MLOps
Machine Learning Operations
What Is MLOps?
- A framework for developing, deploying, and running ML models
- DevOps for AI
- Drift detection and retraining
What MLOps changes in an organization is not model accuracy but 'who changed what, and when, becomes traceable.' Before MLOps, AI projects run only on the owner's laptop. If that person goes on leave, no one can rebuild the model that performed well last month. MLOps keeps data snapshots, hyperparameters, training code, and result metrics together as one package, turning experiments into assets. The focus shifts from competing on performance to competing on reliability.
In practice, maturity is described in stages: Level 0, where people train and deploy by hand; Level 1, where the training pipeline runs automatically; and Level 2, where code changes flow through automated testing all the way to deployment. Most companies in Korea sit between Level 0 and Level 1. The key difference from DevOps is what you have to manage. With ordinary software, version-controlling the code is enough, but with machine learning, code, data, and model all change at the same time. Drift also comes in two kinds: data drift, where the distribution of input data shifts, and concept drift, where the inputs stay the same but the meaning of the right answer changes. Concept drift is usually caught much later.
The most common misconception is that installing a platform gives you MLOps. Before choosing tools, decide 'what counts as performance degradation' and 'who approves retraining'. Many teams monitor only server CPU and response time, but a situation where the infrastructure is healthy and only the predictions are wrong never shows up on that dashboard. If the ground-truth label is only confirmed two months later, wrong predictions keep driving orders and reviews in the meantime. Fully automating retraining is even riskier. Once the model's flawed outputs flow back into the training data, the error compounds, and you can no longer trace when things started to break.
Hyerang ParkExpert in content storytelling and AI-driven brand content
Jaryong KooMarketing PhD and leading authority on AI, data analytics, and personal brandingA concept that applies software DevOps to machine learning, established as AI became widely commercialized.
Consider a hypothetical case. Company A, an online food distributor, has 900 employees and KRW 240 billion (≈ US$179,560,100) in annual revenue. Two years ago it deployed a demand forecasting model and recorded an initial forecast error (MAPE) of 8%. After a year and a half, fresh food spoilage began to rise again. People on the business side were saying "We get the same results whether we use the model or not."
- When Team Lead Kim in data analytics pulled every prediction log from the past 12 months, MAPE had climbed from 8% to 19%. Tracing the cause, the team found the model was still trained on 2022 consumer patterns.
- When the team set out to retrain, they found the training code was 47 files scattered across the laptops of three people. Team Lead Kim asked, "Can anyone reproduce last year's model exactly as it was?" Nobody raised a hand.
- The team added experiment-tracking tools that automatically record data snapshots, parameters, and evaluation metrics, and built a pipeline that cut a manual three-day deployment to two hours. Park, the infrastructure manager, deliberately kept one manual approval step.
- For monitoring, the team replaced server metrics with weekly MAPE and distribution shifts in key input features. When a threshold is crossed, an alert goes to a shared channel for the purchasing and analytics teams, and a rule requires them to reply with the cause within 24 hours.
- Retraining ran automatically, but only models that passed a champion-challenger comparison were promoted to production. In practice, this step caught two challenger models whose performance had dropped because promotional data had gotten mixed in.
- Six months later, MAPE had recovered to the 9% range, and the monthly spoilage rate for fresh food fell from 4.1% to 2.6%. The time it took to spot a performance problem also dropped, from an average of three weeks to under a day.