Get hands-on experience with 20+ free Google Cloud products and $300 in free credit for new customers.

Vertex AI JupyterLab Extremely Slow Load Times on version M127

Issue Summary:
We are experiencing extremely slow load times when launching JupyterLab on version M127 in Vertex AI. The issue persists across multiple fresh instances and configurations.

Steps to Reproduce:

Spin up a new Vertex AI Workbench instance using M127.
Attempt to launch JupyterLab (via the instance page "OPEN JUPYTERLAB" button ).
Observe that JupyterLab takes a few minutes to 10+ minutes to load, compared to seconds on previous versions.
Try launching on an M126 instance—it loads immediately.

Follow the steps above to create a new instance, but uncheck the

 

 

Enable Dataproc Serverless Interactive Sessions
Enable access to Dataproc Spark kernels 

 

 

option OR stop instance and uncheck the option.

Launch speed returns to normal.

Expected Behavior:
JupyterLab should launch within seconds, as it does on M126.

Observed Behavior:

JupyterLab takes several minutes to load on M127.
Issue occurs on both Jupyter 3.0 and Jupyter 4.0.
The UI loading indicator differs between versions (Jupyter swirl in 4.0 vs. GCP logo in 3.0).
Disabling the Dataproc kernel significantly improves startup time, suggesting the delay is tied to Spark kernel initialization.

What We’ve Tried:

Spinning up multiple fresh instances with default settings.
Restarting local environments, clearing caches, and trying different browsers.
Reinstalling Jupyter, JupyterLab, Jupyter Server, and dependencies with the spark kernel still enabled.
Testing across multiple instances and configurations.

Key Finding:
The issue only occurs on M127 instances and appears linked to Dataproc spark kernel initialization. Disabling the Dataproc spark kernel allows Jupyter to start quickly.

Questions:

  • Is this a known issue with M127 instances?
  • Are there any workarounds or fixes beyond disabling the Dataproc kernel?
  • Can anyone confirm experiencing similar behavior?

 

0 1 133
1 REPLY 1

Hi @wrosko,

Welcome to Google Cloud Community!

Looks like you're dealing with some really slow load times for your JupyterLab instances in Vertex AI with the M127 version, linked to Dataproc Spark kernels. Disabling the Spark kernel makes things start up much faster.

Is this a known issue with M127 instances?

You may want to look on Google Cloud's Status Dashboard to see if there are any reported issues or incidents related to M127 instances. This could help you determine if your slow load times with JupyterLab are part of a broader problem. Additionally, you may contact Google Cloud Support to confirm if there is already a known issue and resolution.

Are there any workarounds or fixes beyond disabling the Dataproc kernel?

You may try the following workaround that could help resolve your issue:

  • Increase Resources for the Instance: Try increasing the instance type by adding more vCPUs or memory to address slow initialization, as insufficient resources might be the cause. Additionally, you may use a high-performance persistent disk like an SSD, and ensure the disk is large enough for temporary files or caching data required by the Dataproc kernel.
  • Configure your Network Settings: If you're using a custom VPC, make sure it's properly configured with firewall rules and routing settings that allow communication between your Vertex AI instance and the Dataproc service. Additionally, if you are using Private Google Access, verify that it is enabled for the subnet associated with your Vertex AI instance.

Was this helpful? If so, please accept this answer as “Solution”. If you need additional assistance, reply here within 2 business days and I’ll be happy to help.