jambonne
commented
2 months ago Adding to this; it seems like almost none of the google-cloud-pipeline-components are usable with DockerRunner. From what I can tell, its because the Dockerfile that makes the image used by all gcpc components here executes the aiplatform.remote_runner with an ENTRYPOINT command which can't be overridden by the command submitted by the Docker runner. This causes the command submitted to the container (something like the below:
['python3', '-u', '-m', 'google_cloud_pipeline_components.container.v1.bigquery.query_job.launcher', '--type', 'BigqueryQueryJob', '--project', "{{$.inputs.parameters['project']}}", '--location', "{{$.inputs.parameters['location']}}", '--payload', '{"Concat": ["{", "\\"configuration\\": {", "\\"query\\": ", "{{$.inputs.parameters[\'job_configuration_query\']}}", ", \\"labels\\": ", "{{$.inputs.parameters[\'labels\']}}", "}", "}"]}', '--job_configuration_query_override', '{"Concat": ["{", "\\"query\\": \\"", "{{$.inputs.parameters[\'query\']}}", "\\"", ", \\"query_parameters\\": ", "{{$.inputs.parameters[\'query_parameters\']}}", ", \\"destination_encryption_configuration\\": {", "\\"kmsKeyName\\": \\"", "{{$.inputs.parameters[\'encryption_spec_key_name\']}}", "\\"}", "}"]}', '--gcp_resources', "{{$.outputs.parameters['gcp_resources'].output_file}}", '--executor_input', '{{$}}']to be submitted as args to the ENTRYPOINT script instead of a command. This results in problems with the argparser and renders a user unable to run anything from GCPC with the Docker runner.
I was able to fix this on my machine with a with a pretty minor modification. Modify the docker run command here from the original:
container = client.containers.run(
image=image,
command=command,
detach=True,
stdout=True,
stderr=True,
volumes=volumes,
)to:
container = client.containers.run(
image=image,
entrypoint=[],
command=command,
detach=True,
stdout=True,
stderr=True,
volumes=volumes,
)While small, I can imagine this is a bigger change than is desirable. Would it be possible to add an additional runner class for running GCPC components locally?
github-actions[bot]
Environment
kfp 2.7.0 kfp-dsl 2.1.2 kfp-pipeline-spec 0.3.0 kfp-server-api 2.0.5
Steps to reproduce
Throws the following error:
Expected result
Materials and Reference
The expected result above was achieved by switching out the docker image used by the
google-cloud-pipeline-componentscomponents by default. The error looks like the entrypoint isn't being overridden as intended.google_cloud_pipeline_components/container/v1/aiplatform/remote_runner.pyis used as the executed file instead ofgoogle_cloud_pipeline_components.container.v1.bigquery.query_job.launcherImpacted by this bug? Give it a 👍.