MLOps

MLOps is the operational practice of managing a machine learning model after it's built, keeping it running smoothly in a live service without breaking down.

A chef developing a new recipe and a restaurant serving hundreds of plates of that recipe every day are completely different jobs. Perfecting the recipe is a one-time task, but running the restaurant means endlessly checking ingredient quality, making sure the taste hasn't drifted, and handling rushes when customers pour in. Machine learning models work the same way. Training a model on data is a one-time job, but once that model is hooked up to a live service, someone has to keep watching whether the input data has changed, whether the predictions have started looking off, and how fast the system responds when traffic spikes. MLOps is the name for this whole cycle of watching, fixing, and redeploying.

The scope of this concept has widened over time. Back when the focus was putting number-prediction models into production, it was called MLOps. Once language models that generate text, like chatbots, appeared, the concept expanded to cover managing prompts and model versions too, leading to LLMOps. More recently, as agents that pick their own tools and carry out multi-step tasks have emerged, the concept is expanding again into the next stage beyond that.

In the article, this comes up where Databricks explains that the practice of putting prediction models into production used to be called MLOps, then became LLMOps once language models appeared, and has now evolved into AgentOps for agents. A common misunderstanding is thinking MLOps is a technique for training better models — it's actually the operational process of deploying an already-built model into a service and finding and fixing problems as they arise.