Rflow-tfx

Major Features and Improvements

* Added experimental exit_handler support for KubeflowDagRunner.
* Enabled custom labels to be submitted to CAIP training jobs.
* Enabled custom resource-setting (vCPU and RAM) for containers orchestrating
on Vertex AI.

Breaking Changes

For Pipeline Authors

* N/A

For Component Authors

* N/A

Deprecations

* N/A

Bug Fixes and Other Changes

* `LatestBlessedModelStrategy` gracefully handles the case where there are no
blessed model at all (e.g. first run).
* Fix that the resolver with custom `ResolverStrategy` (assume correctly
packaged) fails.
* Fixed `ElwcBigQueryExampleGen` data serializiation error that was causing an
assertion failure on Beam.
* Added dark mode styling support for InteractiveContext notebook formatters.
* (Python 3.9+) Supports `list` and `dict` in type definition of execution
properties.
* Populate Artifact proto `name` field when name is set on the Artifact python
object.
* Temporarily capped `apache-airflow` version to 2.2.x to avoid dependency
conflict. We will rollback this change once `kfp` releases a new version.
* Fixed a compatibility issue with apache-airflow 2.3.0 that is failing with
"unexpected keyword argument 'default_args'".
* StatisticsGen will raise an error if unsupported StatsOptions (i.e.,
generators or experimental_slice_functions) are passed.
* Fixed a bug in the Artifact attribute setter that was causing the
corresponding getter not to return a value for properties of type JSON_VALUE.

Dependency Updates

| Package Name | Version Constraints | Previously (in `v1.7.0`) | Comments |
| -- | -- | -- | -- |
| `apache-beam[gcp]` | `>=2.38,<3` | `>=2.36,<3` | Synced release train |

Documentation Updates

* N/A

Major Features and Improvements

* Added support for list-type Placeholder.
* Added support for function-based custom component with beam pipeline.

Breaking Changes

For Pipeline Authors

* N/A

For Component Authors

* N/A

Deprecations

* Removed the already-deprecated components.ImporterNode, should use
v1.dsl.Importer instead.
* Deprecated Channel property setters. Use constructor argument instead.

Bug Fixes and Other Changes

* Fixed the cluster spec error in CAIP Tuner on Vertex when
`num_parallel_trials = 1`
* Replaced deprecated assertDictContainsSubset with
assertLessEqual(itemsA, itemsB).
* Updating Keras tutorial to make better use of Keras, and better feature
engineering.
* Merges KFP UI Metadata file if it already exists. Now components can produce
their own UI results and it will be merged with existing visualization.
* Switch Transform component to always use sketch when computing top-k stats.

Dependency Updates

| Package Name | Version Constraints | Previously (in `v1.6.0`) | Comments |
| -- | -- | -- | -- |
| `apache-beam[gcp]` | `~=2.36` | `~=2.35` | Synced release train |

Major Features and Improvements

* N/A

Breaking Changes

For Pipeline Authors

* N/A

For Component Authors

* N/A

Deprecations

* N/A

Bug Fixes and Other Changes

* N/A
Dependency Updates

| Package Name | Version Constraints | Previously (in `v1.6.0`) | Comments |
| -- | -- | -- | -- |

Major Features and Improvements

* N/A

Breaking Changes

For Pipeline Authors

* N/A

For Component Authors

* N/A

Deprecations

* N/A

Bug Fixes and Other Changes

* Fixed `Pusher` issue that didn't copy files other than
`saved_model.pb`.

Documentation Updates

* N/A

Major Features and Improvements

* Added experimental support for TensorFlow Decision Forests models.
* Added Boolean type value artifacts.
* Function components defined with `component` may now have optional/nullable
primitive type return values when `Optional[T]` is used in the return type
OutputDict.
* Supported endpoint overwrite for CAIP Tuner. Users can use the `keras-tuner`
module or any tuner that implements the `keras_tuner.Tuner` interface for
(parallel) tuning on Vertex.

Breaking Changes

For Pipeline Authors

* N/A

For Component Authors

* N/A

Deprecations

* N/A

Bug Fixes and Other Changes
* Pusher now copies the `saved_model.pb` file at last to prevent loading
SavedModel on invalid (partially available) directory state.
* Always disable caching for exit handlers in Kubeflow V2 runner to
reflect latest status of dependent dag.

Dependency Updates

| Package Name | Version Constraints | Previously (in `v1.5.0`) | Comments |
| -- | -- | -- | -- |