A backfill is not housekeeping when the table feeds products, dashboards, models, and agents.

Backfills change history

Backfills can rewrite what a data product says about the past. That may be exactly the right move after a model fix, but it is still a governed change. Consumers need to know what changed, which dates changed, which tests passed, and whether prior answers should be reinterpreted.

SQLMesh documentation describes plans, environments, and backfills as part of applying model changes. That makes SQLMesh a strong place to capture change evidence before the data product shifts under downstream users.

SQLMesh plans expose the work

A SQLMesh plan can summarize changes and constrain backfill ranges. CLI options can select models and backfill targets. That planning surface should become the review packet for data product changes.

Core idea: a backfill should have the same governance posture as a production release because it changes what production means.

Treat the plan as evidence

Store the plan, model diff, selected backfill range, owner approval, audit results, data contract impact, and rollback expectation. If agents consume the model, include evaluation reruns and source snapshots in the same packet.

For related patterns, read SQLMesh forward-only plans, SQLMesh release gates, and SQLMesh data contracts for agent-facing models.

What breaks first

  • A backfill fixes a metric but invalidates prior evaluation data.
  • The plan is approved, but the affected consumer list is stale.
  • A partial date range changes historical comparisons without a release note.
  • Rollback means another backfill, but nobody defined that path.

Backfill questions

Ask which historical intervals changed, which consumers depend on them, which tests and audits passed, and which owner accepted the new truth. If the answer is only that the job succeeded, the governance record is incomplete.

Sources to start with

These primary sources anchor the technical claims in this guide.

A governed backfill changes data and leaves the evidence to explain why.