* fix(sdk): compile ParallelFor in a deterministic manner
During compilataion ParallelFor components end up with randomized names,
which makes it very inconvenient to compare two versions of a pipeline.
This commit fixes this issue.
* fix(sdk): fix new parallel-for test cases
* SDK - Components - Fixed python components that use \n
The escape sequence was being replaced by the `echo` command.
Apparently, unlike in the `bash` shell, the `echo` command of the `sh` shell expands the escape sequences by default and does not support an option to turn it off. (For some reason the -n option works properly even though it should not).
Fixes https://github.com/kubeflow/pipelines/issues/4939
* Fixed the test data
* Fixed the deprecated container component builder
* Fixed the new compiler test case
* Added test
* add placeholder to spec
* add output_directory to pipeline
* respect uri placeholder in file outputs
* wip: add data passing rewriting logic to respect the uri semantics
* merge input_uri and paths when instantiating ContainerOp
* fix
* fix workflow rewriting
* Add topology rewriting
* add a test case, and various fixes
* make the test case more complex
* Fix the case when working with OpsGroup
* Fix test case
* fix resolving test
* fix redundant cmd lines
* fix redundant cmd lines
* resolve comments
* fix file outputs
* resolve comments
* copy file outputs instead of modifying inplace.
Currently were running the python code inline using `python -c <code>`.
This has two issues:
1) Python does not show source code line in exception stack traces
2) inspect.getsource does not work. This method is used in PyTorch JIT for example.
We solve these issues by writing the code into a file before executing it.
The disadvantage of the new approach is that it adds complexity, a filesystem write operation and also requires the `sh` executable to be present (we could replace it with python-based program if needed).
* SDK - Tests - Restored the ParallelFor compiler test data
Fixes https://github.com/kubeflow/pipelines/issues/4102
* Removed the pipeline-sdk-type annotations
* Fixed the test_artifact_passing_using_volume test data
* SDK - Compiler - Added support for volume-based data passing
Currently artifact passing is performed by Argo sidecar containers what download input data and upload output data to artifact repository (usually, S3-compatible blob storage like Minio).
The performance of this method is not optimal and it requires that pod disks have enough capacity to hold all artifact data.
This commit adds support for volume-based data passing.
This method involves using a single milti-write Kubernetes data volume to pass all intermediate data.
Parts of the volume are mounted to the input/output artifact directories, so when the user program reads and writes files, the files actually reside in the data volume.
This method improves the performance and reduces storage resource requirements.
The data volume must exist and support "READ_WRITE_MANY".
Limitations:
* All artifact file names must be the same (e.g. "data"). All auto-generated paths are already consistent. Avoid using any hard-coded paths.
* Passing constant values (text) as arguments for artifact inputs is not supported.
* The feature is experimental.
* Added data_passing_methods.KubernetesVolume
This class represents a configured volume-based artifact passing method.
* Added PipelineConf.data_passing_method
This property allows setting the method that will be used for intermediate data passing.
Added the compiler support for the new feature.
Example:
```python
from kfp.dsl import PipelineConf, data_passing_methods
from kubernetes.client.models import V1Volume, V1PersistentVolumeClaim
pipeline_conf = PipelineConf()
pipeline_conf.data_passing_method = data_passing_methods.KubernetesVolume(
volume=V1Volume(
name='data',
persistent_volume_claim=V1PersistentVolumeClaim('data-volume'),
),
path_prefix='artifact_data/',
)
```
* Added unit test
* Fixed bug in the unit test
Kubernetes does not validate the structures at all...
* Fixed bug in the result structure
* Fixed the test data
The class should be V1PersistentVolumeClaimVolumeSource, not V1PersistentVolumeClaimSpec.
* Fixed the test
Previously the default image was set to an old version of tensorflow image. That image is now outdated. It's also framework-specific and pretty big.
We're switching to the official python image which is small, official and framework-agnostic.
The users can easily switch to the old behavior by just specifying `base_image='tensorflow/tensorflow:1.13.2-py3'` during the component creation.
* SDK - Components - Stabilize JSON serialization by sorting keys
Otherwise serialization of the default values of the component/pipeline inputs is unstable on Python 3.5.
* Fixed the test data
* SDK - Made outputs with original names available in ContainerOp.outputs
Previously, ContainerOp had strict requirements for the output names, so we had to convert all the names before passing them to the ContainerOp constructor. Outputs with non-pythonic names could not be accessed using their original names.
Now ContainerOp supports any output names, so we're now using the original output names.
However to support legacy pipelines, we're also adding output references with pythonic names.
* Fixed the compiler test data
* Fixed the duplicate parameter outputs in the compiled workflow
* Fixed long line
* Stabilized the output naming conflict resolution
* Fix case of missing special outputs
* SDK - Prioritize lib2to3 when stripping type annotations
It's a standard python library (although not well supported) and it doe not leave training spaces.
* Fixed compiler test data
* Add kfp-container-builder sa
* Allow service account to be configurable
* Fix tests
* Fix test
* Use documentation for service account to introduce compatibility with different types of installation
* updated doc
* clean up
* Update container_builder_test.py
* Update _build_image_api.py
* Update kustomization.yaml
* Add executable permission for presubmit tests mkp.sh
* SDK - Reduce python component limitations - no import errors for custom type annotations
By default, create_component_from_func copies the source code of the function and creates a component using that source code. No global imports are captured. This is problematic for the function definition, since any annotation, that uses a type that needs to be imported, will cause error. There were some special provisions for
NamedTuple, InputPath and OutputPath, but even they were brittle (for example, "typing.NamedTuple" or "components.InputPath" annotations still caused failures at runtime).
This commit fixes the issue by stripping the type annotations from function declarations.
Fixes cases that were failing before:
```python
import typing
import collections
MyFuncOutputs = typing.NamedTuple('Outputs', [('sum', int), ('product', int)])
@create_component_from_func
def my_func(
param1: CustomType, # This caused failure previously
param2: collections.OrderedDict, # This caused failure previously
) -> MyFuncOutputs: # This caused failure previously
pass
```
* Fixed the compiler tests
* Fixed crashes on print function
Code `print(line, end="")` was causing error: "lib2to3.pgen2.parse.ParseError: bad input: type=22, value='=', context=('', (2, 15))"
* Using the strip_hints library to strip the annotations
* Updating test workflow yamls
* Workaround for bug in untokenize
* Switched to the new strip_string_to_string method
* Fixed typo.
Co-Authored-By: Jiaxiao Zheng <jxzheng@google.com>
Co-authored-by: Jiaxiao Zheng <jxzheng@google.com>
* [Testing] Use gke 1.15.8 to mitigate workload identity flakiness
* Upgrade gcloud version
* Update image builder image too
* Turn on workload identity
* Update deploy-cluster.sh
* secret sample uses python3 instead
* Increase xgboost time limit
* Revert files with bad format
* Update component and pipelines to use gcloud 279.0.0
* Fix secret sample using python3
* Upgrade frontend integration test image
* Rebuild frontend integration test image
* SDK - Compiler - Fixed ParallelFor name clashes
The ParallelFor argument reference resolving was really broken.
The logic "worked" like this - of the name of the referenced output
contained the name of the loop collection source output, then it was
considered to be the reference to the loop item.
This broke lots of scenarios especially in cases where there were
multiple components with same output name (e.g. the default "Output"
output name). The logic also did not distinguish between references to
the loop collection item vs. references to the loop collection source
itself.
I've rewritten the argument resolving logic, to fix the issues.
* Argo cannot use {{item}} when withParams items are dicts
* Stabilize the loop template names
* Renamed the test case
* Replaced `_instance_to_dict(obj)` with `obj.to_dict()`
* Fixed the capitalization in _python_function_name_to_component_name
It now only changes the case of the first letter.
* Replaced the _extract_component_metadata function with _extract_component_interface
* Stopped adding newline to the component description.
* Handling None inputs and outputs
* Not including emply inputs and outputs in component spec
* Renamed the private attributes that the @pipeline decorator sets
* Changged _extract_pipeline_metadata to use _extract_component_interface
* Fixed issues based on feedback
* Script to set up workload identity for standalone deployment
* Migrate tests to run on standalone + workload identity
* Fix test script
* Switch to static GSAs for testing, because they have name length limit
* Add workload identity binding for argo
* Fix argo workload identity bindings
* Remove user-gcp-sa from tests
* Remove use_gcp_secret from xgboost sample
* Allow debugging tests locally
* Wait for policies to take effect
* Update deploy-pipeline-lite.sh
* Update deploy-pipeline-lite.sh
* [WIP] test gcloud auth list with test-runner sa
* Add namespace
* test again
* Use new image builder
* test again
* Remove debug code
* Remove usages of use_gcp_secret
* Fix unit test and tensorboard pod template
* Add debug code again to test
* Try waiting until workload identity bindings are ready
* Fix some other samples
* Fix parameterized tfx oss sample
* Add retry to image building
* Try fixing tfx oss sample
* Fix compiled tfx oss sample
* Update all google/cloud-sdk to latest
* Try fixing parameterized tfx oss sample again
* Also verify pipeline-runner ksa is working
* Fix parameterized_tfx_oss sample
* Update gcp-workload-identity-setup.sh
* Revert unneeded change
* Pin to new google/cloud-sdk
* Remove wrongly commited binaries
* SDK - Compiler - Move volumes to templates
Argo v2.3.0+ supports per-template volume specs similiar to Kubernetes. Prior to version 2.3.0 Argo only supported workflow-level volume specs.
We had several outstanding issues caused by the need to put all volumes in the same place.
There was also the issue with input parameter reference placeholders in volume specifications which were placed outside their home templates declaring the inputs.
This change fixes those issues.
* Removed dead code line
* SDK - Compiler - Fix large data passing
Stop outputting parameters unless they're consumed as parameters downstream.
This prevents the situaltion when component outputs a big file, but DSL compiler instructs Argo to pick it up as parameter (parameters only hold few kilobytes of data).
As byproduct, this change fixes some minor compiler data passing bugs where some parameters were being passed around, but never consumed (happened with `ResourceOp`, `dsl.Condition` and recursion).
* Replaced ... with `raise AssertionError`
* Fixed small bug
* Removed unused variables
* Fixed names of the mark_upstream_ios_of_* functions
* Fixed detection of parameter output references
* Fixed handling of volumes
Two PRs have been merged that turned out to be slightly incompatible. This PR fixes the failing tests.
Root causes:
* The pipeline parameter default values were not properly serialized when constructing the metadata object.
* The `ParameterMeta` class did not validate the default value type, so the lack of serialization has not been caught. The `ParameterMeta` was replaced by `InputSpec` which has strict type validation.
* Previously we did not have samples with complex pipeline parameter default values (e.g. lists) that could trigger the failures. Then two samples were added that had complex default values.
* Travis does not re-run tests before merging
* Prow does not re-run Travis tests before merging
Currently, the parameter output values are not saved to storage and their values are lost as soon as garbage collector removes the workflow object.
This change makes is so the parameter output values are persisted.
* first working commit
* incrememtal commit
* in the middle of converting loop args constructor to accept pipeline param
* both cases working
* output works, passed doesn't
* about to redo compiler section
* rewrite draft done
* added withparam tests
* removed sdk/python/comp.yaml
* minor
* subvars work
* more tests
* removed unneeded artifact outputs from test yaml
* sort keys
* removed dead artifact code
* Refactor. Expose a public API to append pipeline param without interacting with dsl.Pipeline obj.
* Add unit test and fix.
* Fix docstring.
* Fix test
* Fix test
* Fix two nit problems
* Refactor
* Explicitly added mlpipeline outputs to the components that actually produce them
* Updated samples
* SDK - DSL - Stopped adding mlpipeline artifacts to every compiled template
Fixes https://github.com/kubeflow/pipelines/issues/1421
Fixes https://github.com/kubeflow/pipelines/issues/1422
* Updated the Lighweight sample
* Updated the compiler tests
* Fixed the lightweight sample
* Reverted the change to one contrib/samples/openvino
The sample will still work fine as it is now.
I'll add the change to that file as a separate PR.
* SDK - Added support for raw artifact values to ContainerOp
* `ContainerOp` now gets artifact artguments from command line instead of the constructor.
* Added back input_artifact_arguments to the ContainerOp constructor.
In some scenarios it's hard to provide the artifact arguments through the `command` list when it already has resolved artifact paths.
* Exporting InputArtifactArgument from kfp.dsl
* Updated the sample
* Properly passing artifact arguments as task arguments
as opposed to default input values.
* Renamed input_artifact_arguments to artifact_arguments to reduce confusion
* Renamed InputArtifactArgument to InputArgumentPath
Also renamed input_artifact_arguments to artifact_argument_paths in the ContainerOp's constructor
* Replaced getattr with isinstance checks.
getattr is too fragile and can be broken by renames.
* Fixed the type annotations
* Unlocked the input artifact support in components
Added the test_input_path_placeholder_with_constant_argument test
Jiaxiao Zheng
Chen Sun
Vitalii Vokhmin
Alexey Volkov
radcheb
Victor
Yuan (Bob) Gong
Alexey Volkov
Ning
Kevin Bache