Data versioning and reproducibility

If you cannot reproduce a result, you cannot trust or defend it, and that depends on knowing exactly which data you used. This guide covers data versioning and reproducibility for sourced data.

Available across the EU. DataSupplier sources and delivers this data in all 27 European Union countries — including Germany, France, Spain, Italy, the Netherlands and Poland — and across the EEA, in the format and cadence you need.

Why versioning matters

Data changes: sources update, get corrected, get re-sourced. Without versioning, a model or report cannot be reproduced, audited or debugged, because no one knows which data produced it.

Point-in-time and snapshots

Versioning means being able to retrieve the data as it was at a given time. Snapshots and immutable versions let analyses be re-run on the exact inputs, essential for backtesting and audit.

Lineage ties it together

Versioning works with lineage: knowing not just the version but how it was transformed. Together they make results explainable and reproducible.

Reproducibility in AI

For machine learning, reproducibility requires versioning data alongside code and models. A model is only reproducible if its training data version is known and retrievable.

Versioning external data

External feeds change outside your control, so capturing versions on receipt, with provenance, is the only way to pin down what you used. Revisions especially need handling.

In a managed model

A managed partner can deliver versioned, point-in-time data with lineage, so analyses and models stay reproducible.

Snapshots and point-in-time access

Versioning means being able to retrieve a dataset exactly as it was at a given moment. Immutable snapshots, with clear version identifiers, let an analysis or model run be reproduced on the precise inputs that produced it, which is essential for debugging, audit and backtesting. For external feeds that change outside your control, capturing a version on receipt, with its provenance, is the only reliable way to pin down what you used.

Reproducibility for models

A machine-learning result is only reproducible if the training data version is known and retrievable alongside the code and model. Versioning data together with lineage, what was transformed and how, turns “we think it was trained on roughly this” into a defensible, repeatable record, which increasingly matters under AI-governance expectations.

Sources & further reading

• DAMA-DMBOK: data lineage and lifecycle.
• Industry references on ML reproducibility and data versioning.
• W3C PROV: provenance.
• Internal practice: DataSupplier versioned delivery.