Rerunning without knowing the scope is like fighting a fire in the dark. On the run's detail screen, the Recover… button opens a panel that shows, before you confirm, exactly what will be run again.
Before you start#
- Be the workspace owner; subscription in good standing.
- Have read the log of the correct attempt — see track a run.
- Know the pipeline's load mode: it decides whether reprocessing replaces, merges, or appends.
The scopes#
| Scope | What it does |
|---|---|
| New task attempt | Repeats only that task, in the same run. |
| Rerun task and dependents | Repeats the task and everything that depends on it in the graph. |
| Run the pipeline again | Creates a new run of the published version, from the beginning. |
| Reprocess period | Reprocesses a date window (From day / To day, inclusive). It only truly slices the data when the Bronze table is Append with a key column; without that, the run reloads the entire table and the panel asks you to confirm that you understand. |
The preview matches the effect
The recovery panel lists the tasks that will be run again. If the preview shows a task, it runs; if it doesn't show it, it doesn't run. Check the preview before confirming.
Step by step#
1. Open the run that failed
ETL/ELT → Runs → the run → the task → the attempt.
2. Read the cause of the failure
Don't rerun before understanding the cause. An expired credential isn't solved by repeating the task: fix the connection first.
3. Open Recover…
In the run's detail, use Recover….
4. Choose the scope and review the impact
The panel has three steps: Choose scope → Review impact → Track result. Expected impact shows the tasks that will run, the ones left out, the destination, and the write effect. An unavailable scope appears with the reason, not as a dead button.
5. Confirm
The button describes the operation — for example Confirm rerun of {N} task(s) or Confirm reprocessing of {N} day(s). The recovery becomes a traceable run, like any other.
Expected result#
What the preview showed is what ran. The destination was checked: expected row count, keys, and known values in the test set.
"Mark as success" doesn't produce data
Marking a task as successful changes the run's status, not the content of the destination. Use it only when you know the data is already correct through another path — never as a way to "solve" a load that didn't happen.
If something goes wrong#
| Symptom | What to check | Next action |
|---|---|---|
| I reran and it duplicated rows | Load mode (appending without a key duplicates by definition) | See load modes |
| The preview shows more tasks than I expected | The dependencies in the graph | Choose New task attempt if you want the minimum scope |
| Reprocess period warns that it will reload everything | Whether the Bronze table is Append with a key column | Without that, the period is only recorded; see incremental ingestion and backfill |
| It failed again for the same reason | Whether the cause was fixed at the source | Fix the connection, the permission, or the schema before repeating |
| Billing block message | Subscription status | Settle it in Administration → Plans |
Limits and costs#
- Opening the preview: R$ 0. Confirming the recovery runs and consumes cloud resources like any run.
- There is no promise of time: it depends on the reprocessed volume.
Next task#
- Runs screens — the exact name of each button
- Handle a failure and the DLQ
- Scheduling, retry, and DLQ
Last review: 2026-10-03.