This template demonstrates how to build a complete sentiment analysis pipeline using Feast (Feature Store) with PyTorch and Hugging Face Transformers. It showcases modern MLOps practices for NLP including feature engineering, model serving, and real-time inference.

• Feast Fundamentals: Feature stores, entities, feature views, and services
• NLP Feature Engineering: Text preprocessing and feature extraction patterns
• PyTorch Integration: Using pre-trained Hugging Face models with Feast
• Real-time Serving: Online feature serving for production inference
• MLOps Patterns: Model versioning, performance monitoring, and data governance

• Python 3.8+
• pip or conda for package management

feast init my-sentiment-project -t pytorch_nlp
cd my-sentiment-project

# Install Feast with NLP support (includes PyTorch, transformers, and ML utilities)
pip install feast[nlp]

cd feature_repo
feast apply
feast materialize-incremental $(date -u +"%Y-%m-%dT%H:%M:%S")

feast serve --host 0.0.0.0 --port 6566

python test_workflow.py

• 1000 synthetic text samples with sentiment labels (positive/negative/neutral)
• Engineered features: text length, word count, emoji count, etc.
• User context: aggregated user statistics and behavior patterns
• Dynamic timestamps generated within the past 30 days for realistic demo experience

• Text Features: Content, metadata, and linguistic characteristics
• User Features: Historical sentiment patterns and engagement metrics
• Real-time Features: On-demand sentiment prediction using pre-trained models

• Pre-trained Models: CardiffNLP Twitter-RoBERTa for sentiment analysis
• Embedding Generation: Text vectorization for similarity and clustering
• Confidence Scoring: Prediction confidence and probability distributions

Once you've started the feature server with feast serve, you can query features via HTTP API:

Query stored text and user features:

curl -X POST \
  "http://localhost:6566/get-online-features" \
  -H "Content-Type: application/json" \
  -d '{
    "features": [
      "text_features:text_content",
      "text_features:sentiment_label",
      "user_stats:user_avg_sentiment"
    ],
    "entities": {
      "text_id": ["text_0000", "text_0001"],
      "user_id": ["user_080", "user_091"]
    }
  }'

Example Response:

{
  "metadata": {"feature_names": ["text_id","user_id","sentiment_label","text_content","user_avg_sentiment"]},
  "results": [
    {"values": ["text_0000"], "statuses": ["PRESENT"]},
    {"values": ["user_080"], "statuses": ["PRESENT"]},
    {"values": ["positive"], "statuses": ["PRESENT"]},
    {"values": ["Having an amazing day at the beach with friends!"], "statuses": ["PRESENT"]},
    {"values": [0.905], "statuses": ["PRESENT"]}
  ]
}

Get real-time sentiment analysis:

curl -X POST \
  "http://localhost:6566/get-online-features" \
  -H "Content-Type: application/json" \
  -d '{
    "features": [
      "sentiment_prediction:predicted_sentiment",
      "sentiment_prediction:sentiment_confidence",
      "sentiment_prediction:positive_prob"
    ],
    "entities": {
      "input_text": ["I love this amazing product!", "This service is terrible"],
      "model_name": ["cardiffnlp/twitter-roberta-base-sentiment-latest", "cardiffnlp/twitter-roberta-base-sentiment-latest"]
    }
  }'

Query using predefined feature service:

curl -X POST \
  "http://localhost:6566/get-online-features" \
  -H "Content-Type: application/json" \
  -d '{
    "feature_service": "sentiment_analysis_v2",
    "entities": {
      "text_id": ["text_0000"],
      "user_id": ["user_080"],
      "input_text": ["This is an amazing experience!"],
      "model_name": ["cardiffnlp/twitter-roberta-base-sentiment-latest"]
    }
  }'

Note: Use actual entity combinations from your generated data. Run head data/sentiment_data.parquet to see available text_id and user_id values.

my-sentiment-project/
├── README.md                     # This file
└── feature_repo/
    ├── feature_store.yaml        # Feast configuration
    ├── example_repo.py           # Feature definitions
    ├── test_workflow.py          # Complete demo workflow
    └── data/                     # Generated sample data
        └── sentiment_data.parquet

• text: Unique identifier for text samples
• user: User who created the content

• text_features: Raw text content and engineered features
• user_stats: User-level aggregated statistics and behavior

• sentiment_prediction: Real-time sentiment analysis using PyTorch models
• Features: predicted sentiment, confidence scores, probability distributions, embeddings

• sentiment_analysis_v1: Basic sentiment features for simple models
• sentiment_analysis_v2: Advanced features with user context
• sentiment_training_features: Historical features for model training

This template is configured for local development using SQLite - no external dependencies required!

project: my_project
provider: local                    # Local provider (no cloud)
registry: data/registry.db         # SQLite registry
online_store:
  type: sqlite                     # SQLite online store (NOT Redis)
  path: data/online_store.db       # Local SQLite file
offline_store:
  type: file                       # Local file-based offline store

• ✅ Zero setup - Works immediately after feast init
• ✅ Self-contained - All data in local files
• ✅ No external services - No Redis/cloud required
• ✅ Perfect for demos - Easy to share and understand

from feast import FeatureStore

store = FeatureStore(repo_path=".")

# Get historical features for model training
from datetime import datetime
import pandas as pd

entity_df = pd.DataFrame({
    "text_id": ["text_0000", "text_0001", "text_0002"],
    "user_id": ["user_080", "user_091", "user_052"],  # Use actual generated user IDs
    "event_timestamp": [datetime.now(), datetime.now(), datetime.now()]  # Current timestamps
})

training_df = store.get_historical_features(
    entity_df=entity_df,
    features=[
        "text_features:text_content",
        "text_features:sentiment_label",
        "text_features:text_length",
        "user_stats:user_avg_sentiment",
    ],
).to_df()

print(f"Retrieved {len(training_df)} training samples")
print(training_df.head())

# Get features for online serving (use actual entity combinations)
entity_rows = [
    {"text_id": "text_0000", "user_id": "user_080"},
    {"text_id": "text_0001", "user_id": "user_091"}
]

online_features = store.get_online_features(
    features=store.get_feature_service("sentiment_analysis_v1"),
    entity_rows=entity_rows,
).to_dict()

print("Online features:", online_features)

# Real-time sentiment analysis
prediction_rows = [{
    "input_text": "I love this product!",
    "model_name": "cardiffnlp/twitter-roberta-base-sentiment-latest"
}]

predictions = store.get_online_features(
    features=[
        "sentiment_prediction:predicted_sentiment",
        "sentiment_prediction:sentiment_confidence",
    ],
    entity_rows=prediction_rows,
).to_dict()

Here's a step-by-step walkthrough of the entire template workflow:

# Create new project
feast init my-sentiment-demo -t pytorch_nlp
cd my-sentiment-demo

# Install dependencies
pip install torch>=2.0.0 transformers>=4.30.0

# Navigate to feature repository
cd feature_repo

# Register entities, feature views, and services
feast apply

Expected Output:

Created entity text
Created entity user
Created feature view text_features
Created feature view user_stats
Created on demand feature view sentiment_prediction
Created feature service sentiment_analysis_v1
Created feature service sentiment_analysis_v2

# Load features into online store
feast materialize-incremental $(date -u +"%Y-%m-%dT%H:%M:%S")

Expected Output:

Materializing 2 feature views to 2025-XX-XX XX:XX:XX+00:00 into the sqlite online store.
text_features: ████████████████████████████████████████
user_stats: ████████████████████████████████████████

# Start HTTP feature server
feast serve --host 0.0.0.0 --port 6566

Expected Output:

Starting gunicorn 23.0.0
Listening at: http://0.0.0.0:6566

In a new terminal, test the feature server:

# Check actual entity IDs in your data
python -c "
import pandas as pd
df = pd.read_parquet('data/sentiment_data.parquet')
print('Sample entities:', df.head())
"

# Test with actual entity combinations
curl -X POST \
  "http://localhost:6566/get-online-features" \
  -H "Content-Type: application/json" \
  -d '{
    "features": ["text_features:text_content", "text_features:sentiment_label"],
    "entities": {
      "text_id": ["text_0000"],
      "user_id": ["user_XXX"]
    }
  }' | jq

# In example_repo.py, add to text_features_fv schema:
Field(name="hashtag_count", dtype=Int64, description="Number of hashtags"),
Field(name="mention_count", dtype=Int64, description="Number of @mentions"),
Field(name="url_count", dtype=Int64, description="Number of URLs"),

# In the sentiment_prediction function, change model:
model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
# or
model_name = "distilbert-base-uncased-finetuned-sst-2-english"

@on_demand_feature_view(
    sources=[text_input_request],
    schema=[Field(name="toxicity_score", dtype=Float32)],
)
def toxicity_detection(inputs: pd.DataFrame) -> pd.DataFrame:
    # Implement toxicity detection logic
    pass

1. Cloud Deployment: Use AWS, GCP, or Azure providers instead of local
2. Vector Store: Replace SQLite with Milvus for similarity search
3. Model Serving: Deploy models with KServe or other serving framework
4. Monitoring: Add feature drift detection and model performance tracking

Current Architecture:

• Models load on each request (see sentiment_prediction function)
• CPU-only operation to avoid multiprocessing issues
• SQLite-based storage for fast local access

TODO: Optimization Opportunities:

• Startup-time Model Loading: Load models once at server startup instead of per-request
• Custom Provider: Implement model caching via custom Feast provider
• Model Serving Layer: Use dedicated model servers (TorchServe, MLflow) for heavy models

Production Optimizations:

1. Model Caching: Cache loaded models in memory to avoid repeated loading
2. Batch Inference: Process multiple texts together for efficiency
3. Feature Materialization: Pre-compute expensive features offline
4. Async Processing: Use async patterns for real-time serving

Note: The demo uses SQLite (above). These are examples for production deployment:

# feature_store.yaml for AWS production (requires Redis setup)
project: sentiment_analysis_prod
provider: aws
registry: s3://my-bucket/feast/registry.pb
online_store:
  type: redis                      # Requires separate Redis server
  connection_string: redis://my-redis-cluster:6379
offline_store:
  type: bigquery
  project_id: my-gcp-project

# feature_store.yaml for GCP production (requires cloud services)
project: sentiment_analysis_prod
provider: gcp
registry: gs://my-bucket/feast/registry.pb
online_store:
  type: redis                      # Requires separate Redis server
  connection_string: redis://my-redis-cluster:6379
offline_store:
  type: bigquery
  project_id: my-gcp-project

This template is designed to be extended and customized:

1. Add new feature transformations in example_repo.py
2. Experiment with different models in the sentiment_prediction function
3. Extend the test workflow with additional evaluation metrics
4. Add new data sources (Twitter API, product reviews, etc.)

• Feast Documentation
• Hugging Face Transformers
• PyTorch Documentation

ImportError: No module named 'transformers'

pip install torch transformers

Model download timeout

# Set environment variable for Hugging Face cache
export HF_HOME=/path/to/cache

Feature store initialization fails

# Reset the feature store
feast teardown
feast apply

On-demand features return defaults

• This is expected if PyTorch/transformers aren't installed
• The template includes fallback dummy predictions for demonstration

• Check the Feast GitHub Issues
• Join the Feast Slack Community
• Review the PyTorch Forums

Happy Feature Engineering! 🎉

Built with ❤️ using Feast, PyTorch, and Hugging Face.