Glosarium
Contents

Data Analysis › The Analyst's Toolkit

Versioning Analysis Code

Putting queries and models in source control so a number can be re-run and diffed.

Data Analyst · Mid-level

Also known as: versioning analysis code, SQL in git, source control for queries, version control for analysis

Versioning analysis code means keeping the queries, notebooks and model definitions that produced a number in a git repository, next to the documentation, so they can be re-run, reviewed and compared. It is the difference between a number and an auditable number (version control).

Why it is worth the overhead: a stakeholder asks, six weeks later, how a figure was calculated. If the query is in a repo, you check out the commit from that week and run it. If it is a local file, you rebuild it from memory and hope you get the same answer. The same repository lets a colleague review the logic before it becomes a dashboard, and lets you see exactly which change moved a number when it moves.

The classic mistake is the local file. A query in ~/analysis/final_v3.sql runs on one machine, depends on one person’s database access, and dies with that laptop. The second classic mistake is the query pasted into a ticket or a chat message: no history, no review, and no way to tell which version produced which number.

Practical version:

  • One repository per project or team, with folders that mirror the subject rather than the author.
  • Commit the query together with the finding it produced, so every number has a commit behind it.
  • Use branches and pull requests for anything that changes a published number, so the change is visible before it ships (metric definitions).
  • Keep credentials, connection strings and environment config out of the query, so it runs on someone else’s machine and in CI (reproducible analysis).
  • Record which database and which snapshot the query ran against. The same query against a different database is a different answer.
  • Put the query where it will be found when someone asks where a number came from (data lineage).

Source control will not make an analysis correct. It makes it possible to ask, later, what was actually run.