feature-store

AI/ML Feature Store

Demonstrates ArcadeDB as a unified feature store for production ML systems, replacing the typical scatter of 3+ specialized databases (graph DB, vector DB, time-series DB) with a single multi-model engine serving three ML teams:

• Fraud scoring — graph topology + behavior vectors + transaction velocity
• Product recommendations — collaborative filtering + product embeddings
• Predictive maintenance — equipment dependency graph + sensor aggregates

A cross-cutting Feature Store infrastructure layer records feature snapshots for audit and lineage, demonstrating training-serving consistency.

Prerequisites

• Docker and Docker Compose
• curl and jq
• Java 21+ and Maven 3.x (for the Java demo)
• Node.js 22+ (for the JavaScript demo)

Quickstart

1. Start ArcadeDB

docker compose up -d

2. Create database and load data

./setup.sh

This creates the FeatureStore database, applies the schema, and inserts sample data.

3a. Run queries via curl

./queries/queries.sh

3b. Run queries via Java

cd java
mvn package -q
java -jar target/feature-store.jar

3c. Run queries via JavaScript (PostgreSQL protocol)

cd js
npm install
node feature-store.js

Schema

Type	Kind	Key properties
Account	Vertex	accountId, accountType, signupSource, flagged, behaviorVec
Merchant	Vertex	merchantId, category, riskTier
User	Vertex	userId, preferenceVec
Product	Vertex	productId, name, category, price, embedding
Equipment	Vertex	equipmentId, name, specifications, failureRate
Sensor	Vertex	sensorId, sensorType, unit
TRANSFERRED	Edge	Account → Account (amount, recordedAt)
LINKED_DEVICE	Edge	Account → Account (deviceId)
TRANSACTED	Edge	Account → Merchant (amount, recordedAt)
PURCHASED	Edge	User → Product
DEPENDS_ON	Edge	Equipment → Equipment (criticality)
MONITORED_BY	Edge	Equipment → Sensor
TransactionMetric	Document	accountId, txCount, totalAmount, recordedAt
SensorReading	Document	equipmentId, temperature, vibration, pressure, recordedAt
FeatureSnapshot	Document	entityId, entityType, featureVector, computedAt, modelVersion

Query Patterns

#	Pattern	Language	Signal type	Domain
1	Account Graph Features	SQL MATCH	Graph	Fraud
2	Distance to Flagged Account	SQL MATCH	Graph	Fraud
3	Behavior Similarity Search	SQL + vectorNeighbors	Vector	Fraud
4	Transaction Velocity	SQL	Time-series	Fraud
5	Shared Device Network	Cypher	Graph	Fraud
6	Collaborative Filtering	Cypher	Graph	Recommendations
7	Product Embedding Search	SQL + vectorNeighbors	Vector	Recommendations
8	Category Vector Search	SQL + vectorNeighbors	Vector	Recommendations
9	Equipment Dependency Chain	SQL MATCH	Graph	Maintenance
10	Sensor Anomaly Detection	SQL	Time-series	Maintenance
11	Feature Vector Assembly	SQL (multi-step)	All	Cross-domain

Sample Data

Fraud Domain

• 6 accounts (a1–a3 legit, a4–a5 suspicious, a6 flagged) with 4-d behavior vectors
• 4 merchants (grocery, electronics, gambling, crypto)
• ~12 TRANSFERRED edges with circular money flow pattern (a4↔a5↔a6)
• 4 LINKED_DEVICE edges (a4, a5 share devices with flagged a6)
• 15 TransactionMetric documents across 3 time buckets

Recommendation Domain

• 5 users with 4-d preference vectors
• 8 products across Electronics, Books, Sports with 4-d embeddings
• 14 PURCHASED edges with deliberate overlap (u1+u2 share purchases → collab signal)

Maintenance Domain

• 5 equipment units with dependency chain (eq2, eq3, eq4 depend on eq1)
• 3 sensors monitoring equipment
• 12 SensorReading documents (eq1 showing anomalous high temperature/vibration)

Feature Store Infrastructure

• 3 FeatureSnapshot documents for audit trail

Connectivity

Runner	Protocol	Port
curl / shell	HTTP API	2480
Java (arcadedb-network)	HTTP API	2480
JavaScript (pg)	PostgreSQL wire protocol	5432

The JavaScript module uses the {cypher} prefix for Cypher queries over PostgreSQL protocol.

ArcadeDB Version Notes

This use case targets ArcadeDB 26.4.2. Vector similarity queries use vectorNeighbors('TypeName[property]', vector, k) with an LSM_VECTOR index.

Reference

ArcadeDB AI/ML Feature Store use case