Not able to use Elyra Pipeline Editor Extension with the "HybridContentsManager"

[altruistcoder]

I am working on a use case where I am running Jupyter as a pod in OpenShift, and I want to leverage several JupyterLab extensions such as Elyra.

Currently, I am trying to execute a sample pipeline provided in Elyra examples repo. But, I am facing the below error when I tried to run it in the JupyterLab notebook images having S3 buckets mounted in them (Using same HybridContentsManager):

But, I am able to run the pipeline successfully on JupyterLab notebook images which do not have S3 buckets mounted in them. As far as I am able to figure out with this issue, when I am trying to run the pipeline in an environment where HybridContentsManager is being used, the extension is trying to find files in the /examples/pipelines/ folder which gives the error as shown in the above screenshot, while it tries to find files in the /opt/app-root/src/examples/pipelines/ folder in the other case.

I also faced another similar issue with the jupyterlab-git extension and I got a solution for it which can be referenced here.

So, can you please help me in resolving this issue as I am not able to use the elyra's pipeline extension anymore.

[altruistcoder]

Hello,

I just wanted to check if there's any update on this issue yet?

@akchinSTC @ptitzler @lresende

[kevin-bates]

Hi @altruistcoder - first, I apologize for the lack of response.

I'd like to get some more information that I'm hoping you can provide:

1. Do you have or can you produce the Jupyter console log output, preferably when --debug option is used to start jupyter lab?
2. Can you share your configuration settings? In particular, those that configure the HybridContentsManager, but also, anything related to the notebook_dir or root_dir configurable traits.
3. What is the most recent version of Elyra you're using to reproduce this?

[altruistcoder]

Hello @kevin-bates, please pardon me for replying late.

Here are the answers to your queries:

1. Actually, my JupyterHub is running as a part of the ODH project in the form of pods on my openshift cluster, so I won't be able to add the debug flag.
2. You can find the configuration file for the Jupyter notebook and the jupyter server which we are using here.
3. The latest version of Elyra on which I have tested this issue is version 3.2.2.
   As you can see in this error message it is trying to access the notebook file at the path: /Elyra-Pipelines/examples/pipelines/introduction-to-generic-pipelines/load_data.ipynb which ideally should have been /opt/app-root/src/Elyra-Pipelines/examples/pipelines/introduction-to-generic-pipelines/load_data.ipynb.

So, basically it always try to search for the files in the / directory instead of the actual path. The same issue was happening with the jupyterlab-git extension as well and they needed to make some modifications in their code base to bring the support of HybridContentsManager which can be found here.

[kevin-bates]

Hi @altruistcoder - can you tell me whether or not the path value displayed in the validation error message is the correct path, but it should be prefixed by the root_dir you reference in your ContentsManager's configuration?

Unfortunately, the solution you link to points at a 45-file refactoring of the repo, and I have no idea what portion of that applies to this root-dir issue.

One thought is that we can't rely on os.path for node resources and need to go through the configured ContentsManager. If that's the case, which I suspect it is since s3-hosted files won't be found using os.path, then this is a widespread change that even affects BYO runtimes. Based on this information, I suspect the path displayed in the error message is perfectly correct but just doesn't exist in the local filesystem.

cc: @akchinSTC @lresende

[lresende]

Hi @altruistcoder - can you tell me whether or not the path value displayed in the validation error message is the correct path, but it should be prefixed by the root_dir you reference in your ContentsManager's configuration?

Without actually validating this, I believe the issue is that the backend is accessing files directly instead of using a content service and while this works on a local filesystem, it won't work on a hybrid content manager when the file is only available remotely.

There are probably two issues that we would have to handle here, one is the support for pipeline portability with relative paths and then patching the backend to use content manager to access files.

NOT A CONTRIBUTION

[kevin-bates]

I have confirmed this to be an issue by configuring the s3contents Content Manager. I confirmed general operation by pointing to an existing bucket and adding the following to my Jupyter configuration (from the referenced repo's README):

from s3contents import S3ContentsManager

# Tell Jupyter to use S3ContentsManager for all storage.
c.ServerApp.contents_manager_class = S3ContentsManager
c.S3ContentsManager.access_key_id = "my-cos-user"
c.S3ContentsManager.secret_access_key = "my-cos-password"
c.S3ContentsManager.endpoint_url = "http://<my-cos-server>:<cos-port>"
c.S3ContentsManager.bucket = "kbates"
c.S3ContentsManager.prefix = "notebooks/test"

Upon startup of jupyter lab from my home notebooks directory (typically my root-dir), I see the launcher but get a "Directory not found: ''" dialog box that I dismiss. The server log also contains:

[W 2021-11-18 15:33:29.905 ServerApp] 404 GET /api/contents/Users/kbates/notebooks?content=1&1637278409508 (::1): No such entity: [kbates/notebooks/test/Users/kbates/notebooks]

Since my cwd is /Users/kbates/notebooks, the s3contents is looking for the root-dir relative to the configured S3ConentsManager.prefix within bucket kbates (e.g., kbates/notebooks/test/Users/kbates/notebooks), which doesn't exist. This issue can be resolved by starting jupyter with a notebook-dir of / (e.g., jupyter lab --notebook-dir=/) which then treats kbates/notebooks/test within s3 as the root-dir.

I can create notebooks and execute cells as expected.

I also uploaded a working pipeline, its corresponding notebooks, and dependencies into s3 - mirroring my filesystem structure.

Issues

We have several issues in which submission is immediately blocked by validation - so I suspect many more to follow. Here's an enumeration so far using a generic pipeline (I'm using a numbered list for easier reference and priorities should not be inferred):

1. Drag and drop doesn't work (nothing happens). The server log contains the following, indicating that the components/properties endpoint needs to use the ContentsManager - which makes sense.

   [W 2021-11-18 15:45:47.989 ServerApp] 404 GET /elyra/contents/properties/elyra/my_pipeline/node3.ipynb?1637279147983 (::1): No such file or directory: elyra/my_pipeline/node3.ipynb

2. Right-clicking on a notebook or script and selecting 'Add file to pipeline' also does nothing and produces the same log entry.
3. (no issue) Using the component palette, I can drag an empty notebook node onto the canvas, then open its properties to set the name and runtime image. The browse button to associate the notebook file works fine, as does the runtime-image selection since metadata instances are in the local filesystem.
4. When attempting to run the pipeline (doesn't matter if local or against a platform), validation fails (as in this opening issue) - which also makes sense.

To get past the last issue, we'll need to either address the issue or disable that level of validation to determine what other issues exist. I'm also concerned there may be issues with Papermill since we pass the filepath of the notebook file. I'm not sure if there are other alternatives (like passing the JSON directly). That same kind of issue exists with script execution as well. There are also probable issues related to dependency management.

I suspect platform-specific components may work better.

Needless to say, this is probably a fairly long road, but one we should definitely explore further.