Wandb

W&B is the default choice for ML experiment tracking in production ML teams. The Python SDK handles run initialization, metric logging (wandb.log), artifact versioning, sweep configuration, and model registry operations — all with minimal code changes. For agents in ML pipelines: log training metrics per run, compare runs across experiments, version datasets and model checkpoints, trigger hyperparameter sweeps, query historical results. Free tier for individual use. Self-hostable (W&B Server). Confidence is docs-derived.

Two access patterns: (1) Python SDK (wandb): pip install wandb; wandb.init(project='name'); wandb.log({'loss': 0.1}); wandb.finish(). (2) Public REST API: GraphQL at api.wandb.ai/graphql for querying runs, metrics, artifacts. Python SDK is the primary interface for ML training code; GraphQL for post-hoc analysis. Artifact API: wandb.Artifact for versioning datasets, model checkpoints. Framework integrations for PyTorch, TF, Keras, HuggingFace Trainer.

API key auth: set WANDB_API_KEY environment variable (or wandb.login(key=...)). Keys from wandb.ai/settings → API keys. Per-user keys; entity/project scoping in SDK calls. HTTPS enforced. Service accounts for CI/CD automation. Self-hosted: LDAP/SSO available for enterprise.

SDK raises exceptions on auth failure (CommError), network issues (RetryError, with automatic retry). wandb.run.finish() ensures run data is flushed. Offline mode: WANDB_MODE=offline queues logs; sync later. Rate limiting: effectively unlimited for standard SDK logging. W&B service uptime at status.wandb.ai. Run crash detection: mark run as failed if process exits without wandb.finish().

docs.wandb.ai is comprehensive — SDK reference, guides for all major ML frameworks, sweep configuration, artifact management, REST API reference. Getting started: pip install wandb, wandb login, add 3 lines to training code. Free tier generous. Large community via W&B Discord, GitHub. Extensive example notebooks.