Experiment Tracking vs Model Registry
People think the model registry is where training runs are logged, and it is not. Experiment tracking logs every run, good or bad, with its settings, data and scores. The registry keeps only the models someone chose to keep, with a stage such as testing or live. Each version links back to its run. Track to find the best model, and register it to ship it.
Experiment Tracking
Notes on each training run. You can then compare them and redo the best one.
Experiment tracking means saving a record of every training run, so you can compare runs and repeat the best one. Each record holds the settings used, the version of the code and the data, the scores and the files the run made. A dashboard then shows runs side by side, so you can see which change helped. Without it, people forget which settings gave the best model. It is often the first MLOps tool a team adds.
Model Registry
A list of each model version, its score, and which one is live.
A model registry is a list that keeps every trained model version and its details. For each version it notes who made it, which data and code were used, its test scores and its stage. The stage may be testing, live or archived. Ship tools read the registry to know which version to serve. Moving a model to live, or back to an old version, is a change in the registry that can be signed off and logged. It links training to serving.
In one line
- Experiment Tracking
- Notes on each training run. You can then compare them and redo the best one.
- Model Registry
- A list of each model version, its score, and which one is live.