Skip to content

Fetching multiple Feature Vectors with an Online Serving Index

Tecton allows you to fetch a set of features based on specifying a subset of entity IDs. This functionality is commonly used when multiple candidates need to be scored, such as in a recommendation system.

In this example, we'll show how to retrieve feature vectors for all ads a user has seen in the past week. We'll walk through:

  • Configuring the feature view with an online serving index
  • Retrieving features online
  • Creating training data

Configuring your feature views

First, when defining the feature view, you need to specify the online_serving_index parameter and omit the key you won't use during retrieval. In this case, we will specify the user at feature retrieval time, and get back a row for each ad they have feature values for.

from tecton import stream_window_aggregate_feature_view, Input, FeatureAggregation
from core.entities import user
from ads.entities import ad
from ads.data_sources.ad_impressions_stream import ad_impressions_stream
from datetime import datetime


@stream_window_aggregate_feature_view(
    inputs={"ad_impressions": Input(ad_impressions_stream)},
    entities=[user, ad],
    online_serving_index=["user_uuid"] # Only the user_uuid will be used at retrieval time
    mode="spark_sql",
    aggregation_slide_period="1h",
    aggregations=[FeatureAggregation(column="impression", function="count", time_windows=["1h", "12h", "24h","72h","168h"])],
    online=False,
    offline=False,
    batch_schedule="1d",
    feature_start_time=datetime(2021, 1, 1),
    family='ads',
    tags={'release': 'production'},
    owner="matt@tecton.ai",
    description="The count of impressions between a given user and a given ad"
)
def user_ad_impression_counts(ad_impressions):
    return f"""
        select
            user_uuid as user_id,
            ad_id,
            1 as impression,
            timestamp
        from
            {ad_impressions}
        """

Now that we've specified our serving indices for the Feature View, let's create our Feature Service to enable online retrieval.

from tecton import FeatureService, FeaturesConfig
from feature_repo.shared.features.user_ad_impression_counts_wildcard import user_ad_impression_counts_wildcard

ctr_prediction_service = FeatureService(
    name='ctr_prediction_service',
    description='A Feature Service used for supporting a CTR prediction model.',
    online_serving_enabled=True,
    features=[
        user_ad_impression_counts
    ],
    family='ad_serving',
    tags={'release': 'production'},
    owner="derek@tecton.ai",
)

Fetching wildcard features online

Once those changes have been applied, we can use the Tecton python library to retrieve a dataframe representing all the features that match our user by omitting the ad_id join key.

import tecton

my_fs = tecton.get_feature_service("ctr_prediction_service")

keys = {
        "user_uuid": "sample-user-uuid"
}

response = my_fs.query_features(keys).to_pandas()
print(response.head())

Alternatively, we can use the HTTP API. See the section above for more detail on how to configure the API key.

$ export TECTON_API_KEY='<your_tecton_key>'

$ curl -X POST https://<your_cluster>.tecton.ai/api/v1/feature-service/get-features\
     -H "Authorization: Tecton-key $TECTON_API_KEY" -d\
'{
  "params": {
    "feature_service_name": "ctr_prediction_service",
    "join_key_map": {
      "user_uuid": "sample-user-id",
    }
  }
}'

Creating training sets with wildcard features

Similarly, we can construct our training dataset by providing a prediction context that contains the join key we specified as our serving index.

import tecton

events = spark.read.parquet("dbfs:/event_data.pq").select("user_uuid", "timestamp")

my_fs = tecton.get_feature_service("ctr_prediction_service")

training_set = fs.get_feature_dataframe(events, timestamp_key="timestamp")

print(training_set.to_pandas().head())