Adapting Software Engineering Observability Practices for AI Systems

This chapter explores the challenges and strategies of achieving observability in AI systems, drawing parallels with medical diagnostics. Observability is defined as inferring internal states from external telemetry, with a focus on software engineering and data science aspects. The chapter adopts the AWS observability maturity model to guide organizations through stages of observability development, starting with foundational monitoring. Telemetry, logs, traces, and metrics are identified as key components for monitoring AI systems. The chapter emphasizes the importance of establishing standards for observability to prevent user complaints from being the primary alert signal and outlines the role of logs in capturing AI model inputs and outputs for advanced observability techniques. Logs, traces, and metrics are critical components of AI observability, with logs capturing metadata such as request IDs, model configurations, usage statistics, and outcomes, but requiring contextual integration with metrics and traces for effective debugging. Traces, structured records of request lifecycles, are essential for diagnosing complex failures and optimizing developer velocity, despite challenges in implementation and integration. Metrics, quantitative signals tracking system performance, are categorized into availability, resource, throughput, and AI interaction metrics, aligning with the Four Golden Signals framework.