Worked on stabilizing metrics configuration in the pytorch/torchrec repository, focusing on resolving a subtle bug in MetricsConfig initialization. Addressed the issue of mutable default arguments in Python dataclasses by introducing factory functions, ensuring each MetricsConfig instance received a unique configuration object. This approach prevented cross-instance metric contamination during concurrent or parallel workloads, directly supporting reliability and observability goals without impacting performance. The work involved Python programming, dataclass usage, and applying software design patterns to improve code quality and maintainability. The targeted refactor resulted in safer initialization and more robust metric collection for production environments using TorchRec.