Question

Your team's Vertex AI Pipeline uses a KFP `dsl.ParallelFor` loop to train one model per geographic region in parallel. After the loop, you need a single downstream component to receive the list of validation accuracy metrics produced by every loop iteration so it can select and register the best-performing regional model. What is the correct way to pass the per-iteration outputs to the aggregation component?

Accepted Answer

Wrap the loop output with `dsl.Collected` and pass it as an `Input[List[...]]` (or parameter list) argument to the aggregation component.

Answer

Have each loop iteration write its metric to a hardcoded GCS path, then have the aggregation component list the bucket to read all files.

Answer

Add `.after()` on the ParallelFor task so the aggregation component automatically inherits all iteration outputs as inputs.

Answer

Reference the loop task's output directly (e.g., `loop_task.outputs['metric']`) since KFP returns the last iteration's value.