Hello,
I'm writing as a Colab Pro+ subscriber to report what looks like a recent regression in runtime stability.
One of the main reasons I subscribed to Pro+ is the advertised ability to keep notebooks running for up to 24 hours continuously, including in the background after closing the browser. Over the past 3–4 days, however, my runtimes have been consistently terminating after only 40 mins 1 hours of execution.
These disconnections happen entirely on their own — there is no user interaction involved, and the notebook is still actively running when the session is killed. The runtime doesn't just disconnect; it gets fully torn down, meaning all in-memory state is lost.
On top of the disruption itself, this issue has already cost me approximately 200 compute units that were burned on training runs which never had a chance to complete because the session was terminated mid-execution. My work has been significantly slowed down as a result, since I now have to restart jobs repeatedly and cannot trust the runtime to stay alive long enough to finish anything meaningful.
This is simply not acceptable. The entire reason I pay for Pro+ — over the free tier or standard Pro — is precisely to be able to run long training jobs reliably. If sessions can no longer stay alive for extended periods, the core value of the Pro+ subscription is broken, and I'm effectively paying a premium for a service that isn't delivering what it advertises.
Hello,
I'm writing as a Colab Pro+ subscriber to report what looks like a recent regression in runtime stability.
One of the main reasons I subscribed to Pro+ is the advertised ability to keep notebooks running for up to 24 hours continuously, including in the background after closing the browser. Over the past 3–4 days, however, my runtimes have been consistently terminating after only 40 mins 1 hours of execution.
These disconnections happen entirely on their own — there is no user interaction involved, and the notebook is still actively running when the session is killed. The runtime doesn't just disconnect; it gets fully torn down, meaning all in-memory state is lost.
On top of the disruption itself, this issue has already cost me approximately 200 compute units that were burned on training runs which never had a chance to complete because the session was terminated mid-execution. My work has been significantly slowed down as a result, since I now have to restart jobs repeatedly and cannot trust the runtime to stay alive long enough to finish anything meaningful.
This is simply not acceptable. The entire reason I pay for Pro+ — over the free tier or standard Pro — is precisely to be able to run long training jobs reliably. If sessions can no longer stay alive for extended periods, the core value of the Pro+ subscription is broken, and I'm effectively paying a premium for a service that isn't delivering what it advertises.