detections/_docs/00_problem/restrictions.md

Restrictions

Hardware

• GPU: NVIDIA GPU with compute capability ≥ 6.1 required for TensorRT acceleration. Without a compatible GPU, the system falls back to ONNX Runtime (CPU or CUDA provider).
• GPU memory: TensorRT model conversion uses 90% of available GPU memory as workspace. Minimum ~2 GB GPU memory assumed (default fallback value).
• Concurrency: ThreadPoolExecutor limited to 2 workers — maximum 2 concurrent inference operations.

Software

• Python 3 with Cython 3.1.3 compilation required (setup.py build step).
• ONNX model: azaion.onnx must be available via the Loader service.
• TensorRT engine files are GPU-architecture-specific (filename encodes compute capability and SM count) — not portable across different GPU models.
• OpenCV 4.10.0 for image/video decoding and preprocessing.
• classes.json must exist in the working directory at startup — no fallback if missing.
• Model input: fixed 1280×1280 default for dynamic dimensions (hardcoded in TensorRT engine).
• Model output: maximum 300 detections per frame, 6 values per detection (x1, y1, x2, y2, confidence, class_id).

Environment

• LOADER_URL environment variable (default: http://loader:8080) — Loader service must be reachable for model download/upload.
• ANNOTATIONS_URL environment variable (default: http://annotations:8080) — Annotations service must be reachable for result posting and token refresh.
• Logging directory: Logs/ directory must be writable for loguru file output.
• No local model storage: models are downloaded on demand from the Loader service; converted TensorRT engines are uploaded back for caching.

Operational

• No persistent storage: the service is stateless regarding detection results — all results are returned via HTTP/SSE or forwarded to the Annotations service.
• No TLS at application level: encryption in transit is expected to be handled by infrastructure (reverse proxy / service mesh).
• No CORS configuration: cross-origin requests are not explicitly handled.
• No rate limiting: the service has no built-in throttling.
• No graceful shutdown: in-progress detections are not drained on shutdown; background TensorRT conversion runs in a daemon thread.
• Single-instance state: _active_detections dict and _event_queues list are in-memory — not shared across instances or persistent across restarts.