What Is MLOps?

People say M L Ops all the time, but few can define it clearly. Is it a tool, a job, or a process? In this lesson, you will get a simple definition, and understand why machine learning needs its own kind of operations.

• Not a tool, a practice
• Dev Ops, plus data and models
• Why ML is different

Where MLOps comes from

M L Ops is a combination of three worlds. Dev Ops brought automation, testing, and reliable delivery to software. Data engineering brought pipelines, and trustworthy data. Machine learning brought models that learn from data. M L Ops sits where they meet. It is the practice of building, deploying, and maintaining machine learning systems, reliably, repeatably, and continuously. Machine learning, plus Dev Ops, plus data engineering, equals M L Ops.

Why ML is harder than software

Why is this different from ordinary software? In normal software, behaviour comes from code. If the code does not change, the behaviour does not change. In machine learning, behaviour comes from code, and data, and the trained model. Change the data, and behaviour changes, even if no line of code did. Also, software is tested against exact expectations, but a model is judged by statistics, and it can be wrong, a little, all the time. And its inputs come from the real world, which never stops changing.

The small box in the corner

A famous research paper from Google made this point with one picture. In a real machine learning system, the model code is a small box in the middle. Around it is a huge amount of other work. Data collection, and verification. Feature extraction. Configuration. Serving infrastructure. Monitoring. Resource management. The code for the model itself is often a few percent of the whole system. M L Ops is the discipline for everything around the box.

1. The model codea small part of the whole
2. Data collection and checksfeeds the model
3. Feature extraction and configprepares the inputs
4. Monitoring and servingkeeps it alive

Summary