AI Trend Research

← AI glossary

AI glossary · Basic · Data engineering · MLOps & LLMOps

Data versioning

Data versioning means recording the state of a dataset every time it changes, so you can retrieve the data as it was at any point in time and compare what changed.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-10-09
In plain words It is like the 'version history' in a document editor: if you don't like yesterday's edits, you can roll back to last week's state or see the differences between two versions. But data is much larger than documents, so you need a smart way to record only the changed parts instead of copying everything each time.

How it works

Each time data is modified or added, a version marker (snapshot) is left for the state at that point. A common approach is to leave large files in storage as they are, while the versioning tool records only each file's unique fingerprint (hash) and location. This way, the same file is not stored repeatedly across versions. Tying code versions and data versions together lets you record exactly 'which code and which data this model was trained on'.

Why it matters

The results of a machine learning model depend heavily on its training data, not just its code. If data changes quietly, retraining with the same code produces a different model, and it is hard to find why performance dropped. With data versioning, you can reproduce experiments exactly and roll back problematic data to an earlier state. It also provides evidence when proving 'what data was used for this decision' in regulation or audits.

Things to know

Using a version control tool built for code as is quickly makes the repository heavy because of large data files. So people often use data-specific tools or the 'time travel' (querying past points in time) feature offered by data lakehouse table formats. Keeping every version forever increases storage costs, so retention rules are needed. Data containing personal information may remain in old versions, so care is needed when handling deletion requests.

Code versioning vs. data versioning

Code versioningMainly text files

  • Files are small enough to store whole
  • Easy to see line-by-line differences

Data versioningMainly large datasets

  • Records only hashes and locations to avoid duplicate storage
  • Time travel to query past points in time
  • Also accounts for retention periods and personal data deletion

공통 Tying the two together lets you reproduce which code and data produced a result

Both let you roll back to past states and see differences, but they differ in the size of what they handle and how it is stored.

Example

Say a team retrained its customer inquiry classification model and performance dropped. If they had data versioning in place, they could compare the data version used by the previous model with the new one, find that a large number of wrong labels had entered during recent labeling work, and roll back to the earlier data to retrain. Used together with experiment tracking tools or a model registry, the data version used for each model is recorded automatically.

Related terms

Experiment tracking Model registry Data lineage Data lakehouse MLOps Data quality