Abstract The growing complexity of data processing systems has long led systems designers to imagine systems (e.g. databases, schedulers) which can self-configure and adapt based on environmental cues. In this context, reinforcement learning (RL) methods have since their inception appealed to systems developers. They promise to acquire complex decision policies from raw feedback signals. Despite their conceptual popularity, RL methods are scarcely found in real-world data processing systems. Recently, RL has seen explosive growth in interest due to high profile successes when utilising large neural networks (deep reinforcement learning). Newly emerging machine learning frameworks and powerful hardware accelerators have given rise to a plethora of new potential applications.

In this dissertation, I first argue that in order to design and execute deep RL algorithms efficiently, novel software abstractions are required which can accommodate the distinct computational patterns of communication-intensive and fast-evolving algorithms. I propose an architecture which decouples logical algorithm construction from local and distributed execution semantics. I further present RLgraph, my proof-of-concept implementation of this architecture. In RLgraph, algorithm developers can explore novel designs by constructing a high-level data flow graph through combination of logical components. This dataflow graph is independent of specific backend frameworks or notions of execution, and is only later mapped to execution semantics via a staged build process. RLgraph enables high-performing algorithm implementations while maintaining flexibility for rapid prototyping.

Second, I investigate reasons for the scarcity of RL applications in systems themselves. I argue that progress in applied RL is hindered by a lack of tools for task model design which bridge the gap between systems and algorithms, and also by missing shared standards for evaluation of model capabilities. I introduce Wield, a first-of-its-kind tool for incremental model design in applied RL. Wield provides a small set of primitives which decouple systems interfaces and deployment-specific configuration from representation. Core to Wield is a novel instructive experiment protocol called progressive randomisation which helps practitioners to incrementally evaluate different dimensions of non-determinism. I demonstrate how Wield and progressive randomisation can be used to reproduce and assess prior work, and to guide implementation of novel RL applications.