Hey there, I totally get how frustrating this is—restarting a multi-hour training run three times and watching your compute units tick down is the worst. Let’s dive into why those old scripts aren’t working anymore and share practical, up-to-date fixes for 2024 to keep your Colab session stable for long training jobs.

Why your old scripts failed #

Colab regularly updates its UI structure, so the DOM selectors your scripts used (like #top-toolbar > colab-connect-button) don’t point to the right elements anymore. Those clicks were essentially hitting empty air, which is why they didn’t keep your session alive.

Reliable fixes to try #

• Use an updated auto-keep-alive script (2024-compatible)
  Colab’s current UI has different element identifiers, so here’s a revised script that targets the right connect button and prevents tab dormancy:

  function keepColabAlive() {
    console.log("Keeping Colab session active...");
    // Target the current connect button element
    const connectButton = document.querySelector('colab-connect-button[aria-label="Connect"]');
    if (connectButton) {
      connectButton.click();
    }
    // Simulate a harmless page interaction to stop the browser from sleeping the tab
    document.body.dispatchEvent(new MouseEvent('mousemove'));
  }
  // Run every 5 minutes (300000 ms) to avoid triggering anti-abuse measures
  setInterval(keepColabAlive, 300000);
  

  To use it: Open your Colab notebook, press F12 to open Developer Tools, switch to the Console tab, paste the code, and hit Enter. Keep the browser tab open and active (don’t minimize it—many browsers throttle background tabs).

• Upgrade to Colab Pro/Pro+ if budget allows
  Free Colab sessions have hard time limits (usually ~6-8 hours) and are more likely to be disconnected to free up resources for other users. Colab Pro/Pro+ offers:

  • Longer maximum session durations (up to 24 hours for Pro+)
  • Higher priority access to resources, reducing unexpected disconnections
  • More compute units per month, so you won’t run out mid-training

• Optimize your training workflow to handle disconnections
  Even with the best keep-alive tricks, free sessions might still hit time limits. Protect your progress with these steps:

  • Save checkpoints frequently: Use model.save() or your framework’s checkpointing feature every 1-2 hours, and store them in Google Drive so you can resume training from the last checkpoint instead of starting over.
  • Reduce session load: Use smaller batch sizes if possible, cache datasets locally in Colab to cut down on IO delays, and close any unused tabs or apps on your computer to keep resources focused.

• Prevent browser and system dormancy
  Your computer or browser might put the tab to sleep if it’s inactive, which can trigger a Colab disconnection:

  • Disable sleep mode on your computer: Go to your system’s power settings and set it to "Never" sleep while plugged in.
  • Disable browser tab throttling: For Chrome, you can use extensions like "Tab Wrangler" (configured to not close your Colab tab) or adjust experimental settings to prevent background tab suspension.

备注：内容来源于stack exchange，提问作者Subaru Natsuki