resource availability, data characteristics, and user preferences. The authors characterize *moments of symmetry* where pipelined joins can be easily reordered and *synchronization barriers* that require coordination between different sources. By combining eddies with appropriate join algorithms, such as Ripple Joins, the system can merge the optimization and execution phases, enabling flexible ordering of query operators. The initial implementation demonstrates promising results, showing that eddies perform nearly as well as static optimizers in static scenarios and provide significant improvements in dynamic environments. The paper also discusses the architectural assumptions, the properties of query processing algorithms that allow frequent reordering, and experimental results illustrating the effectiveness of eddies in both static and dynamic environments.