Forum Discussion
Some methods in notebooks do not work when executed from Data Pipelines deployed via SPN
Hi,
I already created the ticket with Microsoft (2504161420001430) and also opened the issue here and also seems to be related to this issue here.
`notebookutils.runtime.context.get("currentWorkspaceName")` works when executed directly from the Notebook, but does not work when executed from the DataPipeline which was deployed in another workspace using Service Principal via Fabric Core APIs (Create Item - https://learn.microsoft.com/en-us/rest/api/fabric/core/items/create-item?tabs=HTTP).
If I reopen the notebook with my personal account and rerun it again from the DataPipeline, then it works. This implies that the issue is coming from the service principal somehow losing permissions/token and is not able to read the current workspace name where it is running in from the notebookutils runtime context.
What I noticed is that this issue is also present for other built-in methods in Fabric. For example: `notebookutils.lakehouse.get()` and `synapsesql()`. Details can be found here: https://github.com/microsoft/fabric-cicd/issues/202#issuecomment-2797384465. Same approach is being followed in the fabric-cicd library and it is using SPN auth + Fabric Core Create Item APIs (as already mentioned and referenced above).
Did anyone else experience the same/similar issues?
Best regards,
Milos
38 Replies
You're absolutely right - this is a known issue when running notebooks in the context of a Service Principal, especially when using notebookutils.runtime.context or mssparkutils.env. I ran into the same thing recently and did a deep dive into how execution context really works in Fabric.
If you're interested, I wrote up my findings (including this bug and a workaround) in this blog post:
Who's Calling? Understanding Execution Context in Microsoft FabricYour workaround was so helpful thank you. I don't suppose you have a workaround for synapsesql() ?
I have noticed today under SPN ownership it no longer can run spark.sql and returns mwc token error. Whereas most of the other library methods with sempy, notebookutils is working even though it gives some cluster issues but the process continues to run successfully.
we are also having the same issue
Having same issue.
Hi mmilosanovic ,
Thanks for sharing the details and for raising a support ticket .
The issue likely stems from the notebook’s runtime context not being fully initialized when triggered via a Service Principal (SPN) using the Fabric Core APIs. This affects methods like notebookutils.runtime.context.get() and others that rely on workspace-level context.
In the meantime, please try the following steps:
-
Ensure the SPN has Contributor or higher role at the workspace level.
-
Verify that the workspace is explicitly included in the SPN’s access scope.
-
Test execution using a user-assigned managed identity if supported.
-
As a workaround, consider passing the required context (e.g., workspace name) as parameters to the notebook via the pipeline.
Please continue monitoring your support case with microsoft and share any updates or guidance as it becomes available, so that other community members who have similar problems to solve it faster.
Thank you.
-
we have the same issue with workspace identity as executing user. There are errors from cluster just with import statements in notebook and on applying run magic command. Even notebookutils.lakehouse.getWithProperties method also fails with 403 error. We were able to execute all of our code artefacts successfully with executing user as SPN (we generated our own SPN and added a secret to it) whereas when deploying same using workspace identity and running through it failed miserably. We even created a secret for the workspace identity and tried to use it by initializing credential class still no success.
Thank you for the detail in your response. You said you were able to run as SPN.... do you know if you were able to run these simple lines?
import jsonimport sempy.fabric as fabricfrom notebookutils import mssparkutils#Instantiate the clientclient = fabric.FabricRestClient()# Get WorkspaceworkspaceId = fabric.resolve_workspace_id([workspace name here])It works by explicitly creating an Service principal token implementation and passing that to fabric rest client. Here is a blog that can helped me with this implementation.
https://fabric.guru/using-service-principal-authentication-with-fabricrestclient
We have also raised a support request for the same issue
Hi mmilosanovic ,
Could you please confirm if the issue has been resolved after raising a support case? If a solution has been found, it would be greatly appreciated if you could share your insights with the community. This would be helpful for other members who may encounter similar issues.
Thank you for your understanding and assistance.
hi Anonymous , not yet resolved. Still in the back and forth with MS support. Last thing I got is this: "I have consulted with the notebook team, and they have confirmed that it is a known issue."
I have asked if there is a workaround and got the following feedback:
- "It has been noted that there are issues with executing notebooks via Service Principal (SPN) authentication, particularly with certain functions such as notebookutils. The internal team is actively investigating this matter to identify any limitations or configuration issues that may be causing these failures, and they will share any updates they have over the link.
- Also, I've informed Santhiya that by 20th May, Pipeline product team are releasing connection experience where users can create SPN connection via it."
So I guess we all monitor the progress here: https://github.com/microsoft/fabric-cicd/issues/248, and also keep an eye for the connection experience updates based on the seconds bullet above.
My team is also experiencing the same issue but when running notebooks called via API e.g. a making a request like this:
https://api.fabric.microsoft.com/v1/workspaces/<workspace_id>/items/<notebook_id>/jobs/instances?jobType=RunNotebook
from a service principal (in our case the managed identity of an ADF factory)
I imagine the Fabric Pipeline is doing the same API call under the hood.
As notebookutils.runtime.context.get("currentWorkspaceId", "") is working and we only have two workspaces at the moment (dev and prd) for a temporary workaround we have hardcoded environment specific values based on the workspace id into the custom package that we are using to manage ELT.
Still deciding whether or not we want to move these hardcoded environment specific values to a JSON file managed via CI/CD or swap back to notebookutils once the current issues are resolved. Would be nice to have less moving parts but also equally nice to have less reliance on MS managed bits and bobs that haven't proved completely reliable yet.
We have also observed the following log output when running a notebook via Service Princiapl that imports papermill (we have a development workflow where we can build out data modelling locally and run it via local spark installs before using CI/CD to push it up to Fabric, use papermill when running locally in place of notebookutils.notebook calls) this one is easier to tidy up just had to make sure we don't import papermill when running in a remote context but might help whoever is looking into this.
Failed to fetch cluster details Traceback (most recent call last): File "/home/trusted-service-user/cluster-env/clonedenv/lib/python3.11/site-packages/synapse/ml/fabric/service_discovery.py", line 110, in get_mlflow_shared_host raise Exception( Exception: Fetch cluster details returns 401:b'' Fetch cluster details returns 401:b'' Traceback (most recent call last): File "/home/trusted-service-user/cluster-env/clonedenv/lib/python3.11/site-packages/synapse/ml/fabric/service_discovery.py", line 152, in set_envs set_fabric_env_config(builder.fetch_fabric_client_param(with_tokens=False)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/trusted-service-user/cluster-env/clonedenv/lib/python3.11/site-packages/synapse/ml/fabric/service_discovery.py", line 72, in fetch_fabric_client_param shared_host = get_fabric_context().get("trident.aiskill.shared_host") or self.get_mlflow_shared_host(pbienv) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/trusted-service-user/cluster-env/clonedenv/lib/python3.11/site-packages/synapse/ml/fabric/service_discovery.py", line 110, in get_mlflow_shared_host raise Exception( Exception: Fetch cluster details returns 401:b'' ## Not In PBI Synapse Platform ##
