iterative / blog-tpi-jupyter

Terraform Provider Iterative + Jupyter + TensorBoard + AWS/Azure/GCP/K8s

Terraform Provider Iterative + Jupyter + TensorBoard

Repository supporting blog: short version on dev.to & full version on hackernoon.com

Painlessly deploy an ML-ready Jupyter server and sync results with your preferred cloud compute provider.

To get started, clone this repo, then in the repo directory run:

export NGROK_TOKEN="..."     # Sign up for free at https://ngrok.com
export TF_LOG_PROVIDER=INFO  # (optional) Control verbosity
terraform init               # Setup local dependencies
terraform apply              # Create cloud resources & upload "shared" workdir
terraform refresh            # Get Jupyter & TensorBoard URLs
# ...
# Have fun!
# (optional) Click "Quit" in Jupyter to shutdown the cloud machine
# ...
terraform destroy # Download "shared" workdir & terminate cloud resources

🛈 Note that it can take a couple of minutes after apply for the machine to be ready.

🛈 Note that Terraform Provider Iterative (TPI) will automatically restart interrupted spot/preemptible instances (including restoring Jupyter's working directory). In such cases, run terraform refresh again to obtain the new Jupyter & TensorBoard URLs.

Requirements

Toggle Version:

• Download terraform (free)
• ngrok credentials (free)
• Cloud credentials (AWS, Azure, GCP, or Kubernetes). For example:
  • AWS_ACCESS_KEY_ID
  • AWS_SECRET_ACCESS_KEY

Alternatives

Some that we are aware of: (Anything missing? Please do open a PR!)

• Google Colab: based on (but not identical to) Jupyter, has resource limits and limited config options (CPU, GPU, RAM, memory, timeouts)
• Binder: no GPU, no config options, for-profit use disallowed, has user guidelines

However there are a few distinct advantages to using terraform over the alternatives:

• Lower cost: use your preferred cloud provider's existing pricing, including on-demand per-second billing and bulk discounts
• Auto-recovery: auto-backup workdir & auto-recover terminated spot instances
• Custom spec: full control over hardware & software requirements via main.tf -- including machine types (CPU, GPU, RAM, storage) & images