Weights & Biases

Weights & Biases (W&B) became the experiment tracking tool of choice for machine learning researchers and production ML teams largely by being genuinely useful from the first 'pip install wandb' command. Founded in San Francisco in 2017, the company built a platform that logs training runs, visualizes metrics, versions datasets and model artifacts, and helps teams reproduce experiments — problems that every ML practitioner encounters but that most organizations were solving with ad-hoc spreadsheets and custom scripts. By the mid-2020s, W&B was used by teams at major AI labs, universities, and enterprise ML organizations worldwide.

The Python SDK integrates directly into training loops with a few lines of code, automatically capturing loss curves, hyperparameters, system metrics (GPU utilization, memory), and custom-logged values. These are streamed to W&B's cloud backend where they're stored, visualized in the web dashboard, and made available for comparison across runs. The Artifacts system versions datasets, models, and evaluation results as immutable objects with lineage tracking — so teams can trace exactly which data and code produced a given model checkpoint. Sweeps provides automated hyperparameter search using Bayesian optimization, grid search, or random search, coordinating distributed training runs across a fleet of machines.

When W&B experiences an outage, training runs don't necessarily crash — the SDK is designed with offline queuing so metrics buffer locally when the backend is unreachable. But this means that dashboards go stale: researchers can't monitor training progress in real time, compare runs across team members, or catch performance regressions as they happen. Artifact uploads fail, breaking model versioning workflows. Sweep agents lose coordination, causing distributed hyperparameter searches to stall or produce duplicated work. Teams using W&B for model registry and deployment triggers find downstream CI/CD pipelines blocked.

Outage.gg tracks W&B service health so ML teams can quickly determine whether dashboard issues reflect a real platform problem. Check the live status page before investigating your training infrastructure.