Vertex AI Batch Prediction Job

The Vertex AI Batch Prediction Job job type submits a bulk inference Batch Prediction Job to Google Vertex AI via POST projects/{project}/locations/{location}/batchPredictionJobs. The template job is identified by its resource name — Polysync stores projects/{project}/locations/{location}/batchPredictionJobs/{jobId} in the Job's External Id.

This job type is supported on the Google Vertex AI platform.

Required job fields

External Id — the template Batch Prediction Job resource name.
Job Type — Vertex AI Batch Prediction Job (set automatically on import).

Captured as job attributes from the template (immutable per run):

Model — the registered model's resource name (projects/{p}/locations/{l}/models/{id}).
Input Config — the input data source.
Output Config — the output destination.

Job discovery

GET projects/{project}/locations/{location}/batchPredictionJobs?pageSize=100

(paginated). inputConfig, outputConfig, and model are captured as job attributes.

Parameter handling

Only the modelParameters field is run-time variable. Polysync sends Input + Input&Output values there:

{
  "displayName": "<polysync-name>_<timestamp>",
  "model": "<Model attribute>",
  "inputConfig":  "<Input Config attribute>",
  "outputConfig": "<Output Config attribute>",
  "modelParameters": {
    "<param-1>": <typed-value>,
    "<param-2>": <typed-value>
  }
}

Direction	Sent in `modelParameters`	Updated from response
`Input`	✅	❌
`Output`	❌	❌ (not supported)
`Input&Output`	✅	❌ (not supported)

Output parameters are not supported. Predictions land in outputConfig (GCS / BigQuery).

Execution flow

Polysync POSTs the body above; the resource name becomes the Polysync RunId.
Status is polled via GET projects/{project}/locations/{location}/batchPredictionJobs/{jobId} with the same status mapping as Vertex AI Custom Job.
Cancel is supported via POST {resourceName}:cancel.

Monitor URL

https://console.cloud.google.com/vertex-ai/locations/{location}
  /batch-predictions/{jobId}?project={projectId}

Best practices

Maintain one template Batch Prediction Job per (model, input shape) combination — inputConfig / outputConfig are cloned per run.
Use modelParameters only for inference hyperparameters the model actually accepts; unknown keys are silently ignored.

Troubleshooting

Job fails immediately with INVALID_ARGUMENT — inputConfig points to a path the runtime identity can't read. Grant the Vertex AI service agent / your impersonated SA Storage Object Viewer on the bucket.
Empty predictions — your input format does not match what the model expects; check the model's Predict Schemata.

Documentation