Online vs offline feature stores in MLOps - CLI Comparison

Introduction

Feature stores help manage data features used in machine learning. Online feature stores provide fast access to fresh data for real-time predictions. Offline feature stores store historical data for training and batch processing.

When you need to serve real-time predictions with the latest data in a web app or mobile app

When you want to train machine learning models using historical data stored in a data warehouse

When you want to keep feature data consistent between training and serving environments

When you want to reduce data engineering work by centralizing feature management

When you want to monitor feature data quality and freshness over time

Commands

This command creates an offline feature store using Parquet files stored at /data/offline_features. Offline stores hold historical feature data for training.

Terminal

mlflow feature-store create-offline-store --name example_offline_store --type parquet --path /data/offline_features

Expected Output

Created offline feature store 'example_offline_store' with type 'parquet' at path '/data/offline_features'

→

--name - Sets the name of the offline feature store

→

--type - Specifies the storage type for offline features

→

--path - Defines the storage location for offline features

This command creates an online feature store using Redis for fast, low-latency access to features during model serving.

Terminal

mlflow feature-store create-online-store --name example_online_store --type redis --host 127.0.0.1 --port 6379

Expected Output

Created online feature store 'example_online_store' with type 'redis' at 127.0.0.1:6379

→

--name - Sets the name of the online feature store

→

--type - Specifies the storage type for online features

→

--host - Defines the Redis server host

→

--port - Defines the Redis server port

This command ingests historical user feature data into the offline feature store from a Parquet file for training use.

Terminal

mlflow feature-store ingest --store example_offline_store --feature-group user_features --file user_features.parquet

Expected Output

Ingested 10000 records into feature group 'user_features' in offline store 'example_offline_store'

→

--store - Specifies which feature store to ingest data into

→

--feature-group - Defines the feature group name

→

--file - Specifies the data file to ingest

This command starts serving the user_features from the online feature store for real-time model predictions.

Terminal

mlflow feature-store serve --store example_online_store --feature-group user_features

Expected Output

Serving feature group 'user_features' from online store 'example_online_store' on port 8080

→

--store - Specifies which online feature store to serve from

→

--feature-group - Defines the feature group to serve

Key Concept

If you remember nothing else from this pattern, remember: offline feature stores hold historical data for training, while online feature stores provide fast access to fresh data for real-time predictions.

Common Mistakes

Using the online feature store for batch training data ingestion

Online stores are optimized for low-latency access, not large-scale batch storage, which can cause performance issues

Use the offline feature store to ingest and store historical data for training

Not keeping feature definitions consistent between online and offline stores

This causes training-serving skew, where models see different feature values during training and prediction

Define features once and use the same definitions in both stores

Serving features from the offline store in real-time applications

Offline stores have higher latency and may have stale data, causing slow or inaccurate predictions

Serve features from the online store for real-time prediction needs

Summary

Create offline feature stores to hold historical data for training machine learning models.

Create online feature stores to serve fresh features quickly for real-time predictions.

Ingest data into offline stores for batch processing and serve features from online stores during model serving.

Practice

1. What is the main purpose of an online feature store in MLOps?

easy

A. To backup model checkpoints

B. To store historical data for model training

C. To provide fast, real-time features for model predictions

D. To monitor model performance metrics

Solution

1. Step 1: Understand the role of online feature stores

   Online feature stores serve features quickly to models during prediction time, enabling real-time decisions.

2. Step 2: Differentiate from offline feature stores

   Offline feature stores hold historical data used for training, not for real-time serving.

3. Final Answer:

   To provide fast, real-time features for model predictions -> Option C

4. Quick Check:

   Online feature store = real-time features [OK]

Hint: Online = real-time data for predictions [OK]

Common Mistakes:

• Confusing online with offline feature stores
• Thinking online stores hold historical training data
• Mixing feature stores with model storage

2. Which of the following is a correct characteristic of an offline feature store?

easy

A. Stores historical feature data for model training

B. Automatically updates features during live inference

C. Provides low-latency access for real-time predictions

D. Is used to deploy models to production

Solution

1. Step 1: Identify offline feature store purpose

   Offline feature stores keep historical data used to train machine learning models.

2. Step 2: Eliminate incorrect options

   Low-latency and live inference updates are for online stores; deployment is unrelated.

3. Final Answer:

   Stores historical feature data for model training -> Option A

4. Quick Check:

   Offline feature store = historical training data [OK]

Hint: Offline = historical data for training [OK]

Common Mistakes:

• Confusing offline with online feature store roles
• Assuming offline stores serve real-time predictions
• Mixing feature storage with model deployment

3. Given this scenario: A model needs features for prediction within milliseconds. Which feature store query is correct?

medium

A. Query the offline feature store for batch data

B. Query the online feature store for real-time features

C. Query the model registry for feature values

D. Query the training dataset directly

Solution

1. Step 1: Identify the requirement for low latency

   Prediction within milliseconds requires fast access to features, which online stores provide.

2. Step 2: Match query to feature store type

   Online feature stores serve real-time features; offline stores and training data are too slow.

3. Final Answer:

   Query the online feature store for real-time features -> Option B

4. Quick Check:

   Real-time prediction needs online store [OK]

Hint: Real-time prediction = online store query [OK]

Common Mistakes:

• Using offline store for real-time prediction
• Confusing model registry with feature store
• Querying training data directly during prediction

4. You notice your model predictions are slow. You find the system queries the offline feature store during inference. What is the best fix?

medium

A. Switch queries to the online feature store for low latency

B. Increase the batch size in the offline store queries

C. Add more features to the offline store

D. Retrain the model with fewer features

Solution

1. Step 1: Identify cause of slow predictions

   Querying offline store during inference causes latency because it is not optimized for real-time access.

2. Step 2: Choose the fix for low latency

   Switching to the online feature store provides fast, real-time feature access, improving prediction speed.

3. Final Answer:

   Switch queries to the online feature store for low latency -> Option A

4. Quick Check:

   Slow predictions fixed by using online store [OK]

Hint: Use online store for inference speed [OK]

Common Mistakes:

• Trying to fix latency by changing batch size
• Adding features does not improve speed
• Retraining model unrelated to feature store latency

5. You want to ensure your ML system uses consistent features during training and prediction. How should you combine online and offline feature stores?

hard

A. Use only the online store for both training and prediction

B. Store features separately in each model without sharing

C. Use the offline store for serving features and the online store for training

D. Use the offline store for training data and the online store for serving features in production

Solution

1. Step 1: Understand consistency needs

   Consistent features mean training and prediction use the same data definitions and values.

2. Step 2: Apply best practice for feature stores

   Offline stores hold historical data for training; online stores serve features quickly during prediction.

3. Step 3: Combine stores correctly

   Use offline store for training datasets and online store for real-time serving to maintain consistency and performance.

4. Final Answer:

   Use the offline store for training data and the online store for serving features in production -> Option D

5. Quick Check:

   Offline for training + online for serving = consistency [OK]

Hint: Train offline, serve online for consistent features [OK]

Common Mistakes:

• Using only online store for training causes inconsistency
• Serving from offline store causes latency
• Not sharing feature definitions between stores