def test_build_component_inputs_spec(self):
pipeline_params = [
dsl.PipelineParam(name='input1', param_type='Dataset'),
dsl.PipelineParam(name='input2', param_type='Integer'),
dsl.PipelineParam(name='input3', param_type='String'),
dsl.PipelineParam(name='input4', param_type='Float'),
]
expected_dict = {
'inputDefinitions': {
'artifacts': {
'input1': {
'artifactType': {
'instanceSchema':
'properties:\ntitle: kfp.Dataset\ntype: object\n'
}
}
},
'parameters': {
'input2': {
'type': 'INT'
},
'input3': {
'type': 'STRING'
},
'input4': {
'type': 'DOUBLE'
}
}
}
}
expected_spec = pipeline_spec_pb2.ComponentSpec()
json_format.ParseDict(expected_dict, expected_spec)
component_spec = pipeline_spec_pb2.ComponentSpec()
dsl_component_spec.build_component_inputs_spec(component_spec,
pipeline_params)
self.assertEqual(expected_spec, component_spec)
def _group_to_dag_spec(
self,
group: dsl.OpsGroup,
inputs: Dict[str, List[Tuple[dsl.PipelineParam, str]]],
outputs: Dict[str, List[Tuple[dsl.PipelineParam, str]]],
dependencies: Dict[str, List[_GroupOrOp]],
pipeline_spec: pipeline_spec_pb2.PipelineSpec,
deployment_config: pipeline_spec_pb2.PipelineDeploymentConfig,
rootgroup_name: str,
) -> None:
"""Generate IR spec given an OpsGroup.
Args:
group: The OpsGroup to generate spec for.
inputs: The inputs dictionary. The keys are group/op names and values are
lists of tuples (param, producing_op_name).
outputs: The outputs dictionary. The keys are group/op names and values
are lists of tuples (param, producing_op_name).
dependencies: The group dependencies dictionary. The keys are group/op
names, and the values are lists of dependent groups/ops.
pipeline_spec: The pipeline_spec to update in-place.
deployment_config: The deployment_config to hold all executors.
rootgroup_name: The name of the group root. Used to determine whether the
component spec for the current group should be the root dag.
"""
group_component_name = dsl_utils.sanitize_component_name(group.name)
if group.name == rootgroup_name:
group_component_spec = pipeline_spec.root
else:
group_component_spec = pipeline_spec.components[group_component_name]
# Generate component inputs spec.
if inputs.get(group.name, None):
dsl_component_spec.build_component_inputs_spec(
group_component_spec, [param for param, _ in inputs[group.name]])
# Generate component outputs spec.
if outputs.get(group.name, None):
group_component_spec.output_definitions.CopyFrom(
dsl_component_spec.build_component_outputs_spec(
[param for param, _ in outputs[group.name]]))
# Generate task specs and component specs for the dag.
subgroups = group.groups + group.ops
for subgroup in subgroups:
subgroup_task_spec = getattr(subgroup, 'task_spec',
pipeline_spec_pb2.PipelineTaskSpec())
subgroup_component_spec = getattr(subgroup, 'component_spec',
pipeline_spec_pb2.ComponentSpec())
is_loop_subgroup = (isinstance(group, dsl.ParallelFor))
is_recursive_subgroup = (
isinstance(subgroup, dsl.OpsGroup) and subgroup.recursive_ref)
# Special handling for recursive subgroup: use the existing opsgroup name
if is_recursive_subgroup:
subgroup_key = subgroup.recursive_ref.name
else:
subgroup_key = subgroup.name
subgroup_task_spec.task_info.name = dsl_utils.sanitize_task_name(
subgroup_key)
# human_name exists for ops only, and is used to de-dupe component spec.
subgroup_component_name = dsl_utils.sanitize_component_name(
getattr(subgroup, 'human_name', subgroup_key))
subgroup_task_spec.component_ref.name = subgroup_component_name
if isinstance(subgroup, dsl.OpsGroup) and subgroup.type == 'condition':
condition = subgroup.condition
operand_values = []
subgroup_inputs = inputs.get(subgroup.name, [])
subgroup_params = [param for param, _ in subgroup_inputs]
tasks_in_current_dag = [subgroup.name for subgroup in subgroups]
dsl_component_spec.build_component_inputs_spec(
subgroup_component_spec,
subgroup_params,
)
dsl_component_spec.build_task_inputs_spec(
subgroup_task_spec,
subgroup_params,
tasks_in_current_dag,
)
for operand in [condition.operand1, condition.operand2]:
operand_values.append(self._resolve_value_or_reference(operand))
condition_string = '{} {} {}'.format(operand_values[0],
condition.operator,
operand_values[1])
subgroup_task_spec.trigger_policy.CopyFrom(
pipeline_spec_pb2.PipelineTaskSpec.TriggerPolicy(
condition=condition_string))
# Generate dependencies section for this task.
if dependencies.get(subgroup.name, None):
group_dependencies = list(dependencies[subgroup.name])
group_dependencies.sort()
subgroup_task_spec.dependent_tasks.extend(
[dsl_utils.sanitize_task_name(dep) for dep in group_dependencies])
if isinstance(subgroup, dsl.ParallelFor):
# Remove loop arguments related inputs from parent group component spec.
input_names = [param.full_name for param, _ in inputs[subgroup.name]]
for input_name in input_names:
if _for_loop.LoopArguments.name_is_loop_argument(input_name):
dsl_component_spec.pop_input_from_component_spec(
group_component_spec, input_name)
if subgroup.items_is_pipeline_param:
# These loop args are a 'withParam' rather than 'withItems'.
# i.e., rather than a static list, they are either the output of
# another task or were input as global pipeline parameters.
pipeline_param = subgroup.loop_args.items_or_pipeline_param
input_parameter_name = pipeline_param.full_name
if pipeline_param.op_name:
subgroup_task_spec.inputs.parameters[
input_parameter_name].task_output_parameter.producer_task = (
dsl_utils.sanitize_task_name(pipeline_param.op_name))
subgroup_task_spec.inputs.parameters[
input_parameter_name].task_output_parameter.output_parameter_key = (
pipeline_param.name)
else:
subgroup_task_spec.inputs.parameters[
input_parameter_name].component_input_parameter = (
input_parameter_name)
# Correct loop argument input type in the parent component spec.
# The loop argument was categorized as an artifact due to its missing
# or non-primitive type annotation. But it should always be String
# typed, as its value is a serialized JSON string.
dsl_component_spec.pop_input_from_component_spec(
group_component_spec, input_parameter_name)
group_component_spec.input_definitions.parameters[
input_parameter_name].type = pipeline_spec_pb2.PrimitiveType.STRING
# Additional spec modifications for dsl.ParallelFor's subgroups.
if is_loop_subgroup:
self._update_loop_specs(group, subgroup, group_component_spec,
subgroup_component_spec, subgroup_task_spec)
# Add importer node when applicable
for input_name in subgroup_task_spec.inputs.artifacts:
if not subgroup_task_spec.inputs.artifacts[
input_name].task_output_artifact.producer_task:
type_schema = type_utils.get_input_artifact_type_schema(
input_name, subgroup._metadata.inputs)
importer_name = importer_node.generate_importer_base_name(
dependent_task_name=subgroup_task_spec.task_info.name,
input_name=input_name)
importer_task_spec = importer_node.build_importer_task_spec(
importer_name)
importer_comp_spec = importer_node.build_importer_component_spec(
importer_base_name=importer_name,
input_name=input_name,
input_type_schema=type_schema)
importer_task_name = importer_task_spec.task_info.name
importer_comp_name = importer_task_spec.component_ref.name
importer_exec_label = importer_comp_spec.executor_label
group_component_spec.dag.tasks[importer_task_name].CopyFrom(
importer_task_spec)
pipeline_spec.components[importer_comp_name].CopyFrom(
importer_comp_spec)
subgroup_task_spec.inputs.artifacts[
input_name].task_output_artifact.producer_task = (
importer_task_name)
subgroup_task_spec.inputs.artifacts[
input_name].task_output_artifact.output_artifact_key = (
importer_node.OUTPUT_KEY)
# Retrieve the pre-built importer spec
importer_spec = subgroup.importer_specs[input_name]
deployment_config.executors[importer_exec_label].importer.CopyFrom(
importer_spec)
# Add component spec if not exists
if subgroup_component_name not in pipeline_spec.components:
pipeline_spec.components[subgroup_component_name].CopyFrom(
subgroup_component_spec)
# Add task spec
group_component_spec.dag.tasks[
subgroup_task_spec.task_info.name].CopyFrom(subgroup_task_spec)
# Add executor spec, if applicable.
container_spec = getattr(subgroup, 'container_spec', None)
if container_spec:
if compiler_utils.is_v2_component(subgroup):
compiler_utils.refactor_v2_container_spec(container_spec)
executor_label = subgroup_component_spec.executor_label
if executor_label not in deployment_config.executors:
deployment_config.executors[executor_label].container.CopyFrom(
container_spec)
# Add AIPlatformCustomJobSpec, if applicable.
custom_job_spec = getattr(subgroup, 'custom_job_spec', None)
if custom_job_spec:
executor_label = subgroup_component_spec.executor_label
if executor_label not in deployment_config.executors:
deployment_config.executors[
executor_label].custom_job.custom_job.update(custom_job_spec)
pipeline_spec.deployment_spec.update(
json_format.MessageToDict(deployment_config))
def _create_pipeline_spec(
self,
args: List[dsl.PipelineParam],
pipeline: dsl.Pipeline,
) -> pipeline_spec_pb2.PipelineSpec:
"""Creates the pipeline spec object.
Args:
args: The list of pipeline arguments.
pipeline: The instantiated pipeline object.
Returns:
A PipelineSpec proto representing the compiled pipeline.
Raises:
NotImplementedError if the argument is of unsupported types.
"""
compiler_utils.validate_pipeline_name(pipeline.name)
deployment_config = pipeline_spec_pb2.PipelineDeploymentConfig()
pipeline_spec = pipeline_spec_pb2.PipelineSpec()
pipeline_spec.pipeline_info.name = pipeline.name
pipeline_spec.sdk_version = 'kfp-{}'.format(kfp.__version__)
# Schema version 2.0.0 is required for kfp-pipeline-spec>0.1.3.1
pipeline_spec.schema_version = '2.0.0'
dsl_component_spec.build_component_inputs_spec(pipeline_spec.root, args)
root_group = pipeline.groups[0]
opsgroups = self._get_groups(root_group)
op_name_to_parent_groups = self._get_groups_for_ops(root_group)
opgroup_name_to_parent_groups = self._get_groups_for_opsgroups(root_group)
condition_params = self._get_condition_params_for_ops(root_group)
op_name_to_for_loop_op = self._get_for_loop_ops(root_group)
inputs, outputs = self._get_inputs_outputs(
pipeline,
root_group,
op_name_to_parent_groups,
opgroup_name_to_parent_groups,
condition_params,
op_name_to_for_loop_op,
)
dependencies = self._get_dependencies(
pipeline,
root_group,
op_name_to_parent_groups,
opgroup_name_to_parent_groups,
opsgroups,
condition_params,
)
for opsgroup_name in opsgroups.keys():
self._group_to_dag_spec(
opsgroups[opsgroup_name],
inputs,
outputs,
dependencies,
pipeline_spec,
deployment_config,
root_group.name,
)
return pipeline_spec
def _group_to_dag_spec(
self,
group: dsl.OpsGroup,
inputs: Dict[str, List[Tuple[dsl.PipelineParam, str]]],
outputs: Dict[str, List[Tuple[dsl.PipelineParam, str]]],
dependencies: Dict[str, List[_GroupOrOp]],
pipeline_spec: pipeline_spec_pb2.PipelineSpec,
rootgroup_name: str,
) -> None:
"""Generate IR spec given an OpsGroup.
Args:
group: The OpsGroup to generate spec for.
inputs: The inputs dictionary. The keys are group/op names and values are
lists of tuples (param, producing_op_name).
outputs: The outputs dictionary. The keys are group/op names and values
are lists of tuples (param, producing_op_name).
dependencies: The group dependencies dictionary. The keys are group/op
names, and the values are lists of dependent groups/ops.
pipeline_spec: The pipeline_spec to update in-place.
rootgroup_name: The name of the group root. Used to determine whether the
component spec for the current group should be the root dag.
"""
group_component_name = dsl_utils.sanitize_component_name(group.name)
if group.name == rootgroup_name:
group_component_spec = pipeline_spec.root
else:
group_component_spec = pipeline_spec.components[
group_component_name]
deployment_config = pipeline_spec_pb2.PipelineDeploymentConfig()
# Generate component inputs spec.
if inputs.get(group.name, None):
dsl_component_spec.build_component_inputs_spec(
group_component_spec,
[param for param, _ in inputs[group.name]])
# Generate component outputs spec.
if outputs.get(group.name, None):
group_component_spec.output_definitions.CopyFrom(
dsl_component_spec.build_component_outputs_spec(
[param for param, _ in outputs[group.name]]))
# Generate task specs and component specs for the dag.
subgroups = group.groups + group.ops
for subgroup in subgroups:
subgroup_task_spec = getattr(subgroup, 'task_spec',
pipeline_spec_pb2.PipelineTaskSpec())
subgroup_component_spec = getattr(
subgroup, 'component_spec', pipeline_spec_pb2.ComponentSpec())
is_recursive_subgroup = (isinstance(subgroup, dsl.OpsGroup)
and subgroup.recursive_ref)
# Special handling for recursive subgroup: use the existing opsgroup name
if is_recursive_subgroup:
subgroup_key = subgroup.recursive_ref.name
else:
subgroup_key = subgroup.name
subgroup_task_spec.task_info.name = dsl_utils.sanitize_task_name(
subgroup_key)
# human_name exists for ops only, and is used to de-dupe component spec.
subgroup_component_name = dsl_utils.sanitize_component_name(
getattr(subgroup, 'human_name', subgroup_key))
subgroup_task_spec.component_ref.name = subgroup_component_name
if isinstance(subgroup,
dsl.OpsGroup) and subgroup.type == 'condition':
condition = subgroup.condition
operand_values = []
subgroup_inputs = inputs.get(subgroup.name, [])
subgroup_params = [param for param, _ in subgroup_inputs]
tasks_in_current_dag = [
subgroup.name for subgroup in subgroups
]
dsl_component_spec.build_component_inputs_spec(
subgroup_component_spec,
subgroup_params,
)
dsl_component_spec.build_task_inputs_spec(
subgroup_task_spec,
subgroup_params,
tasks_in_current_dag,
)
for operand in [condition.operand1, condition.operand2]:
operand_values.append(
self._resolve_value_or_reference(operand))
condition_string = '{} {} {}'.format(operand_values[0],
condition.operator,
operand_values[1])
subgroup_task_spec.trigger_policy.CopyFrom(
pipeline_spec_pb2.PipelineTaskSpec.TriggerPolicy(
condition=condition_string))
# Generate dependencies section for this task.
if dependencies.get(subgroup.name, None):
group_dependencies = list(dependencies[subgroup.name])
group_dependencies.sort()
subgroup_task_spec.dependent_tasks.extend([
dsl_utils.sanitize_task_name(dep)
for dep in group_dependencies
])
# Add importer node when applicable
for input_name in subgroup_task_spec.inputs.artifacts:
if not subgroup_task_spec.inputs.artifacts[
input_name].task_output_artifact.producer_task:
type_schema = type_utils.get_input_artifact_type_schema(
input_name, subgroup._metadata.inputs)
importer_name = importer_node.generate_importer_base_name(
dependent_task_name=subgroup_task_spec.task_info.name,
input_name=input_name)
importer_task_spec = importer_node.build_importer_task_spec(
importer_name)
importer_comp_spec = importer_node.build_importer_component_spec(
importer_base_name=importer_name,
input_name=input_name,
input_type_schema=type_schema)
importer_task_name = importer_task_spec.task_info.name
importer_comp_name = importer_task_spec.component_ref.name
importer_exec_label = importer_comp_spec.executor_label
group_component_spec.dag.tasks[
importer_task_name].CopyFrom(importer_task_spec)
pipeline_spec.components[importer_comp_name].CopyFrom(
importer_comp_spec)
subgroup_task_spec.inputs.artifacts[
input_name].task_output_artifact.producer_task = (
importer_task_name)
subgroup_task_spec.inputs.artifacts[
input_name].task_output_artifact.output_artifact_key = (
importer_node.OUTPUT_KEY)
# Retrieve the pre-built importer spec
importer_spec = subgroup.importer_specs[input_name]
deployment_config.executors[
importer_exec_label].importer.CopyFrom(importer_spec)
# Add component spec if not exists
if subgroup_component_name not in pipeline_spec.components:
pipeline_spec.components[subgroup_component_name].CopyFrom(
subgroup_component_spec)
# Add task spec
group_component_spec.dag.tasks[
subgroup_task_spec.task_info.name].CopyFrom(subgroup_task_spec)
# Add executor spec
container_spec = getattr(subgroup, 'container_spec', None)
if container_spec:
if compiler_utils.is_v2_component(subgroup):
compiler_utils.refactor_v2_container_spec(container_spec)
executor_label = subgroup_component_spec.executor_label
if executor_label not in deployment_config.executors:
deployment_config.executors[
executor_label].container.CopyFrom(container_spec)
pipeline_spec.deployment_spec.update(
json_format.MessageToDict(deployment_config))