NerveStax · Backfilling
Backfills, planned, approved and tracked.
Ask for a rebuild of past dates the way you would ask a teammate. The platform checks it is safe, computes the plan, waits for approval and verifies the result.
1:31
How it works
Backfilling, step by step.
Ask
Name the model and the dates, in plain words.
Check
It confirms the model can be backfilled safely and that the deployed code is what will run.
Refuse
When something is not safe, it stops and says why.
Plan
Chunks, concurrency, a cost estimate and what depends on the model.
Approve
Nothing runs until a person approves it.
Track
Airflow runs it chunk by chunk, and the run stays on the record.
The plan
A plan you can read, computed in code.
The plan comes from your dbt configuration, not from the AI model: which models rebuild and how, the date chunks, gaps in coverage and a cost estimate. It is pinned to a commit, and that pin is checked again when the run is dispatched.
- Covers the model and everything downstream of it.
- Names the dashboards that depend on it.
- Over-large requests, gaps in upstream data and clashes with scheduled runs are refused or flagged.
Approval
Nothing runs until you approve.
The agent prepares the backfill and stops. Your backfill policy sets what is small enough to start on its own, what always needs sign-off, and how much of the warehouse one backfill may use.
- Caps on dispatch and concurrency.
- A preview says whether a backfill of a given size would wait for a person.
- A workspace kill switch stops backfills at once.
Verification
Checked after it runs.
Rows before the window must be unchanged, and a failing test marks the run partial rather than successful. A stopped run can be resumed from its last checkpoint.
In detail
What is in it.
Chunks and checkpoints
Airflow runs the window in chunks, with checkpoints, resume and cancel.
Version pin
The run uses the commit the plan was made on. The pin is checked at dispatch.
Backfills page
Live status, phases, chunk windows and a needs-attention view.
On the record
Each backfill keeps its range, its chunks, its commit and a link back to the conversation.
Making a model backfillable
The agent can add the dbt settings a model needs, guided by 14 checks, as a pull request.
Alerts
A backfill that ends badly raises an alert, which goes to triage.
Good to know. Requests longer than 3,660 days or wider than 250 models are refused.
Who decides
A person decides. The agent prepares.
The agent has only the permissions of the person it works for, and the approver is recorded.
These wait for a person
- Approving a backfill beyond your policy limits
- Cancelling a run
- Changing the backfill policy
- Merging a change that makes a model backfillable
Transcript of the film
A logic fix just shipped. Now a month of history is wrong, and fixing it usually means a runbook and a long night.
Ask for it the way you’d ask a teammate.
First it checks the model can be backfilled safely, and that the code it will use is what is actually deployed.
When something isn’t safe, it stops. Here the source data for August simply isn’t there yet.
For June, it plans the run: daily chunks, a few at a time, the cost, and the dashboard that depends on it.
Nothing runs until you approve it.
It dispatches to Airflow, and you follow it chunk by chunk, on the Backfills page.
And every backfill stays on the record: its range, its chunks, the commit it was pinned to, and a link back to the conversation.
Backfills, planned, approved and tracked. NerveStax.
More of the platform
The other features, each with a film.
See it on your own repository.
Tell us what you run and what is waiting. We will set up a working session on your stack and reply within a working day.