ML Feature Platform Architecture

Replace {{PLACEHOLDERS}} with your environment values, then deploy to your stack.

# ML Feature Platform Architecture

> AI Architecture · {{ORGANIZATION_NAME}}

## Overview

Architecture for batch + online feature engineering, feature store, and model training/serving integration.

## Architecture Diagram

```
┌──────────────┐     batch      ┌──────────────┐     train      ┌──────────────┐
│ Data Sources │ ─────────────▶│ Feature Store│ ─────────────▶│ ML Training  │
│ Lake/Warehouse│               │  (offline)   │               │  Pipeline    │
└──────────────┘               └──────┬───────┘               └──────┬───────┘
                                      │                               │
                              stream  │                               ▼
                                      ▼                        ┌──────────────┐
                               ┌──────────────┐   online     │ Model Registry│
                               │ Online Store │ ◀────────────│ + Serving     │
                               │ Redis/Dynamo │              └──────────────┘
                               └──────────────┘
```

## Feature Definition Workflow

1. Data scientist defines feature in {{FEATURE_REPO}} (YAML/Python)
2. Register metadata: owner, freshness SLA, entity key, description
3. CI validates: no leakage, point-in-time correctness tests
4. Deploy batch transformation job to {{SPARK}}/dbt
5. Materialize to offline store (Parquet/BigQuery) + online store sync

## Point-in-Time Correctness

- Training datasets use `as_of_timestamp` joins - never future data
- Feature views versioned: `customer_features_v{{VERSION}}`
- Backfill workflow for historical retrains

## Online Serving Workflow

1. Model inference request with entity ID
2. Fetch precomputed features from online store (< {{P99_MS}} ms)
3. On cache miss: compute from stream or fallback to defaults
4. Return prediction + feature contribution metadata (optional SHAP)

## Monitoring

- Feature drift detection (PSI, null rate)
- Serving latency and cache hit rate
- Training/serving skew alerts

## Stack Placeholders

| Component | {{YOUR_TOOL}} |
|-----------|---------------|
| Feature store | Feast / Tecton / SageMaker FS |
| Offline storage | {{LAKE/WH}} |
| Online storage | Redis / DynamoDB |
| Serving | KServe / SageMaker / custom |