YouTube Sentiment Analysis MLOps Project

Minimal backend + Chrome extension to analyze YouTube comments with a HuggingFace model.

Features

• Fetch top-level YouTube comments via YouTube Data API v3
• Sentiment analysis (multilingual model)
• FastAPI endpoint /analyze
• Chrome extension popup to trigger analysis and show counts
• MLflow for model tracking

Setup

1. Create .env from the template:

   cp .env.template .env

2. Fill in your API key in .env:

   YOUTUBE_API_KEY=...
   YOUTUBE_VIDEO_ID=...   # optional for tests
   

3. Install dependencies:

   uv sync

Run the API

uv run uvicorn --app-dir src youtube_sentiment.main:app --reload --port 8001

Test the API

curl -X POST http://127.0.0.1:8001/analyze \
  -H "Content-Type: application/json" \
  -d '{"video_id":"VIDEO_ID","max_comments":50}'

Run tests

uv run pytest

Integration tests (require .env values):

RUN_INTEGRATION_TESTS=1 uv run pytest -m integration

Chrome Extension

1. Go to chrome://extensions
2. Enable Developer mode
3. Click Load unpacked and select this repo folder
4. Open a YouTube watch page
5. Click the extension icon → Analyze Comments

Notes

• The model cache is stored in .hf-cache/ (override with HF_CACHE_DIR).

DVC (Data Versioning)

• DVC is used to track datasets, comment dumps, and model artifacts without putting large files in Git.
• This project is initialized for DVC, but no pipelines are defined yet (we will add them later).
• DVC remote is currently a local placeholder; swap to s3://... or gs://... when you move storage to S3 or GCS.

MLflow Model Registry (Local)

1. Start MLflow server:

   mlflow server \
     --backend-store-uri sqlite:///mlflow.db \
     --default-artifact-root ./mlruns \
     --host 127.0.0.1 \
     --port 5000

2. Register the pretrained model:

   uv run python scripts/register_model.py

CI/CD Secrets

• MLFLOW_TRACKING_URI (GitHub Actions secret): MLflow tracking server URL used by CI workflows.
• KUBECONFIG (GitHub Actions secret): kubeconfig contents for deploy workflow.

Deployment (Kubernetes)

1. Build & push happens on main via CI (GHCR).
2. Deploy manually via GitHub Actions → Deploy Backend.
   • Requires KUBECONFIG secret.
   • Deploys only if a Production model exists in MLflow.
   • Optionally provide image_tag to deploy a specific SHA.

Local apply (optional):

kubectl apply -f k8s/configmap-dev.yaml
kubectl apply -f k8s/secret.yaml
kubectl apply -f k8s/deployment.yaml
kubectl apply -f k8s/service.yaml

Architecture

• Doc: docs/architecture.md
• Link: Architecture Doc