MLOps
MLOps applies DevOps discipline — version control, automated testing, CI/CD, monitoring — to machine learning systems, covering the parts that are unique to ML: versioning datasets and models, tracking experiments, and watching for a model's real-world accuracy quietly decaying over time.
Standard DevOps practices — version control, automated testing, CI/CD — apply to ML systems too, but ML adds problems regular software doesn’t have. A model’s behavior depends on the data it was trained on, not just the code, so dataset versioning matters as much as code versioning. Training the same model twice can produce different results, so teams need experiment tracking — a record of exactly which data, code, and parameters produced which model, so results are reproducible and comparable.
The problem that has no equivalent in regular software is model drift: a model that was 95% accurate at launch can quietly become 80% accurate six months later, not because anything broke, but because the real world changed and no longer matches the data it was trained on. Regular software monitoring watches for errors and crashes; MLOps monitoring has to watch for a model that’s still running perfectly and returning results that are just quietly wrong.
For teams running LLMs specifically, MLOps overlaps with LLM observability — logging prompts, responses, token costs, and latency well enough to debug a bad output or a cost spike after the fact, not just confirm the service was “up.”
Related terms
Related reading