MLOps

MLOps (machine learning operations) is the set of practices for taking machine learning models from experiment to production and keeping them working there. It applies DevOps ideas such as version control, CI/CD and monitoring to three moving parts instead of one: the model, the data it learned from, and the code around it.

How it works

A typical MLOps workflow adds a second loop to the usual DevOps cycle:

1. Collect and validate training data, and version it.

2. Train and evaluate the model, and store it with its metrics in a model registry.

3. Deploy the approved model behind an API or batch job.

4. Monitor latency, errors and data drift, and retrain when quality drops.

Serving 2 million API requests per day means about 23 requests per second on average, with higher peaks to plan for.

Why it matters in production

A model can fail without any error. If live data drifts away from the training data, predictions get worse while the service still returns HTTP 200. Teams need data and model checks alongside normal service monitoring, and a repeatable path to retrain and redeploy.

See also

• LLMOps

• Platform engineering

• Kubernetes

• AI observability

On the One2N blog

• Understanding MLOps from a software engineers' perspective

• From DevOps to MLOps: scaling ML models to 2 million API requests per day

References

1. Origin of the term

2. Hidden technical debt in machine learning systems (Sculley et al., NeurIPS 2015)