Tutorial: Stream fraud velocity features with Feast

For the complete documentation index see: llms.txt

All documentation pages available in markdown.

Objectives

By the end of this tutorial, you will be able to:

Define velocity features on a PushSource and write fresh values to Aerospike with store.push().
Identify the freshness responsibility and boundary of a push-based feature pipeline.
Configure Aerospike record TTL, then read pushed values and inspect their event timestamps.

Introduction

Imagine a payment app with fraud protection. On the backend, a fraud detection model runs whenever a customer starts a transaction. It has to return an approve or decline decision fast enough to keep checkout smooth. To evaluate the transaction, the model needs recent history for the account behind it:

How many transactions did this account place in the last five minutes?
How much was spent across those transactions?
How many distinct merchants were involved?

Those values change continuously as new transactions arrive, so the model needs an answer that is both current and available in a single fast read.

The feature values the model needs are called velocity features: rolling aggregates that capture recent behavior. Two approaches fall short, for different reasons:

On-the-fly aggregation: The model reads the account’s recent transactions and aggregates them while scoring. That work lands inside the request and can exceed the latency budget.
Scheduled batch materialization: A job recomputes the aggregates on a fixed interval and loads the results into the online store. Scoring stays fast, but a five-minute window recomputed hourly is stale for most of the hour.

Both point at the same fix, and the difference is cadence rather than architecture: compute the aggregates outside the request, but do it continuously instead of on a schedule. Feast models this with a PushSource. A background process consumes transactions as they arrive, maintains the rolling windows, and calls store.push() to write each new value into Aerospike. The fraud model reads the stored value with get_online_features() at scoring time and performs no aggregation of its own.

Why push pre-computed values: The tradeoff is that freshness now depends on that background process. If it stalls, Feast keeps returning the last value written. Push cadence, event-timestamp monitoring, and Aerospike record time to live (TTL) bound how old a served value can become.

In this tutorial, you push one velocity snapshot manually. It does not run a transaction consumer or maintain rolling windows.

Key terms

Term	Meaning
Velocity features	Rolling aggregates that capture recent account behavior, such as transaction count and spend in the last five minutes.
Rolling window	A time-based sliding aggregate that updates continuously as new events arrive.
Velocity snapshot	The latest velocity feature values for one entity at a point in time, written to Aerospike with `store.push()`.
`PushSource`	Feast data source for features written through `store.push()`.
`store.push()`	API that writes feature values to the online store (and optionally offline).
`feast apply`	Command that reads your project’s Python definitions and registers the entities, feature views, and feature services in the Feast registry.
`FeatureService`	Named bundle of feature views your model requests together.
`ttl_seconds`	Online-store time to live (TTL) stamped on Aerospike records at write time.

When to use this pattern

Use this tutorial when your model needs features that reflect events from the last few minutes and you have, or plan to build, a process that computes those values before scoring.

If the Aerospike online store layout is unfamiliar, start with Tutorial: Serve real-time Feast features with Aerospike. If you need velocity features on a dedicated memory namespace, see the tiering tutorial after this one.

What you do

Define a push-backed feature view and register it with feast apply.
Push a velocity snapshot manually and read it back through the Python SDK.
Fetch features over HTTP using feast serve.

The wrap-up explains how to connect a stream job after the demo works.

On the next page, you install Feast and Aerospike locally.