The Development Environment

The first question in any ML project is: "Where do I write the code?" For Google Cloud, the answer is Vertex AI Workbench.

It comes in two flavors, and the exam expects you to know exactly when to use each.

---

1. Managed Notebooks (The Modern Way)

Keyword: "JupyterLab-as-a-Service".

• Architecture: It runs on a Google-managed container system.
• Pros:
  • Zero generic Ops: No OS patches, no SSH needed.
  • Integrations: Built-in widget for BigQuery, GCS, and Experiments.
  • Kernel Interconnect: You can execute code on a separate Dataproc cluster directly from the notebook cell.
• Cons: Limited customization of the underlying OS (because you aren't supposed to touch the OS).

Use Case: 90% of Data Science work. "I just want to write Python."

---

2. User-Managed Notebooks (The Legacy/Power Way)

Keyword: "Deep Learning VM (DLVM)".

• Architecture: It is literally a Compute Engine VM with Jupyter installed.
• Pros:
  • Full Control: You get sudo root access. You can install custom NVIDIA drivers, Docker containers, or weird C++ libraries.
  • Networking: You can put it inside a specific VPC subnet with custom firewall rules.
• Cons: You manage the VM. You pay for it even if idle (unless you set up idle shutdown).

Use Case: "I need a specific version of CUDA 10.1 that isn't supported by Managed Notebooks anymore."

---

3. Best Practices

1. Idle Shutdown: Always enable this. If the notebook isn't used for 3 hours, shut down the VM to save money.
2. Hardware Accelerator: You can attach 1 T4 GPU to the notebook for prototyping. Don't attach 8 GPUs (too expensive). Use a Training Job for that.

---

Knowledge Check