Sacred

Sacred is a Python library that structures, records, and reproduces computational experiments. Its primary architectural mechanism is configuration injection, which inspects function signatures and local scopes to supply hyperparameters without configuration dictionaries. By pairing automatic git and environment tracking with modular storage observers, it turns standalone training scripts into auditable execution records.

Where Sacred fits

Sacred sits between dedicated configuration managers like Hydra and full-lifecycle tracking platforms like MLflow or Weights & Biases. Unlike client-server platforms that rely on background logging daemons and frequent HTTP requests, Sacred runs entirely in-process. Metrics and metadata are captured directly inside the Python runtime with negligible CPU overhead. For training routines where database writes could stall GPU execution, an asynchronous queue observer delegates persistence to a separate worker thread to preserve training throughput.

This embedded design eliminates server management, but lacks the visual tooling of modern platforms. Sacred provides no integrated web dashboard, leaving users dependent on community-built frontends like Omniboard or Sacredboard that receive infrequent updates. Furthermore, its parameter injection inspects local function scopes at runtime. While this removes boilerplate from small scripts, it bypasses static type analysis and degrades IDE autocompletion across large multi-module repositories.

What it does well

• Automated environment auditing: records git commit hashes, working tree status, active package versions, and host hardware specs on every execution to ensure reproducible runs.
• Parameter injection: maps configuration scopes decorated with @ex.config directly into function arguments by name, removing manual dictionary parsing.
• Decoupled storage writes: routes run telemetry through QueueObserver to a background worker process, ensuring metric logging never bottlenecks GPU loops.
• Flexible backend storage: persists artifacts, logs, and metadata across MongoDB, Amazon S3, Google Cloud Storage, SQL databases, TinyDB, and local directories.
• Deterministic seed control: automatically generates and injects explicit random seeds across Python, NumPy, and deep learning runtimes to prevent unseeded drift.