LangFlow IDE custom workbench needs two ~pods~ (containers?): one for frontend, other as backend

[chaturvedi-kna]

Hi Guys,

I am looking to deploy LangFlow IDE as custom workbench and use it to connect to model running using Ollama runtime on ODH, but I am bit confused if that can be done or not as it has two pods one for frontend, other as backend are running as part of the deployment.

Related link

It would be great if someone can provide some guidance.

[jiridanek]

In Kubeflow Notebooks 1, you are allowed to specify more than one container in your pod, so this part is implemented. The problem is that when ODH Dashboard is used to spawn the workbench, it cannot do this for you and only one container is permitted there. This means that in ODH/RHOAI there is no way to spawn the workbench you want through Dashboard. You'd have to upload Notebook CR manually and users would need to take care not to try to edit it in the Dashboard.

For upcoming Kubeflow Notebooks 2, I think the situation is a little better than in Notebooks 1; even though I am not aware of any specific thinking around deploying multiple containers, the workaround I described above should be somewhat more robust in practice, with users not being able to break it for themselves once it's been set up.

If you'd like this for Red Hat OpenShift AI, then please fill a https://issues.redhat.com/projects/RHOAIRFE/issues ticket.

[chaturvedi-kna]

Hi @jiridanek Thanks for the reply.

I understand that two containers in one workbench is possible by applying CR but in this case of Langflow it has two pods architecture so is there any workaround for the same.

In case it is not possible as workbench if we go ahead and deploy it as an application on OCP, I want to confirm that you foresee any issue when Langflow has to communicate with model which is running on ODH (with respect to auth or something else).

[jiridanek]

For future reference, consider asking in the Kubeflow upstream; for example we had a discussion with @shalberd that lead to opening

• feature: allow custom pod labels in notebooks, pipeline components, serving pods kubeflow/community#800

because the concern was not specific to ODH, but it was a general Kubeflow concern. And so is running these fancy workbenches.

In case it is not possible as workbench if we go ahead and deploy it as an application on OCP

That sounds like a good way forward to me, yes. Does it have any negatives for you over being able to use it as a workbench?

[chaturvedi-kna]

Sorry for the delayed reply I was on a trip, I did proceed with OCP deployment.
Thought everything would be on a single page if used workbenches.

For future reference, consider asking in the Kubeflow upstream

Sure, I will keep that in mind from next time on..