# Build Feature Store

> Build a feature store using Feast for centralized feature management, configure offline and online stores for batch and real-time serving, define feature views with transformations, and implement point-in-time correct joins for ML pipelines. Use when managing features for multiple ML models, ensuring training-serving consistency, serving low-latency features for real-time inference, reusing feature definitions across projects, or building a feature catalog for discovery and governance.

- Skill: `pjt222/build-feature-store-7` (Agent Skill)
- Install (CLI): `npx skillmds@latest add pjt222/build-feature-store-7`
- Raw SKILL.md: https://api.skillmd.com/api/skills/pjt222/build-feature-store-7/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Product & Planning
- License: MIT
- Author: pjt222 (https://skillmd.com/u/pjt222)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/pjt222/build-feature-store-7

---


# Build Feature Store


> See [Extended Examples](references/EXAMPLES.md) for complete configuration files and templates.

以 Feast 實作集中式特徵管理，確訓練與推理間特徵服務之一致。

## 適用時機

- 為多團之多 ML 模型管特徵
- 確特徵訓服一致
- 實作時點正確之歷史特徵
- 為實時推理提低延之特徵
- 跨項目復用特徵定義
- 版特徵轉換
- 建可發現與治理之特徵目錄
- 止訓練管線中之特徵洩

## 輸入

- **必要**：原數源（數據庫、數據湖、數據倉）
- **必要**：裝 Feast 之 Python 環境
- **必要**：離線倉後端（BigQuery、Snowflake、Redshift、或 Parquet 文件）
- **必要**：在線倉後端（Redis、DynamoDB、Cassandra、或開發用 SQLite）
- **選擇性**：特徵轉換邏輯（Python、SQL、Spark）
- **選擇性**：實體鍵定義（user_id、product_id 等）
- **選擇性**：Feast 伺服器部署之 Kubernetes 集群

## 步驟

### 步驟一：初始化 Feast 特徵庫

立 Feast 項目結構並配存儲後端。

```bash
# Install Feast with required extras
pip install 'feast[redis,postgres]'  # Add backends as needed

# Initialize new feature repository
feast init my_feature_repo
cd my_feature_repo

# Directory structure created:
# my_feature_repo/
# ├── feature_store.yaml       # Configuration
# ├── features.py              # Feature definitions
# └── data/                    # Sample data (dev only)
```

配 `feature_store.yaml`：

```yaml
# feature_store.yaml
project: customer_analytics
registry: data/registry.db  # SQLite for dev, use S3/GCS for prod
provider: local

# Offline store for training data
offline_store:
  type: postgres
# ... (see EXAMPLES.md for complete implementation)
```

生產配，以雲後端：

```yaml
# feature_store.prod.yaml
project: customer_analytics
registry: s3://feast-registry/prod/registry.db
provider: aws

offline_store:
  type: bigquery
  project_id: my-gcp-project
# ... (see EXAMPLES.md for complete implementation)
```

**預期：** Feast 庫已初始化附配文件，樣本特徵定義已創，離／在線倉已配，註冊路可達。

**失敗時：** 驗數據庫／Redis 憑（`psql -U feast_user -h localhost`），查連接字串格式，確數據庫存（`CREATE DATABASE feature_store`），驗 S3/BigQuery/DynamoDB 之雲權，測存儲後端連通，查 Feast 版與後端之相容（`feast version`）。

### 步驟二：定實體與數源

創實體定義並連原數源。

```python
# entities.py
from feast import Entity, ValueType

# Define entities (primary keys for features)
customer = Entity(
    name="customer",
    description="Customer entity",
    value_type=ValueType.INT64,
# ... (see EXAMPLES.md for complete implementation)
```

定數源：

```python
# data_sources.py
from feast import FileSource, BigQuerySource, RedshiftSource
from feast.data_format import ParquetFormat
from datetime import timedelta

# Development: File-based source
customer_transactions_source = FileSource(
    path="data/customer_transactions.parquet",
# ... (see EXAMPLES.md for complete implementation)
```

**預期：** 實體定義引正 ID 列，數源成連原數，event_timestamp_column 於源中存，created_timestamp_column 允時點查詢。

**失敗時：** 驗源數據文件存且可讀，查 BigQuery/Redshift 憑與表存取，確時戳列格式正確（Unix 時戳或 ISO8601），驗 Kafka 連通與題存，查源與實體間之綱相容。

### 步驟三：定附轉換之特徵視圖

創定原數如何化為 ML 就緒特徵之特徵視圖。

```python
# feature_views.py
from feast import FeatureView, Field
from feast.types import Float32, Int64, String, Bool
from datetime import timedelta
from entities import customer, product
from data_sources import customer_features_source

# Simple feature view without transformations
# ... (see EXAMPLES.md for complete implementation)
```

**預期：** 特徵視圖成註，綱合源數，轉換無錯執，TTL 值合用例，on-demand 視圖合批與請特徵。

**失敗時：** 驗欄名與源列完合，查 dtype 相容（Int64 對 Int32），確實體引存，以樣本數驗轉換邏輯，查算中除零，驗請源綱合推理載荷。

### 步驟四：應特徵定義並物化特徵

部特徵定義至註冊並物化至在線倉。

```bash
# Apply feature definitions to registry
feast apply

# Expected output:
# Created entity customer
# Created feature view customer_stats
# Created on demand feature view customer_segments

# ... (see EXAMPLES.md for complete implementation)
```

程序化物化：

```python
# materialize_features.py
from feast import FeatureStore
from datetime import datetime, timedelta

# Initialize feature store
fs = FeatureStore(repo_path=".")

# Materialize all feature views
# ... (see EXAMPLES.md for complete implementation)
```

**預期：** 特徵定義應至註冊而無衝，物化作業成完，在線倉以特徵填，特徵新鮮於所配 TTL 內。

**失敗時：** 查離線倉查詢成（`feast feature-views describe customer_stats`），驗時段有數，確在線倉可寫（Redis/DynamoDB 權），查特徵名於視圖間無重，驗實體鍵於源數存，監物化作業日誌求錯，查本地倉之盤空。

### 步驟五：為訓練取特徵

取時點正確之歷史特徵以供模型訓練。

```python
# get_training_data.py
from feast import FeatureStore
import pandas as pd
from datetime import datetime

# Initialize feature store
fs = FeatureStore(repo_path=".")

# ... (see EXAMPLES.md for complete implementation)
```

時點正確性驗：

```python
# validate_pit_correctness.py
import pandas as pd
from datetime import datetime, timedelta

def validate_point_in_time_correctness(training_df, entity_df):
    """
    Ensure features don't leak future information.
    """
# ... (see EXAMPLES.md for complete implementation)
```

**預期：** 歷史特徵成取，entity_df 時戳保全，已物化特徵無 NaN，時點正確性保（無未來數洩），特徵服務合邏輯組特徵。

**失敗時：** 查 entity_df 具必列（實體名加 event_timestamp），驗特徵視圖名合註冊，確離線倉於所請時段有數，查時區失配（用 UTC），驗實體 ID 於源數存，察日誌求 SQL 查詢錯，驗特徵視圖 TTL 覆所請時段。

### 步驟六：為實時推理服特徵

自在線倉取低延特徵以供模型服務。

```python
# serve_features.py
from feast import FeatureStore
import time

# Initialize feature store
fs = FeatureStore(repo_path=".")

def get_inference_features(customer_ids: list, request_data: dict = None):
# ... (see EXAMPLES.md for complete implementation)
```

FastAPI 整合：

```python
# api.py
from fastapi import FastAPI
from pydantic import BaseModel
from feast import FeatureStore
import mlflow

app = FastAPI()
fs = FeatureStore(repo_path=".")
# ... (see EXAMPLES.md for complete implementation)
```

**預期：** 在線特徵於單實體十毫秒內取，批取有效擴展，on-demand 轉換正執，請時特徵與批特徵合，API 快應（端到端小於五十毫秒）。

**失敗時：** 查在線倉已填（若空則物化），驗 Redis/DynamoDB 連通與延，確實體鍵於在線倉存，查冷啟問題（暖快取），驗 on-demand 轉換邏輯，監在線倉之記憶/CPU，查服務與在線倉間之網延。

## 驗證

- [ ] Feast 庫已初始化並配
- [ ] 離／在線倉成連
- [ ] 實體定義合源數
- [ ] 特徵視圖於註冊已註
- [ ] On-demand 轉換正執
- [ ] 物化無錯完
- [ ] 歷史特徵取附時點正確性
- [ ] 在線特徵以低延服（小於十毫秒）
- [ ] 特徵新鮮於所配 TTL 內
- [ ] 訓服一致已驗
- [ ] 特徵目錄可為發現存取

## 常見陷阱

- **特徵洩**：歷史特徵中用未來數——恆驗時點正確，用 created_timestamp 列
- **轉換不一致**：訓練對服務之異邏輯——用 Feast on-demand 視圖以求一致
- **陳特徵**：在線倉未定期物化——立排程物化作業（cron/Airflow）
- **實體鍵缺**：訓集中實體不在在線倉——確全面物化，優雅處缺鍵
- **類型失配**：綱類與源數不合——apply 前驗 dtype，用明 Field 定義
- **在線取慢**：網延或在線倉超載——服特徵倉與推理服同地，用連池
- **大特徵視圖**：物化數百萬實體慢——按日分區，用增量物化，優離線查詢
- **無特徵版**：破壞變影生產模型——版特徵視圖，維反向相容
- **時區混**：混時區致誤合——時戳恆用 UTC
- **忽 TTL**：服過期特徵——設合之 TTL，監特徵新鮮

## 相關技能

- `track-ml-experiments` - 於 MLflow 實驗中記特徵元數據
- `orchestrate-ml-pipeline` - 排程特徵物化作業
- `version-ml-data` - 版特徵工程之原數源
- `deploy-ml-model-serving` - 特徵倉與模型服務之整合
- `serialize-data-formats` - 擇特徵之高效存儲格式
- `design-serialization-schema` - 為特徵源設計綱