Porting applications into Kubernetes is not always a walk in the park, operators help to abstract complex logic behind domain specific naming and configuration. But how should they behave if the application itself relies on its own distributed process orchestration ? How should controllers and clients be designed when jobs are scaling up to 10.000 pods or more ? Traditional HPC software represents a perfect storm of all the aspects that may push its portability to the limit: unique assumptions, high cardinality of processes distributed across contended nodes and cores, heavy memory, cpu and network footprint. This talk will present some of the fundamental operator design patterns behind the successful integration of established HPC software into large managed production clusters. The audience will walk away with a set of key principles and ideas useful for creating efficient custom controllers and optimizing them for the different challenges of massively parallel batch processing.