Streaming import

Streaming import lets you make real-time updates to feature values. This method is useful when having the latest available data for online serving is a priority. For example, you can import streaming event data and, within a few seconds, Vertex AI Feature Store (Legacy) makes that data available for online serving scenarios.

If you must backfill data or if you compute feature values in batch, use batch import. Compared to streaming import requests, batch import requests can handle larger payloads but take longer to complete.

For information about the oldest feature value timestamp that you can import, see Vertex AI Feature Store (Legacy) in Quotas and limits. You can't import feature values for which the timestamps indicate future dates or times.

Example use case

An online retail organization might provide a personalized shopping experience by using the current activity of a user. As users navigate through the website, you can capture their activity into a featurestore and then, soon after, serve all that information for online predictions. This real-time import and serving can help you show useful and relevant recommendations to customers during their shopping session.

Online storage node usage

Writing feature values to an online store uses the featurestore's CPU resources (online storage nodes). Monitor your CPU usage to check that demand doesn't exceed supply, which can lead to serving errors. We recommend around a 70% usage rate or lower to avoid these errors. If you regularly exceed that value, you can update your featurestore to increase the number of nodes or use autoscaling. For more information, see Manage featurestores.

Streaming import

Write a value to a particular feature. The feature value must be included as part of the import request. You can't stream data directly from a data source.

If you're writing to recently created features, wait a few minutes before you do so because the new features might not have propagated yet. If you don't, you might see a resource not found error.

You can import feature values for only one entity per write. For any specific project and region, you can simultaneously write feature values for multiple entities within a maximum of ten different entity types. This limit includes streaming import requests to all featurestores in a given project and region. If you exceed this limit, Vertex AI Feature Store (Legacy) might not write all of your data to the offline store. If this occurs, Vertex AI Feature Store (Legacy) logs the error in the Logs Explorer. For more information, see Monitor offline storage write errors for streaming import.


To import feature values for existing features, send a POST request by using the featurestores.entityTypes.writeFeatureValues method. If the names of the source data columns and the destination feature IDs are different, include the sourceField parameter. Note that featurestores.entityTypes.writeFeatureValues lets you import feature values for only one entity at a time.

Before using any of the request data, make the following replacements:

  • LOCATION: Region where the featurestore is created. For example, us-central1.
  • PROJECT: Your project ID.
  • FEATURESTORE_ID: ID of the featurestore.
  • ENTITY_TYPE_ID: ID of the entity type.
  • FEATURE_ID: ID of an existing feature in the featurestore to write values for.
  • VALUE_TYPE: The value type of the feature.
  • VALUE: Value for the feature.
  • TIME_STAMP (optional): The time at which the feature was generated. The timestamp must be in the RFC3339 UTC format.

HTTP method and URL:


Request JSON body:

  "payloads": [
      "entityId": "ENTITY_ID",
      "featureValues": {
        "FEATURE_ID": {
          "VALUE_TYPE": VALUE,
          "metadata": {"generate_time": "TIME_STAMP"}

To send your request, choose one of these options:


Save the request body in a file named request.json, and execute the following command:

curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json; charset=utf-8" \
-d @request.json \


Save the request body in a file named request.json, and execute the following command:

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
-Method POST `
-Headers $headers `
-ContentType: "application/json; charset=utf-8" `
-InFile request.json `
-Uri "" | Select-Object -Expand Content

You should receive a successful status code (2xx) and an empty response.


To learn how to install or update the Vertex AI SDK for Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation.

from import aiplatform

def write_feature_values_sample(
    project: str, location: str, entity_type_id: str, featurestore_id: str

    aiplatform.init(project=project, location=location)

    my_entity_type = aiplatform.featurestore.EntityType(
        entity_type_name=entity_type_id, featurestore_id=featurestore_id

    my_data = {
        "movie_01": {
            "title": "The Shawshank Redemption",
            "average_rating": 4.7,
            "genre": "Drama",


Additional languages

You can install and use the following Vertex AI client libraries to call the Vertex AI API. Cloud Client Libraries provide an optimized developer experience by using the natural conventions and styles of each supported language.

What's next