Split View: Feature Store 구축 완전 가이드: Feast 아키텍처·온라인/오프라인 서빙·ML 파이프라인 통합

Complete Guide to Building a Feature Store: Feast Architecture, Online/Offline Serving, and ML Pipeline Integration

• Introduction
• 1. Why You Need a Feature Store
  • Problems Without Feature Management
  • What a Feature Store Solves
• 2. Feature Store Architecture
  • Offline Store
  • Online Store
  • Registry
• 3. Deep Dive into the Feast Framework
  • Core Concepts
• 4. Feast Installation and Project Setup
  • Installation and Initialization
  • Project Configuration (feature_store.yaml)
• 5. Feature View and Entity Definitions
  • Defining Entities and Feature Views
  • Feature Service Definition
• 6. Online/Offline Serving Implementation
  • Offline Serving (Training Data Generation)
  • Online Serving (Real-time Inference)
  • Materialization (Offline to Online Sync)
• 7. ML Pipeline Integration
  • Feature Pipeline with Airflow
  • Kubeflow Pipelines Integration
• 8. Feature Store Comparison
  • Key Differences Summary
• 9. Failure Scenarios and Resolution Strategies
  • Preventing Training-Serving Skew
  • Latency Optimization
  • Ensuring Data Consistency
• 10. Operations Checklist
  • Design Phase
  • Pipeline Configuration
  • Monitoring
  • Security and Governance
• Conclusion

Introduction

The production performance of ML models depends more on feature quality and consistency than on model architecture itself. The moment feature transformation logic written by a data scientist in a Jupyter Notebook subtly diverges from the logic implemented in the serving server, Training-Serving Skew emerges and model performance degrades dramatically.

A Feature Store is the core infrastructure that solves this problem at the architectural level. In this guide, we focus on Feast, the leading open-source Feature Store, covering everything from architecture design to online/offline serving implementation, ML pipeline integration with Airflow and Kubeflow, and comparative analysis with competing platforms like Tecton and Hopsworks. This is not a simple installation tutorial -- it is a production guide for teams managing millions of entities and dozens of Feature Views.

1. Why You Need a Feature Store

Problems Without Feature Management

Operating ML systems without a Feature Store leads to recurring issues:

1. Duplicated feature logic: The code computing avg_order_amount_30d in the training pipeline differs from the serving server code. Subtle differences in NULL handling, aggregation windows, and timezone handling cause skew.
2. Lack of feature discoverability: Team A has already computed user_click_rate_7d, but Team B recomputes it because they have no way to discover it. Feature assets become siloed across the organization.
3. No time travel: You cannot answer "What was this user's feature value 3 weeks ago?" -- making retraining and debugging impossible.
4. Serving latency: Computing features in real-time from multiple data sources during inference pushes p99 latency to hundreds of milliseconds.

What a Feature Store Solves

A Feature Store serves as the Single Source of Truth for features. By generating both offline training data and online serving data from a single feature definition, logic inconsistencies are eliminated at the root. A central registry enables feature search and reuse, while Point-in-Time Joins accurately reproduce features at any past point in time.

2. Feature Store Architecture

The core architecture of a Feature Store consists of three components: the Offline Store, the Online Store, and the Registry.

Offline Store

Stores large volumes of historical feature data and performs Point-in-Time Joins for training data generation. Uses data warehouses or data lakes like BigQuery, Snowflake, Redshift, or S3/Parquet as backends. Must efficiently scan data at the tens-of-terabyte scale.

Online Store

A low-latency key-value store for real-time serving. Uses Redis, DynamoDB, or Bigtable as backends, returning the latest feature values for entity keys within p99 of 10ms. Data is synchronized from the offline store through the Materialization process.

Registry

A central catalog storing metadata for entities, Feature Views, Feature Services, and more. Can be operated as file-based (local, S3, GCS) or SQL-based (PostgreSQL, MySQL). SQL-based registries are recommended for production to prevent concurrent access conflicts.

3. Deep Dive into the Feast Framework

Feast (Feature Store) is an open-source project initiated in 2019 by Gojek and Google, now managed under the Linux Foundation. Feast's core strength is its pluggable architecture -- you can add the Feature Store layer while continuing to use your existing infrastructure (Spark, Kafka, Redis, Snowflake, etc.).

Core Concepts

• Entity: A business object to which features are attached (e.g., user, product, driver)
• Feature View: A logical grouping of features derived from the same source
• Feature Service: A bundle of features used by a specific model
• Data Source: The origin of feature data (BigQuery, Parquet, Kafka, etc.)
• Materialization: The process of synchronizing data from the offline store to the online store

4. Feast Installation and Project Setup

Installation and Initialization

# Install Feast with Redis and PostgreSQL support
pip install 'feast[redis,postgres]'

# Initialize a project
feast init feature_repo
cd feature_repo

# Directory structure
# feature_repo/
#   feature_store.yaml    -- Project configuration
#   definitions.py        -- Entity and Feature View definitions
#   data/                 -- Sample data

Project Configuration (feature_store.yaml)

project: my_ml_platform
provider: gcp
registry:
  registry_type: sql
  path: postgresql://feast:feast@db-host:5432/feast_registry
  cache_ttl_seconds: 60
online_store:
  type: redis
  connection_string: redis-host:6379,password=secret
offline_store:
  type: bigquery
  dataset: feast_offline
entity_key_serialization_version: 2

Three key points in this configuration: First, the SQL-based registry supports concurrent access from multiple teams. Second, the online store uses Redis to guarantee millisecond-level responses. Third, the offline store uses BigQuery to handle large-scale historical joins.

5. Feature View and Entity Definitions

Defining Entities and Feature Views

from datetime import timedelta
from feast import Entity, FeatureView, Field, FileSource, BigQuerySource
from feast.types import Float32, Float64, Int64, String

# Entity definitions
customer = Entity(
    name="customer_id",
    description="Unique customer identifier",
)

driver = Entity(
    name="driver_id",
    description="Unique driver identifier",
)

# BigQuery source definition
customer_stats_source = BigQuerySource(
    name="customer_stats_source",
    table="my_project.feast_dataset.customer_stats",
    timestamp_field="event_timestamp",
    created_timestamp_column="created_timestamp",
)

# Feature View definition
customer_stats_fv = FeatureView(
    name="customer_stats",
    entities=[customer],
    ttl=timedelta(days=3),
    schema=[
        Field(name="total_orders", dtype=Int64),
        Field(name="avg_order_amount", dtype=Float64),
        Field(name="lifetime_value", dtype=Float64),
        Field(name="preferred_category", dtype=String),
        Field(name="churn_risk_score", dtype=Float32),
    ],
    source=customer_stats_source,
    online=True,
    tags={
        "team": "growth",
        "version": "v2",
    },
)

driver_stats_source = BigQuerySource(
    name="driver_stats_source",
    table="my_project.feast_dataset.driver_stats",
    timestamp_field="event_timestamp",
)

driver_stats_fv = FeatureView(
    name="driver_stats",
    entities=[driver],
    ttl=timedelta(hours=6),
    schema=[
        Field(name="avg_rating", dtype=Float64),
        Field(name="total_trips", dtype=Int64),
        Field(name="acceptance_rate", dtype=Float64),
        Field(name="avg_delivery_time_min", dtype=Float32),
    ],
    source=driver_stats_source,
    online=True,
)

Feature Service Definition

Bundle the features used by a specific model into a Feature Service for organized management.

from feast import FeatureService

# Feature Service for churn prediction model
churn_prediction_svc = FeatureService(
    name="churn_prediction_service",
    features=[
        customer_stats_fv[["total_orders", "avg_order_amount", "lifetime_value", "churn_risk_score"]],
    ],
    tags={
        "model": "churn_prediction_v3",
        "owner": "growth-team",
    },
)

# Feature Service for driver matching model
driver_matching_svc = FeatureService(
    name="driver_matching_service",
    features=[
        driver_stats_fv[["avg_rating", "acceptance_rate", "avg_delivery_time_min"]],
        customer_stats_fv[["preferred_category"]],
    ],
)

6. Online/Offline Serving Implementation

Offline Serving (Training Data Generation)

Offline serving uses Point-in-Time Joins to accurately retrieve feature values at specific past points in time. This is the core mechanism that prevents Feature Leakage.

from feast import FeatureStore
import pandas as pd

store = FeatureStore(repo_path="feature_repo/")

# Entity dataframe with labels for training
entity_df = pd.DataFrame({
    "customer_id": [1001, 1002, 1003, 1001, 1002],
    "event_timestamp": pd.to_datetime([
        "2026-01-15 10:00:00",
        "2026-01-15 11:00:00",
        "2026-01-16 09:00:00",
        "2026-02-01 10:00:00",
        "2026-02-01 11:00:00",
    ]),
    "churned": [0, 1, 0, 1, 0],  # Labels
})

# Generate training data with Point-in-Time Join
training_df = store.get_historical_features(
    entity_df=entity_df,
    features=[
        "customer_stats:total_orders",
        "customer_stats:avg_order_amount",
        "customer_stats:lifetime_value",
        "customer_stats:churn_risk_score",
    ],
).to_df()

print(training_df.head())
# customer_id | event_timestamp     | churned | total_orders | avg_order_amount | ...
# 1001        | 2026-01-15 10:00:00 | 0       | 42           | 35.50            | ...

Online Serving (Real-time Inference)

Online serving returns the latest feature values with millisecond-level latency.

# Online feature retrieval
online_features = store.get_online_features(
    features=[
        "customer_stats:total_orders",
        "customer_stats:avg_order_amount",
        "customer_stats:churn_risk_score",
    ],
    entity_rows=[
        {"customer_id": 1001},
        {"customer_id": 1002},
    ],
).to_dict()

print(online_features)
# Example output:
# {
#     "customer_id": [1001, 1002],
#     "total_orders": [45, 12],
#     "avg_order_amount": [35.50, 28.00],
#     "churn_risk_score": [0.15, 0.82],
# }

Materialization (Offline to Online Sync)

# Full Feature View materialization
feast materialize 2026-01-01T00:00:00 2026-03-10T00:00:00

# Incremental materialization (only changes since last run)
feast materialize-incremental 2026-03-10T00:00:00

7. ML Pipeline Integration

Feature Pipeline with Airflow

Automating Feast Materialization as an Airflow DAG enables reliable operations.

from airflow import DAG
from airflow.decorators import task
from datetime import datetime, timedelta

default_args = {
    "owner": "ml-platform",
    "retries": 3,
    "retry_delay": timedelta(minutes=5),
}

with DAG(
    dag_id="feast_materialization_pipeline",
    default_args=default_args,
    schedule_interval="0 */4 * * *",  # Every 4 hours
    start_date=datetime(2026, 1, 1),
    catchup=False,
) as dag:

    @task()
    def validate_source_data():
        """Validate source data quality"""
        from great_expectations import get_context
        context = get_context()
        result = context.run_checkpoint(checkpoint_name="feature_source_check")
        if not result.success:
            raise ValueError("Source data quality validation failed")
        return True

    @task()
    def materialize_features():
        """Materialize from offline store to online store"""
        from feast import RepoConfig, FeatureStore
        from feast.infra.online_stores.redis import RedisOnlineStoreConfig
        from feast.repo_config import RegistryConfig

        repo_config = RepoConfig(
            project="my_ml_platform",
            provider="gcp",
            registry=RegistryConfig(
                registry_type="sql",
                path="postgresql://feast:feast@db-host:5432/feast_registry",
            ),
            online_store=RedisOnlineStoreConfig(
                connection_string="redis-host:6379",
            ),
        )
        store = FeatureStore(config=repo_config)
        store.materialize_incremental(end_date=datetime.utcnow())
        return True

    @task()
    def validate_online_store():
        """Validate online store feature values"""
        from feast import FeatureStore
        store = FeatureStore(repo_path="feature_repo/")

        result = store.get_online_features(
            features=["customer_stats:total_orders"],
            entity_rows=[{"customer_id": 1001}],
        ).to_dict()

        if result["total_orders"][0] is None:
            raise ValueError("Features not loaded in online store")
        return True

    @task()
    def notify_completion():
        """Send Slack notification"""
        import requests
        requests.post(
            "https://hooks.slack.com/services/YOUR/WEBHOOK/URL",
            json={"text": "Feast Materialization completed successfully"},
        )

    validate_source_data() >> materialize_features() >> validate_online_store() >> notify_completion()

Kubeflow Pipelines Integration

In Kubeflow Pipelines, Feast tasks can be defined as individual components.

from kfp import dsl
from kfp.dsl import component, Output, Dataset

@component(
    base_image="python:3.10",
    packages_to_install=["feast[redis,postgres]>=0.40.0"],
)
def feast_materialize_op(
    project_name: str,
    registry_path: str,
    redis_connection: str,
):
    from feast import RepoConfig, FeatureStore
    from feast.infra.online_stores.redis import RedisOnlineStoreConfig
    from feast.repo_config import RegistryConfig
    from datetime import datetime

    config = RepoConfig(
        project=project_name,
        provider="gcp",
        registry=RegistryConfig(
            registry_type="sql",
            path=registry_path,
        ),
        online_store=RedisOnlineStoreConfig(
            connection_string=redis_connection,
        ),
    )
    store = FeatureStore(config=config)
    store.materialize_incremental(end_date=datetime.utcnow())

@component(
    base_image="python:3.10",
    packages_to_install=["feast[redis,postgres]>=0.40.0", "scikit-learn"],
)
def train_model_op(
    project_name: str,
    model_output: Output[Dataset],
):
    from feast import FeatureStore
    import pandas as pd
    from sklearn.ensemble import GradientBoostingClassifier
    import pickle

    store = FeatureStore(repo_path="feature_repo/")
    entity_df = pd.read_parquet("gs://my-bucket/training_entities.parquet")

    training_df = store.get_historical_features(
        entity_df=entity_df,
        features=[
            "customer_stats:total_orders",
            "customer_stats:avg_order_amount",
            "customer_stats:churn_risk_score",
        ],
    ).to_df()

    X = training_df.drop(columns=["customer_id", "event_timestamp", "churned"])
    y = training_df["churned"]

    model = GradientBoostingClassifier(n_estimators=200)
    model.fit(X, y)

    with open(model_output.path, "wb") as f:
        pickle.dump(model, f)

@dsl.pipeline(name="feast-ml-training-pipeline")
def feast_training_pipeline():
    materialize_task = feast_materialize_op(
        project_name="my_ml_platform",
        registry_path="postgresql://feast:feast@db-host:5432/feast_registry",
        redis_connection="redis-host:6379",
    )
    train_task = train_model_op(
        project_name="my_ml_platform",
    )
    train_task.after(materialize_task)

8. Feature Store Comparison

Key Differences Summary

• Feast: Maximum flexibility. Each component can be chosen to fit existing infrastructure. The team bears the operational burden directly.
• Tecton: Turnkey solution. Powerful streaming feature pipelines but expensive. Best for organizations where real-time ML is critical.
• Hopsworks: Strong data governance and audit logging, preferred in regulated industries like finance and healthcare. The RonDB-based online store achieves latency at just 15% of SageMaker's levels.
• SageMaker Feature Store: Convenient for organizations already deeply embedded in the AWS ecosystem, but carries vendor lock-in risk.

9. Failure Scenarios and Resolution Strategies

Preventing Training-Serving Skew

Training-Serving Skew does not fully disappear even after adopting a Feature Store. Here are common scenarios and mitigation strategies.

Scenario 1: Stale features due to TTL expiry

If the online store TTL is 6 hours but the Materialization batch fails and does not run for 12 hours, some features return null. The mitigation is to send immediate alerts on Materialization failure and set TTL to at least 3x the Materialization interval.

Scenario 2: Compatibility breakage on feature definition changes

Changing the aggregation window of avg_order_amount from 30 days to 90 days breaks compatibility with already-trained models. The solution is to never modify existing features -- instead, add new ones (e.g., avg_order_amount_90d).

Scenario 3: Timezone mismatch

If offline training data uses UTC but online source data uses local timezones, feature values will differ. All timestamps must be unified to UTC.

Latency Optimization

Causes and solutions for high online serving latency:

• Cause: Too many Feature Views queried in a single request.
• Solution: Bundle only needed features with Feature Services, and use batch queries (pass multiple entities in a single get_online_features call).
• Cause: Redis cluster hotkey issues.
• Solution: Use hash tags in entity keys to distribute keys evenly across the cluster.

Ensuring Data Consistency

Data consistency between the offline and online stores depends on Materialization. To reinforce this:

1. Run sampling validation tasks after Materialization to compare feature values between offline and online stores.
2. Build feature drift monitoring to detect abnormal changes in feature distributions.
3. Integrate data quality tools like Great Expectations into your source data pipelines.

10. Operations Checklist

A checklist for operating a production Feature Store reliably:

Design Phase

• Entity design: Have you defined entity keys that align with your business domain?
• TTL settings: Is each Feature View's TTL consistent with the data refresh cycle?
• Offline/online separation: Not all Feature Views need to be in the online store. Have you identified features that can be set to online=False?
• Registry: Are you using a SQL-based registry to prevent concurrent access conflicts?

Pipeline Configuration

• Materialization schedule: Is periodic Materialization configured via Airflow or Cron?
• Failure alerts: Are Slack/PagerDuty alerts configured for Materialization failures?
• Source data validation: Are you pre-validating source data quality with Great Expectations or similar tools?
• Incremental Materialization: Are you using materialize-incremental to avoid full reprocessing?

Monitoring

• Online store latency: Are you monitoring p50, p95, and p99 latency?
• Feature freshness: Are you tracking the latest Materialization timestamp for each Feature View?
• Feature drift: Is there monitoring to detect changes in feature distributions?
• Null rate: Are you tracking the null return rate for online feature queries?

Security and Governance

• RBAC: Is Feature View access separated by team?
• Audit logs: Are feature definition change histories recorded?
• PII masking: Is appropriate masking applied to features containing personal information?

Conclusion

A Feature Store is the core infrastructure that elevates an ML system to the next level of maturity. Feast, with its open-source flexibility and pluggable architecture, serves as an excellent starting point for most organizations. However, adopting a Feature Store should never become an end in itself. The focus must remain on clear business value: ensuring feature logic consistency, preventing Training-Serving Skew, and improving feature reusability.

By combining pipeline integration with Airflow or Kubeflow, Redis-based online serving, and SQL registry metadata management, you can build a production Feature Store capable of reliably operating dozens of Feature Views and millions of entities. Depending on your organization's scale and requirements, managed solutions like Tecton or Hopsworks are also worth evaluating.