Importer of Examples does not work with KubeflowV2DagRunner

System information

• Have I specified the code to reproduce the issue: Yes
• Environment in which the code is executed: Interactive Notebook (beam) on Google Cloud that Trigger Pipelines in Kubeflow
• TensorFlow version: 2.5.0
• TFX Version: 1.0.0rc1
• Python version: 3.7
• Python dependencies (from pip freeze output):

...

Describe the current behavior

We would like to use an importer to reload Examples already generated.

1 - It perfectly works when using the local BeamDagRunner. 2 - It does not work when using the KubeflowV2DagRunner. The pipeline is created but the importer fails with the following message Unmapped PropertyType specified in Yaml definition: 0; Failed to parse the artifact yaml schema. Project: XXX

Note: It works for all other artifact types.

Describe the expected behavior

It should work when using KubeflowV2DagRunner.

Standalone code to reproduce the issue

import os
import uuid

from kfp.v2.google import client

from tfx.dsl.components.common.importer import Importer
from tfx.types import standard_artifacts
from tfx.orchestration.beam.beam_dag_runner import BeamDagRunner
from tfx.orchestration.kubeflow.v2.kubeflow_v2_dag_runner import KubeflowV2DagRunner, KubeflowV2DagRunnerConfig
from tfx.orchestration import pipeline
from tfx.orchestration import metadata
from tfx.components import StatisticsGen


####################################################################################################################
# /!\ To Fill with your Values /!\ 
_project_id = 'XXX'
_gcp_region = 'europe-west4'
_gcs_bucket_name = 'XXX'
_examples_to_import = "gs://XXX"


####################################################################################################################
# Pipeline Variables
_pipeline_name = f"test-examples-importer-{str(uuid.uuid4())[0:4]}"
_pipeline_root = 'gs://{}/pipeline_root/{}'.format(_gcs_bucket_name, _pipeline_name)
_module_root = 'gs://{}/pipeline_module/{}'.format(_gcs_bucket_name, _pipeline_name)
_temp_location = 'gs://{}/pipeline_tmp/{}'.format(_gcs_bucket_name, _pipeline_name)
_pipeline_definition_file = 'training_pipeline.json'
_beam_bq_pipeline_args = [
    '--runner=DirectRunner',
    '--direct_running_mode=multi_processing',
    '--direct_num_workers=0',
    '--temp_location=' + _temp_location,
    '--project=' + _project_id,
    '--region=' + _gcp_region
]


####################################################################################################################
# Define a Function to Create a Simple Pipeline with 2 Components:
# - an Importer (of Examples)
# - a StatisticGen
def _create_pipeline(pipeline_name: str, 
                     pipeline_root: str, 
                     beam_pipeline_args: list,
                     examples_to_import: str) -> pipeline.Pipeline:
        
    examples_importer = Importer(source_uri=examples_to_import,
                                 artifact_type=standard_artifacts.Examples,
                                 properties={'split_names': '["train", "eval"]'},
                                 custom_properties={"tfx_version": "1.0.0"},
                                 reimport=False)

    statistics_gen = StatisticsGen(examples=examples_importer.outputs['result'])

    return pipeline.Pipeline(
          pipeline_name=pipeline_name,
          pipeline_root=pipeline_root,
          components=[
              examples_importer, 
              statistics_gen
            ],
        beam_pipeline_args=beam_pipeline_args,
        metadata_connection_config=metadata.sqlite_metadata_connection_config(os.path.join(os.environ['HOME'], 'tfx_metadata', pipeline_name, 'metadata.db'))
      )


####################################################################################################################
# Create the Pipeline
pipeline = _create_pipeline(_pipeline_name, _pipeline_root, _beam_bq_pipeline_args, _examples_to_import)


####################################################################################################################
# Run Locally -> OK
BeamDagRunner().run(pipeline)


####################################################################################################################
# Run on Vertex Pipelines -> KO
runner = KubeflowV2DagRunner(config=KubeflowV2DagRunnerConfig(), output_filename=_pipeline_definition_file)
runner.run(pipeline)
pipelines_client = client.AIPlatformClient(project_id=_project_id, region=_gcp_region,)
_ = pipelines_client.create_run_from_job_spec(_pipeline_definition_file)


####################################################################################################################

Other info / logs

The importer fails with the following message:

Unmapped PropertyType specified in Yaml definition: 0; Failed to parse the artifact yaml schema. Project: XXX

[hilcj]

From GCP Vertex AI - Pipeline team: we are working on a backend fixation that will address the issue, ETA end of next week.

[hilcj]

Hello @hilcj, any news on the fix ? thx

Hi @axelborja,

I just tested my end that the issue seems fixed.

Please give a try and feel free to reach us if you found the issue persist from your end 😃