Skip to content

Wrap up

For the complete documentation index see: llms.txt

All documentation pages available in markdown.

You defined push-backed velocity features, wrote and read one velocity snapshot, and verified the HTTP feature server path. You also identified what the manual demo does not cover, as described in Production streaming pattern: event consumption, rolling-window state, retries, and aggregator recovery.

Stop Aerospike

  1. Stop the feature server if it is still running (Ctrl+C).

  2. Stop and remove the Aerospike container:

    Terminal window
    docker rm -f aerospike

Streaming integration

In production, a background process replaces push_velocity.py. The process consumes a transaction stream, maintains rolling-window aggregates, and calls store.push() on a flush interval. Common implementations include a Kafka consumer, an Apache Flink job using PyFlink, or a Spark Structured Streaming job. Each implementation aggregates events into a DataFrame with the same columns as txn_velocity_5m, then calls:

store.push("txn_velocity_push", dataframe, to=PushMode.ONLINE)

Use PushMode.ONLINE_AND_OFFLINE when you also need the pushed values written to an offline store for training history. First configure a writable batch source for the PushSource. A production data warehouse is recommended for durable training history, but Feast also supports writable file-based offline storage.