A run is not a list of tasks: it is a graph. Bronze reads the source; Silver reads Bronze; Gold reads Silver (or several). Declaring those dependencies is what lets the product order the tasks and — more importantly — not fake success when something broke earlier.
Where the dependency is declared#
In the Transformations step of the pipeline wizard,
each Silver/Gold stage says what it depends on (depends on). The reference is the
destination table of the previous task (e.g.: bronze:tb_vendas), not the name of the
file or of the source table.
On the run's page, the Graph tab shows the tasks and the arrows between them (track a run).
The dependency gate (what changes in practice)#
The next layer does NOT run when the source failed
If the Bronze of vendas failed in this run, the Silver that depends on it does not
run — even if the Bronze table exists in the datalake with the previous day's
load. Before this rule, the Silver "passed" by reading stale data and reported a
number that didn't match the day's load. Now it shows up as
blocked by dependency, with the reason.
This holds for both executors (the scheduled run and the manual run): the rule lives in a single module and is the same on both paths.
How the name is resolved#
To decide whether a task depends on a Bronze that failed, the product compares the declared reference with the Bronzes of the version that ran, in this order:
| Answer | When | What happens |
|---|---|---|
| Identity | the reference is exactly the Bronze's destination table (no accents, no case) | that is the dependency; it failed → it blocks |
| Alias | the reference matches a form derived from the source (the file name, with or without extension) and points to a single Bronze | that is the dependency; it failed → it blocks |
| Ambiguous | the same alias matches two Bronzes (e.g.: clientes.csv and clientes.json) | the product does not choose: it blocks if any of them failed and records the ambiguity |
| Nonexistent | no Bronze matches | the task is not blocked (there is no failed source), but the run records that the declared dependency does not exist — fix the declaration |
An alias never overrides an identity. If you want full predictability, always declare by the destination table.
Best practices#
- Name destinations without collisions. Two sources with the same base name
(
clientes.csvandclientes.json) generate ambiguous aliases; give them distinct destinations (tb_clientes_crm,tb_clientes_erp). - A Gold that reads two Silvers declares both. The run order follows the graph; the gate considers every declared source.
- An upstream failure is the day's failure. When a Bronze fails, handle the recovery starting from it; rerunning only the Gold does not bring in new data.
Related: Track a run · Recover from a failure · Publish and version · Pipeline and medallion.