Why Notebooks Become Inaccessible

The most common reasons are:

• Exceeding CPU or RAM limits: If your notebook uses more memory than allocated, the session can crash or be terminated

• Browser or network interruptions: Your notebook interface runs through your browser, so local issues (like Wi-Fi drops or browser crashes) can affect connectivity

• Inactive sessions: If a session is idle for too long, it may be auto-terminated depending on platform settings

Reconnecting to Your Notebook

If your session drops or the browser disconnects:

• Restart the session if needed: If the session was terminated, start a new one and reattach your storage volume to access saved files

• Reload your browser: In some cases, refreshing the tab is enough to reestablish the connection

• Use saved checkpoints or export history: If you enabled model checkpoints or regularly saved outputs to disk, you can easily resume work

Best Practices to Avoid Session Loss

1. Save Outputs to disk

Large plots or datasets held in memory can cause slowdowns or crashes. Save them to your attached volume with:

Example
HTML
Copy
plt.savefig('/workspace/output/my_plot.png')

2. Use Model Checkpoints

When training models, use built-in checkpointing (like PyTorch’s torch.save() or TensorFlow’s checkpoint manager) to periodically save progress.

This way, even if your session ends unexpectedly, you won’t have to start over.

3. Manage Memory Efficiently

Remove unused variables and clear memory during long sessions:

Example
HTML
Copy
del big_tensor
import gc
gc.collect()

4. Run Long Jobs via SSH

For long-running scripts or training loops, consider converting your notebook to a Python script and running it in the background using tools like screen or tmux via SSH:

Example
HTML
Copy
jupyter nbconvert --to script my_notebook.ipynb
ssh into your instance
screen
python my_notebook.py

This keeps your job running even if your browser disconnects.

Still Having Issues?

If you’re still seeing disconnects or crashes after following these tips, reach out to our support team at [email protected] — we’re here to help.