Adding Streaming Responses with Weights & Biases
We’re building a real-time model monitoring system that sends streaming responses using Weights & Biases. This matters because keeping an eye on your model’s behavior while it processes data is essential. No one wants to wake up one day to find their model’s accuracy has plummeted because they skipped over this step.
Prerequisites
- Python 3.11+
- Weights & Biases version 0.15.0+
- Flask for web framework:
pip install flask - Install Weights & Biases:
pip install wandb
Step 1: Setup Your Environment
# Create a virtual environment
python -m venv myenv
source myenv/bin/activate # On Windows, use `myenv\Scripts\activate`
pip install wandb flask
Why this matters: Creating a virtual environment keeps your dependencies isolated. It’s like building a bubble around your project — no one can mess with your packages, and you won’t end up with conflicts. Trust me, I had a project once where I accidentally broke everything with a simple upgrade. Mistakes happen, but we can avoid them!
Step 2: Initialize Weights & Biases
import wandb
# Initialize Weights & Biases
wandb.init(project="streaming-responses")
Why do we initialize Weights & Biases here? This sets up your project in W&B and starts logging metrics. You’ll want to see how your model performs in real time — it’s like having a window into your model’s soul. If you don’t set this up, you’ll just be left in the dark, staring at hard-to-read logs.
Step 3: Create a Simple Flask App
from flask import Flask, Response
import wandb
app = Flask(__name__)
@app.route('/stream')
def stream():
def generate():
for i in range(10):
wandb.log({"iteration": i, "accuracy": i * 10}) # Simulated metrics
yield f"data: Iteration {i}, Accuracy: {i * 10}\n\n"
return Response(generate(), mimetype='text/event-stream')
if __name__ == '__main__':
app.run(debug=True)
Why Flask? It’s lightweight and perfect for serving up data without the overhead of larger frameworks. In this app, we define a route `/stream` that sends updates. Every iteration, we log data to Weights & Biases and yield a streamed response. The data format is essential; otherwise, the client-side won’t understand it.
Step 4: Run Your Flask Application
# Run the Flask app
python app.py
When you run this command, your Flask application should start up. If you get a port error, it’s probably because you have another service running on that port. Either kill that service or change the port in the `app.run()` method.
Step 5: Setting Up the Frontend to Receive Streaming Responses
Streaming Responses
Streaming Responses from Weights & Biases
Why do we need a frontend? To display the data in real-time. Here, we use JavaScript’s `EventSource` object to establish a connection to your Flask app. It automatically handles connection drops and keeps trying to reconnect, which is a plus. Your model’s accuracy metrics will be displayed right on the webpage. If you forget to set the correct URL, you’ll get a 404, and no one likes that.
The Gotchas
- Network Latency: Streaming data can introduce delays. Make sure your server is optimized for performance; otherwise, you’ll be waiting ages for updates.
- Logging Overhead: Too many logs can bloat your storage. Set up a retention policy on W&B to manage this right from the start. Otherwise, you’ll end up with a hefty bill.
- CORS Issues: If you’re serving your Flask app on a different port than your frontend, you might face Cross-Origin Resource Sharing issues. Use Flask-CORS to fix this.
- Debugging: Debugging streamed responses can be tricky. Use Chrome’s Developer Tools to inspect network requests. Seriously, it saved me hours of head-scratching.
Full Code
import wandb
from flask import Flask, Response
# Initialize Weights & Biases
wandb.init(project="streaming-responses")
app = Flask(__name__)
@app.route('/stream')
def stream():
def generate():
for i in range(10):
wandb.log({"iteration": i, "accuracy": i * 10}) # Simulated metrics
yield f"data: Iteration {i}, Accuracy: {i * 10}\n\n"
return Response(generate(), mimetype='text/event-stream')
if __name__ == '__main__':
app.run(debug=True)
Here’s the full working example. This includes everything from initializing W&B to serving data via Flask. You can tweak the range or the data you log based on your needs.
What’s Next?
Consider implementing authentication for your Flask app. Anyone could access your endpoints right now, and that’s not great for security. Adding authentication layers will help secure your application from unauthorized access.
FAQ
- What if I can’t see any data in W&B?
Double-check your API key. Sometimes, you just need to log in viawandb.login(). - Can I use this for larger datasets?
Yes, but you might want to consider batching your logs to avoid overwhelming W&B with too many writes. - How do I stop the streaming?
Just close the connection from the client-side, or in a real-world scenario, you might want to implement a stop condition on the server.
Data Sources
- Weights & Biases Documentation
- AI Mind — How to Use Weights and Biases?
- MLOps: Weights & Biases Experiments Tracking
Last updated April 14, 2026. Data sourced from official docs and community benchmarks.
đź•’ Published: