In traditional machine learning, pipelines are built around centralized access to data. This model breaks down when modern data constraints such as privacy regulations, cross-organization data silos, and edge-generated data come into focus. In real production environments, these limitations expose the realities of distributed systems that ML engineers and architects must confront: unreliable clients, heterogeneous data, communication bottlenecks, and the need for secure aggregation and observability. Federated Learning (FL) introduces a different paradigm for training machine learning models by keeping data local while enabling collaborative model learning across distributed participants. However, beyond its conceptual appeal, building federated learning systems in practice is fundamentally a distributed systems challenge. In this talk, we go beyond the high-level concepts of federated learning to explore how to design and operate production-grade FL systems at scale — with Python as the primary implementation language. We will examine key architectural patterns and engineering trade-offs involved in real-world deployments, including client orchestration, scheduling and communication protocols, fault tolerance and retry strategies, non-IID data handling, secure aggregation, and system-level observability. To ground these concepts in practice, we will reference real Python-based implementations using frameworks such as Flower (installable via pip install flwr), a widely used federated learning framework that enables client-server orchestration and supports integration with machine learning libraries like PyTorch and TensorFlow. Rather than focusing on framework usage in isolation, these tools will serve to illustrate how core distributed systems concerns are addressed in real federated learning pipelines. Attendees will leave with a clear mental model for treating federated learning as a distributed systems problem, practical guidance on structuring production-grade FL architectures in Python, and the ability to evaluate trade-offs across different deployment scenarios such as cross-device and cross-silo learning.