pipelines/sdk/python/kfp/deprecated/compiler/_op_to_template.py

365 lines
14 KiB
Python

# Copyright 2019 The Kubeflow Authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import json
import re
import warnings
import yaml
import copy
from collections import OrderedDict
from typing import Union, List, Any, Callable, TypeVar, Dict
from kfp.deprecated.compiler._k8s_helper import convert_k8s_obj_to_json
from kfp.deprecated import dsl
from kfp.deprecated.dsl._container_op import BaseOp
# generics
T = TypeVar('T')
def _process_obj(obj: Any, map_to_tmpl_var: dict):
"""Recursively sanitize and replace any PipelineParam (instances and
serialized strings) in the object with the corresponding template variables
(i.e. '{{inputs.parameters.<PipelineParam.full_name>}}').
Args:
obj: any obj that may have PipelineParam
map_to_tmpl_var: a dict that maps an unsanitized pipeline
params signature into a template var
"""
# serialized str might be unsanitized
if isinstance(obj, str):
# get signature
param_tuples = dsl.match_serialized_pipelineparam(obj)
if not param_tuples:
return obj
# replace all unsanitized signature with template var
for param_tuple in param_tuples:
obj = re.sub(param_tuple.pattern,
map_to_tmpl_var[param_tuple.pattern], obj)
# list
if isinstance(obj, list):
return [_process_obj(item, map_to_tmpl_var) for item in obj]
# tuple
if isinstance(obj, tuple):
return tuple((_process_obj(item, map_to_tmpl_var) for item in obj))
# dict
if isinstance(obj, dict):
return {
_process_obj(key, map_to_tmpl_var):
_process_obj(value, map_to_tmpl_var) for key, value in obj.items()
}
# pipelineparam
if isinstance(obj, dsl.PipelineParam):
# if not found in unsanitized map, then likely to be sanitized
return map_to_tmpl_var.get(
str(obj), '{{inputs.parameters.%s}}' % obj.full_name)
# k8s objects (generated from swaggercodegen)
if hasattr(obj, 'attribute_map') and isinstance(obj.attribute_map, dict):
# process everything inside recursively
for key in obj.attribute_map.keys():
setattr(obj, key, _process_obj(getattr(obj, key), map_to_tmpl_var))
# return json representation of the k8s obj
return convert_k8s_obj_to_json(obj)
# do nothing
return obj
def _process_base_ops(op: BaseOp):
"""Recursively go through the attrs listed in `attrs_with_pipelineparams`
and sanitize and replace pipeline params with template var string.
Returns a processed `BaseOp`.
NOTE this is an in-place update to `BaseOp`'s attributes (i.e. the ones
specified in `attrs_with_pipelineparams`, all `PipelineParam` are replaced
with the corresponding template variable strings).
Args:
op {BaseOp}: class that inherits from BaseOp
Returns:
BaseOp
"""
# map param's (unsanitized pattern or serialized str pattern) -> input param var str
map_to_tmpl_var = {(param.pattern or str(param)):
'{{inputs.parameters.%s}}' % param.full_name
for param in op.inputs}
# process all attr with pipelineParams except inputs and outputs parameters
for key in op.attrs_with_pipelineparams:
setattr(op, key, _process_obj(getattr(op, key), map_to_tmpl_var))
return op
def _parameters_to_json(params: List[dsl.PipelineParam]):
"""Converts a list of PipelineParam into an argo `parameter` JSON obj."""
_to_json = (lambda param: dict(name=param.full_name, value=param.value)
if param.value else dict(name=param.full_name))
params = [_to_json(param) for param in params]
# Sort to make the results deterministic.
params.sort(key=lambda x: x['name'])
return params
def _inputs_to_json(
inputs_params: List[dsl.PipelineParam],
input_artifact_paths: Dict[str, str] = None,
artifact_arguments: Dict[str, str] = None,
) -> Dict[str, Dict]:
"""Converts a list of PipelineParam into an argo `inputs` JSON obj."""
parameters = _parameters_to_json(inputs_params)
# Building the input artifacts section
artifacts = []
for name, path in (input_artifact_paths or {}).items():
artifact = {'name': name, 'path': path}
if name in artifact_arguments: # The arguments should be compiled as DAG task arguments, not template's default values, but in the current DSL-compiler implementation it's too hard to make that work when passing artifact references.
artifact['raw'] = {'data': str(artifact_arguments[name])}
artifacts.append(artifact)
artifacts.sort(
key=lambda x: x['name']) #Stabilizing the input artifact ordering
inputs_dict = {}
if parameters:
inputs_dict['parameters'] = parameters
if artifacts:
inputs_dict['artifacts'] = artifacts
return inputs_dict
def _outputs_to_json(op: BaseOp, outputs: Dict[str, dsl.PipelineParam],
param_outputs: Dict[str,
str], output_artifacts: List[dict]):
"""Creates an argo `outputs` JSON obj."""
if isinstance(op, dsl.ResourceOp):
value_from_key = "jsonPath"
else:
value_from_key = "path"
output_parameters = []
for param in set(outputs.values()): # set() dedupes output references
output_parameters.append({
'name': param.full_name,
'valueFrom': {
value_from_key: param_outputs[param.name]
}
})
output_parameters.sort(key=lambda x: x['name'])
ret = {}
if output_parameters:
ret['parameters'] = output_parameters
if output_artifacts:
ret['artifacts'] = output_artifacts
return ret
# TODO: generate argo python classes from swagger and use convert_k8s_obj_to_json??
def _op_to_template(op: BaseOp):
"""Generate template given an operator inherited from BaseOp."""
# Display name
if op.display_name:
op.add_pod_annotation('pipelines.kubeflow.org/task_display_name',
op.display_name)
# Caching option
op.add_pod_label('pipelines.kubeflow.org/enable_caching',
str(op.enable_caching).lower())
# NOTE in-place update to BaseOp
# replace all PipelineParams with template var strings
processed_op = _process_base_ops(op)
if isinstance(op, dsl.ContainerOp):
output_artifact_paths = OrderedDict(op.output_artifact_paths)
# This should have been as easy as output_artifact_paths.update(op.file_outputs), but the _outputs_to_json function changes the output names and we must do the same here, so that the names are the same
output_artifact_paths.update(
sorted(((param.full_name, processed_op.file_outputs[param.name])
for param in processed_op.outputs.values()),
key=lambda x: x[0]))
output_artifacts = [{
'name': name,
'path': path
} for name, path in output_artifact_paths.items()]
# workflow template
template = {
'name': processed_op.name,
'container': convert_k8s_obj_to_json(processed_op.container)
}
elif isinstance(op, dsl.ResourceOp):
# no output artifacts
output_artifacts = []
# workflow template
processed_op.resource["manifest"] = yaml.dump(
convert_k8s_obj_to_json(processed_op.k8s_resource),
default_flow_style=False)
template = {
'name': processed_op.name,
'resource': convert_k8s_obj_to_json(processed_op.resource)
}
# inputs
input_artifact_paths = processed_op.input_artifact_paths if isinstance(
processed_op, dsl.ContainerOp) else None
artifact_arguments = processed_op.artifact_arguments if isinstance(
processed_op, dsl.ContainerOp) else None
inputs = _inputs_to_json(processed_op.inputs, input_artifact_paths,
artifact_arguments)
if inputs:
template['inputs'] = inputs
# outputs
if isinstance(op, dsl.ContainerOp):
param_outputs = processed_op.file_outputs
elif isinstance(op, dsl.ResourceOp):
param_outputs = processed_op.attribute_outputs
outputs_dict = _outputs_to_json(op, processed_op.outputs, param_outputs,
output_artifacts)
if outputs_dict:
template['outputs'] = outputs_dict
# pod spec used for runtime container settings
podSpecPatch = {}
# node selector
if processed_op.node_selector:
copy_node_selector = copy.deepcopy(processed_op.node_selector)
for key, value in processed_op.node_selector.items():
if re.match('^{{inputs.parameters.*}}$', key) or re.match(
'^{{inputs.parameters.*}}$', value):
if not 'nodeSelector' in podSpecPatch:
podSpecPatch['nodeSelector'] = {}
podSpecPatch["nodeSelector"][key] = value
del copy_node_selector[
key] # avoid to change the dict when iterating it
if processed_op.node_selector:
template['nodeSelector'] = copy_node_selector
# tolerations
if processed_op.tolerations:
template['tolerations'] = processed_op.tolerations
# affinity
if processed_op.affinity:
template['affinity'] = convert_k8s_obj_to_json(processed_op.affinity)
# metadata
if processed_op.pod_annotations or processed_op.pod_labels:
template['metadata'] = {}
if processed_op.pod_annotations:
template['metadata']['annotations'] = processed_op.pod_annotations
if processed_op.pod_labels:
template['metadata']['labels'] = processed_op.pod_labels
# retries
if processed_op.num_retries or processed_op.retry_policy:
template['retryStrategy'] = {}
if processed_op.num_retries:
template['retryStrategy']['limit'] = processed_op.num_retries
if processed_op.retry_policy:
template['retryStrategy']['retryPolicy'] = processed_op.retry_policy
if not processed_op.num_retries:
warnings.warn('retry_policy is set, but num_retries is not')
backoff_dict = {}
if processed_op.backoff_duration:
backoff_dict['duration'] = processed_op.backoff_duration
if processed_op.backoff_factor:
backoff_dict['factor'] = processed_op.backoff_factor
if processed_op.backoff_max_duration:
backoff_dict['maxDuration'] = processed_op.backoff_max_duration
if backoff_dict:
template['retryStrategy']['backoff'] = backoff_dict
# timeout
if processed_op.timeout:
template['activeDeadlineSeconds'] = processed_op.timeout
# initContainers
if processed_op.init_containers:
template['initContainers'] = processed_op.init_containers
# sidecars
if processed_op.sidecars:
template['sidecars'] = processed_op.sidecars
# volumes
if processed_op.volumes:
template['volumes'] = [
convert_k8s_obj_to_json(volume) for volume in processed_op.volumes
]
template['volumes'].sort(key=lambda x: x['name'])
# Runtime resource requests
if isinstance(op, dsl.ContainerOp) and ('resources' in op.container.keys()):
for setting, val in op.container['resources'].items():
for resource, param in val.items():
if (resource in ['cpu', 'memory', 'amd.com/gpu', 'nvidia.com/gpu'] or re.match('^{{inputs.parameters.*}}$', resource))\
and re.match('^{{inputs.parameters.*}}$', str(param)):
if not 'containers' in podSpecPatch:
podSpecPatch['containers'] = [{
'name': 'main',
'resources': {}
}]
if setting not in podSpecPatch['containers'][0][
'resources']:
podSpecPatch['containers'][0]['resources'][setting] = {
resource: param
}
else:
podSpecPatch['containers'][0]['resources'][setting][
resource] = param
del template['container']['resources'][setting][resource]
if not template['container']['resources'][setting]:
del template['container']['resources'][setting]
if isinstance(op, dsl.ContainerOp) and op._metadata and not op.is_v2:
template.setdefault('metadata', {}).setdefault(
'annotations',
{})['pipelines.kubeflow.org/component_spec'] = json.dumps(
op._metadata.to_dict(), sort_keys=True)
if hasattr(op, '_component_ref'):
template.setdefault('metadata', {}).setdefault(
'annotations',
{})['pipelines.kubeflow.org/component_ref'] = json.dumps(
op._component_ref.to_dict(), sort_keys=True)
if hasattr(op, '_parameter_arguments') and op._parameter_arguments:
template.setdefault('metadata', {}).setdefault(
'annotations',
{})['pipelines.kubeflow.org/arguments.parameters'] = json.dumps(
op._parameter_arguments, sort_keys=True)
if isinstance(op, dsl.ContainerOp) and op.execution_options:
if op.execution_options.caching_strategy.max_cache_staleness:
template.setdefault('metadata', {}).setdefault(
'annotations',
{})['pipelines.kubeflow.org/max_cache_staleness'] = str(
op.execution_options.caching_strategy.max_cache_staleness)
if podSpecPatch:
template['podSpecPatch'] = json.dumps(podSpecPatch)
return template