Publishing pipeline endpoints with SDK V2

Anirban Ray 6

Hi,

I created a basic two step pipeline with v2 SDK. It takes two inputs, first and last name. The first step prints Hello {first name}, and the second step prints Bye {last name}.

I ran this pipeline successfully giving my name as inputs, and then clicked on publish pipeline. That was successful, but I notice that now my name is being considered as a default value of the inputs. I do not want that, and I want the inputs to be mandatory user inputs.

In other words, I want to know how can I publish the pipeline without any default values, and not a particular run of the pipeline?

To clarify why I do not want default values as inputs of one run, this is a dummy pipeline. But the real pipelines can take many things, possibly sensitive, as inputs, and I do not want those to be available to everyone as default value.

Also, a related question, is there a way to programmatically publish a pipeline endpoint?

Thanks.

romungi-MSFT 44,661 Reputation points Microsoft Employee

2023-01-11T12:18:35.4166667+00:00

@Anirban Ray Is it possible to add details of the dummy pipeline to understand why the same inputs are taken in the new pipeline? I believe you might have not set is_deterministic to false in the component, by default it is true. I believe this is similar to allow_reuse from v1 SDK. Please see this page for details.

Anirban Ray 6

Hi @romungi-MSFT,

I feel may be you misinterpreted my question. For your reference, here's the sample pipeline:

import pathlib

import azure.ai.ml as az_ml
import azure.ai.ml.dsl as az_ml_dsl
import azure.ai.ml.entities as az_ml_ent
import azure.identity as az_id


subscription_id = "<subscription_id>"
resource_group_name = "<resource_group_name>"
workspace_name = "workspace_name"

instance_name = "<instance_name>"

try:
    credential = az_id.DefaultAzureCredential()
    credential.get_token("https://management.azure.com/.default")
except Exception:
    credential = az_id.InteractiveBrowserCredential()

try:
    ml_client = az_ml.MLClient.from_config(credential=credential)
except Exception:
    ml_client = az_ml.MLClient(
        credential,
        subscription_id=subscription_id,
        resource_group_name=resource_group_name,
        workspace_name=workspace_name,
    )

greet_user_step = az_ml.command(
    name="greet_user_step_name",
    description="greet user step name",
    display_name="greet user step display name",
    command='echo "hello ${{inputs.user_name}}"',
    environment=az_ml_ent.Environment(
        name="ubuntu_environment_name",
        version="ubuntu environment version",
        description="ubuntu environment description",
        image="library/ubuntu:latest",
    ),
    inputs={"user_name": az_ml.Input(type="string", optional=False)},
    instance_count=1,
    shm_size="1g",
    timeout=300,
    is_deterministic=False,
)

@az_ml_dsl.pipeline(
    name="greet_user_step_name",
    version="greet user step version",
    display_name="greet user step display name",
    description="greet user step description",
)
def greet_user_pipeline(name_of_user: str) -> None:
    _ = greet_user_step(user_name=name_of_user)

pipeline_job = greet_user_pipeline("John Doe")
pipeline_job.settings.default_compute = instance_name

submitted_run = ml_client.create_or_update(pipeline_job)

After this ran successfully, if I click Publish, then I get this.
Screenshot 2023-01-12 at 7.42.00 PM

As you can see, it stores the input for the sample job as default value ( John Doe ), and it's no longer a necessary argument as it was designed earlier. Also can be verified from Pipeline endpoints page:

Screenshot 2023-01-12 at 7.43.57 PM

Let me know if the problem is clear now. This setting of default value was my first question, and how to publish programmatically instead of clicking UI buttons was the second question.

Thanks.

romungi-MSFT 44,661 Reputation points Microsoft Employee

2023-01-13T10:38:42.0533333+00:00
I think you need to submit the pipeline job instead of the pipeline. I think from the above screen shot you are publishing a pipeline endpoint which takes the input as used in the ml_client.create_or_update()

Try:

ml_client.jobs.create_or_update(pipeline_job)
Priyanka Choudhary 15 Reputation points

2023-01-13T13:03:58.68+00:00
Hi @romungi-MSFT

We want to have published pipelines with no default value for pipeline parameters.

Let's say we create a pipeline that has database credentials as one of the pipeline parameters. Now we would not want to publish those database credentials as John Doe was published as the default value for name_of_user in the example shared by @Anirban Ray .

So we have two questions:

Can we publish pipelines without any default values for pipeline parameters?

Can we publish pipelines programmatically from v2 sdk instead of the portal?

Our end goal is to resubmit the published pipeline using pipeline endpoints and allow user to pass different values for the pipeline parameters when they resubmit. e.g. user may want to use a different database to pull their data from, so they have the option to provide db credentials as pipeline parameters.

The solution you provided ml_client.jobs.create_or_update(pipeline_job) will create a new pipeline every time. We want to reuse an already published pipeline.

Let me know if I was able to provide some clarity on our issue.
Patricia Inyang (Student) 5 Reputation points

2023-12-02T06:56:11.14+00:00

Please was an answer ever provided to this?

We want to be able to register a pipeline without actually submitting it as a job.

We also want to be able to submit a pipeline job without having to create the pipeline in the future.
Obaid UrRehman 166 Reputation points

2024-03-13T10:49:40.14+00:00

Hi, facing the same problem. Is there solution to it?

We have script that submits the pipeline, and now in CI we just want to deploy that pipeline instead of resubmiting it. (Like we used to do with SDKv1) Is there a way to do this?

Share via

Publishing pipeline endpoints with SDK V2

Your answer