[v2compat] finalize pipeline run MLMD schema

[Bobgy]

TODOs:

• Design, figure out the exact list of items we need to change (size/M)

Known work items:

• [v2compat] move component emitted metadata to metadata struct field #5788 (size/XS)
  • chore(launcher): revert metadata to custom properties & artifact name -> event path #5842
• [v2compat] log MLMD artifact state during execution #5789 (size/S)
• [v2compat] remove redundant fields “pipeline_name”, “pipeline_run_id” from execution custom properties
• [v2compat] Unify pipeline name and pipeline context name #5978
• make scalar metrics distinguishable. UPDATE: figured out that we treat all properties of type float as scalar metrics in metadata (custom properties).
• [v2compat] move types from kfp.XXX namespace to system.XXX namespace #5985
• [v2compat] re-evaluate execution custom properties schema #5986
• [v2compat] use declared inputs/outputs in MLMD model #5803 (decided this is not necessary)
• P2 pipeline root task execution

[Bobgy]

/assign @capri-xiyue

[Bobgy]

b/187556737

[Bobgy]

One problem I found from #5668, is that

pipelines/samples/test/metrics_visualization_v2.py

Line 62 in 8936431

metrics.log_metric('accuracy', (result*100.0))

For now, we log a custom property called "accuracy" in the artifact, but if we do not know in advance which metric names are logged, we cannot know which custom properties correspond to scalar metrics.

Therefore, it seems we need more structured info for metrics.

[zijianjoy]

For now, we log a custom property called "accuracy" in the artifact, but if we do not know in advance which metric names are logged, we cannot know which custom properties correspond to scalar metrics.
Therefore, it seems we need more structured info for metrics.

SDK allows user to define any mount of log_metric call, therefore scalar metric is a dictionary instead of single value. See this example:

See this SDK implementation: https://github.com/kubeflow/pipelines/blob/master/sdk/python/kfp/dsl/io_types.py#L112-L119

In order to visualize this scalar metrics, we need to create a table: one column for metric name, another column for metric value.

[zijianjoy]

But we also need to distinguish confidenceMetrics and confusionMatrix from scalar metrics. Otherwise they will appear from the same dictionary which we need to explicitly filter: https://github.com/kubeflow/pipelines/blob/master/sdk/python/kfp/dsl/test_data/expected_io_types_classification_metrics.json

[Bobgy]

Needs further work item breakdown