This paper presents a series of case studies and topics from recent experiments in the context of a deployed CTR prediction system at Google. The focus is on topics that have received less attention but are crucial in a working system, such as memory savings, performance analysis, confidence in predictions, calibration, and feature management. The authors highlight the close relationship between theoretical advances and practical engineering in this industrial setting and detail several directions that did not yield significant benefits despite promising results elsewhere. Key contributions include the use of FTRL-Proximal for online learning, per-coordinate learning rates, probabilistic feature inclusion, and methods for automated feature management. The paper also discusses the challenges of applying traditional machine learning methods in a complex dynamic system and provides insights into practical engineering solutions.This paper presents a series of case studies and topics from recent experiments in the context of a deployed CTR prediction system at Google. The focus is on topics that have received less attention but are crucial in a working system, such as memory savings, performance analysis, confidence in predictions, calibration, and feature management. The authors highlight the close relationship between theoretical advances and practical engineering in this industrial setting and detail several directions that did not yield significant benefits despite promising results elsewhere. Key contributions include the use of FTRL-Proximal for online learning, per-coordinate learning rates, probabilistic feature inclusion, and methods for automated feature management. The paper also discusses the challenges of applying traditional machine learning methods in a complex dynamic system and provides insights into practical engineering solutions.