'Server Connection Error' on GCP (AI Platform Notebook)

Question

I am facing some issues with GCP and the AI Platform (Jupyterlab) It seems that I am unable to maintain a stable connection with the server for a long time. I keep getting those 'server connection error' message. From there two possibilities:

• either nothing happens and my cell keeps running or
• the cells have stopped running and I can see the status 'No Kernel!' on the top right of the notebook. Whenever I select a kernel (python 3) again, depending on my luck I can either keep working, or the cell will display the running status (with the * on the left of it) but the kernel status on the bottom left will stay on : 'connected' (instead of 'busy'). For the latter, I need to restart the kernel and run all the cells again, which can be very long.

Sometimes this happens as soon as I run the first cell after (re)starting the instance, sometimes a bit later. The longest stable period I was able to work on the notebook without any issue was 20, 30-ish minutes, which is quite annoying.

Configuration of my main instance : - 16 CPUs - 60gb of RAM - A P100 NVIDIA GPU

I have tried different types of instance and I keep having the same problem, network at home is stable.

error message

Longheaded

were you able to resolve the issue? – Befool

Answer 1

What operating system and browser are you using at work? I had the same problem as you did on Ubuntu 18 with the Firefox browser. When I switched to Windows with Chrome the error did not reoccur, even though it was the same network.

Deserted

Switching from Firefox to Chrome on Mac eliminated 'Server Connection Errors' for me as well. – Ljoka

I frequently receive this error when using Windows with Chrome, and with an instance whose disk has plenty of room. @gogasca Any other insights? – Frayne

Answer 2

I had a similar issue today: according to the google docs the cause for this is that the docker/ Jupyter service is not starting.

The cause why these services couldn't be started in our specific case was a full disk.

Abuzz