MLOps without the buzzwords
Models drift, data changes, and yesterday's accuracy is no guarantee. What it really takes to keep AI working in production.
Shipping a model is the start, not the finish. The real work is keeping it accurate, fast and cheap as the world underneath it changes.
The minimum that counts as MLOps
Not a platform — a set of habits that make change safe.
- 1
Version the data as well as the code
A model you can't reproduce is a model you can't debug.
- 2
Monitor inputs, not just outputs
Ground truth arrives late; input distributions arrive immediately.
- 3
Make retraining boring
A scheduled, tested job — not a heroic afternoon.
- 4
Keep a rollback
The previous model, deployable in minutes, tested regularly.
Monitor the inputs, not just the outputs
Most model failures show up in the data first. We monitor input distributions for drift so we catch problems before accuracy visibly drops.
Make retraining boring
Retraining should be a routine, automated pipeline — not a heroic quarterly project. Reproducible data, versioned models and one-click rollback turn a scary operation into a Tuesday.
Make retraining boring
Boring
- A scheduled, tested job
- Versioned data and code for every run
- Automatic evaluation before promotion
- A rollback tested regularly, not theoretically
Exciting, in the bad way
- A notebook someone runs by hand
- No record of which data produced which model
- Promotion decided by looking at one metric
- A rollback nobody has ever executed