8 guides tagged “production”.
A real, honest production architecture: an ONNX image classifier served by FastAPI on Railway, loaded from object storage at startup, with one shared inference session on CPU — and what we'd improve.
Your model scores 94% in the notebook and falls apart in production. The cause is usually not the model — it's a preprocessing mismatch between training and inference. Here's how to find and prevent it.
A practical guide to Dockerizing a FastAPI inference service: a lean multi-stage build, why you shouldn't bake the model into the image, sane defaults for uvicorn, and the mistakes that bloat images.
Models change more often than code, and a bad model can be worse than a bug. A practical guide to versioning model artifacts and rolling back fast when a new model underperforms.
Processing requests one at a time wastes hardware; batching them trades a little latency for a lot of throughput. How dynamic batching works, when it helps, and when a single shared session is enough.
A model that passed every test can still rot in production as the world changes. What to monitor for an ML service — latency, data drift, prediction drift — and how to start when you have zero instrumentation today.
An ML endpoint that accepts file uploads is an attack surface. Practical hardening for inference APIs — input validation, size and rate limits, and the defenses that matter before you take traffic.
When your service loads a model from object storage at boot, cold starts get slow. Why ML services start slowly, and the practical levers — image size, lazy loading, warm instances, and model size — to fix it.