Platform For AI:Feature Database FeatureDB

High-speed VPC connectivity

FeatureDB provides high-speed VPC connectivity through PrivateLink. After successful configuration, FeatureStore uses PrivateLink to connect your VPC to the FeatureDB service. You can then access FeatureDB over a private connection from within your VPC using the FeatureStore SDK, improving data read/write performance and reducing access latency.

Configure high-speed VPC connectivity as follows:

• Method 1: If you have not created a FeatureDB data source, click Manage Data Source . On the Manage Data Source page, click Create Data Source . When creating a FeatureDB data source, fill in the VPC Direct Connection Configurations section with your VPC and Zone and vSwitch. For detailed steps, see Create an online data source: FeatureDB.

• Method 2: If you have already created a FeatureDB data source, click Manage Data Source . On the Manage Data Source page, click the existing feature_db data source to go to its product page. Click VPC Direct Connection Configurations, enter your VPC and Zone and vSwitch, then click Confirm.

Write features

For offline features, use the FeatureStore Python SDK with DataWorks to run daily scheduled tasks that sync data from MaxCompute to FeatureDB.

By default, real-time feature writes perform full-row updates. If the written data includes only partial fields, missing fields are set to empty. To update only specified fields and merge them into existing data, configure as follows:

• When writing via Java SDK: Specify the InsertMode as InsertMode.PartialFieldWrite.

  for (int i = 0; i < 100;i++) {
      featureView.writeFeatures(writeData, InsertMode.PartialFieldWrite);
  }

• When writing via Flink Connector: Set the parameter insert_mode to partial_field_write.

Read features

Use the FeatureStore SDK (Go/Java) or EasyRec Processor to read features.

The FeatureStore SDK (Go/Java) supports KV point-in-time joins for offline and real-time features. Specify the feature’s JoinID (primary key) and feature name to complete a KV query and retrieve target feature data in milliseconds. The SDK also supports KKV queries for behavior sequence features. Specify a UserID to retrieve assembled sequence feature results.

EasyRec Processor integrates the FeatureStore C++ SDK, enabling full ingestion of FeatureDB feature data into memory and millisecond-level polling to update real-time features in memory for higher read performance.

Monitoring metrics

If you use FeatureDB as an online store, after creating a feature view, click Data Monitoring next to the target view to view metrics such as read/write QPS and RT. (Note: Writes from offline feature view data sync do not appear in these metrics.)

Real-time feature pipeline

FeatureStore’s storage service consists of three components: Feature Service (access layer), message queue (DataHub), and FeatureDB.

For real-time features, you can call the Feature Service via the FeatureStore Java SDK or Flink Connector to write feature data into FeatureDB. Data written through Feature Service is also synced to your MaxCompute table for real-time feature sample export and further model training.

You can read feature data stored in FeatureDB using the FeatureStore Java/Go SDKs or pull all features into local cache via EasyRec Processor for higher read performance. Real-time features deliver the latest feature information within milliseconds.

Real-time feature data lifecycle settings

When creating a real-time feature view, set the data lifecycle for the FeatureDB table using Feature Lifecycle . Rows expire and are automatically cleaned up within seconds after reaching their lifecycle duration.

Data time-to-live (TTL) currently supports two configuration methods:

• Method 1: Do not specify an Event Time field. In this case, lifetime starts from the data write time, and rows are automatically cleaned up upon expiration.

• Method 2: Select a feature field as the Event Time field (unit: milliseconds). Let event_time be the value of the Event Time field, time_now be the current time, and time_ttl = time_now - ttl be the threshold event_time for expiration. FeatureDB handles written data as follows:

  • If using PartialFieldWrite mode for partial-field updates, lifetime is calculated from the actual write time.

  • If event_time > time_now + 15 min: Data is not written. (This prevents timestamp discrepancies between systems by allowing a 15-minute buffer.)

  • If time_ttl < event_time <= time_now + 15 min: Data is written normally. Lifetime starts from event_time, and the row is automatically cleaned up upon expiration.

  • If 0 < event_time < time_ttl: Data is written but immediately cleaned up. Note: event_time is in milliseconds. If your Event Time field uses seconds, data falls into this case and fails to persist.

  • If event_time <= 0: Lifetime is calculated from the actual write time.

  • If the value is invalid (cannot be converted to an integer): Data is not written.

  • The Event Time field is registered but no value is provided for the Event Time field: data is written normally, and its lifetime is calculated based on the actual write time.

  • If no Event Time field is registered: Data is written normally, and lifetime is calculated from the actual write time.

  • Additionally, FeatureDB uses the event_time value as the row’s timestamp (ts). To update data for a given key, the new Event Time value must be greater than or equal to the previous value. If the new event_time is less than the original, the key’s data will not be updated.