Laptop vs Cloud

Step 2 of Prototyping: "Is my computer big enough?"

1. Vertical Scaling (The Notebook)

If your dataframe fits in RAM (e.g., 10GB), you process it in the notebook. If it is slightly larger (e.g., 100GB), you can:

1. Resize the VM: In User-Managed Notebooks, you can stop the VM, change machine type from n1-standard-4 to m1-ultramem-40, and restart.
2. Attach GPU: Add a T4 or V100.

2. Horizontal Scaling (The Executor)

The Executor feature in Workbench allows you to "Offload" a cell.

• You confirm your code works on a sample (100 rows).
• You click "Execute on Cluster."
• Managed Notebooks spins up a Dataproc cluster or a Training Job, runs the notebook on the full dataset, and saves the output.

Why? It keeps your development environment cheap (small VM) while allowing bursts of high power for heavy jobs.

3. Custom Containers

Sometimes pip install isn't enough. You need system libraries (e.g., libsound, ffmpeg).

1. Dockerfile: Write a Dockerfile extending the Deep Learning Container.
2. Build: gcloud builds submit.
3. Use: When creating the Notebook (or Training Job), specify "Custom Container Image."

Knowledge Check