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

533 lines
20 KiB
Python

# Copyright 2018 Google LLC
#
# 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.
from collections import defaultdict
import copy
import inspect
import kfp.dsl as dsl
import re
import string
import tarfile
import tempfile
import yaml
class Compiler(object):
"""DSL Compiler.
It compiles DSL pipeline functions into workflow yaml. Example usage:
```python
@dsl.pipeline(
name='name',
description='description'
)
def my_pipeline(a: dsl.PipelineParam, b: dsl.PipelineParam):
pass
Compiler().compile(my_pipeline, 'path/to/workflow.yaml')
```
"""
def _sanitize_name(self, name):
return re.sub('-+', '-', re.sub('[^-0-9a-z]+', '-', name.lower())).lstrip('-').rstrip('-') #from _make_kubernetes_name
def _param_full_name(self, param):
if param.op_name:
return param.op_name + '-' + param.name
return self._sanitize_name(param.name)
def _build_conventional_artifact(self, name):
return {
'name': name,
'path': '/' + name + '.json',
's3': {
# TODO: parameterize namespace for minio service
'endpoint': 'minio-service.kubeflow:9000',
'bucket': 'mlpipeline',
'key': 'runs/{{workflow.uid}}/{{pod.name}}/' + name + '.tgz',
'insecure': True,
'accessKeySecret': {
'name': 'mlpipeline-minio-artifact',
'key': 'accesskey',
},
'secretKeySecret': {
'name': 'mlpipeline-minio-artifact',
'key': 'secretkey'
}
},
}
def _op_to_template(self, op):
"""Generate template given an operator inherited from dsl.ContainerOp."""
processed_args = None
if op.arguments:
processed_args = list(map(str, op.arguments))
for i, _ in enumerate(processed_args):
if op.argument_inputs:
for param in op.argument_inputs:
full_name = self._param_full_name(param)
processed_args[i] = re.sub(str(param), '{{inputs.parameters.%s}}' % full_name,
processed_args[i])
input_parameters = []
for param in op.inputs:
one_parameter = {'name': self._param_full_name(param)}
if param.value:
one_parameter['value'] = str(param.value)
input_parameters.append(one_parameter)
# Sort to make the results deterministic.
input_parameters.sort(key=lambda x: x['name'])
output_parameters = []
for param in op.outputs.values():
output_parameters.append({
'name': self._param_full_name(param),
'valueFrom': {'path': op.file_outputs[param.name]}
})
output_parameters.sort(key=lambda x: x['name'])
template = {
'name': op.name,
'container': {
'image': op.image,
}
}
if processed_args:
template['container']['args'] = processed_args
if input_parameters:
template['inputs'] = {'parameters': input_parameters}
template['outputs'] = {}
if output_parameters:
template['outputs'] = {'parameters': output_parameters}
# Generate artifact for metadata output
# The motivation of appending the minio info in the yaml
# is to specify a unique path for the metadata.
# TODO: after argo addresses the issue that configures a unique path
# for the artifact output when default artifact repository is configured,
# this part needs to be updated to use the default artifact repository.
output_artifacts = []
output_artifacts.append(self._build_conventional_artifact('mlpipeline-ui-metadata'))
output_artifacts.append(self._build_conventional_artifact('mlpipeline-metrics'))
template['outputs']['artifacts'] = output_artifacts
if op.command:
template['container']['command'] = op.command
# Set resources.
if op.memory_limit or op.cpu_limit or op.memory_request or op.cpu_request:
template['container']['resources'] = {}
if op.memory_limit or op.cpu_limit:
template['container']['resources']['limits'] = {}
if op.memory_limit:
template['container']['resources']['limits']['memory'] = op.memory_limit
if op.cpu_limit:
template['container']['resources']['limits']['cpu'] = op.cpu_limit
if op.memory_request or op.cpu_request:
template['container']['resources']['requests'] = {}
if op.memory_request:
template['container']['resources']['requests']['memory'] = op.memory_request
if op.cpu_request:
template['container']['resources']['requests']['cpu'] = op.cpu_request
return template
def _get_groups_for_ops(self, root_group):
"""Helper function to get belonging groups for each op.
Each pipeline has a root group. Each group has a list of operators (leaf) and groups.
This function traverse the tree and get all ancestor groups for all operators.
Returns:
A dict. Key is the operator's name. Value is a list of ancestor groups including the
op itself. The list of a given operator is sorted in a way that the farthest
group is the first and operator itself is the last.
"""
def _get_op_groups_helper(current_groups, ops_to_groups):
root_group = current_groups[-1]
for g in root_group.groups:
current_groups.append(g)
_get_op_groups_helper(current_groups, ops_to_groups)
del current_groups[-1]
for op in root_group.ops:
ops_to_groups[op.name] = [x.name for x in current_groups] + [op.name]
ops_to_groups = {}
current_groups = [root_group]
_get_op_groups_helper(current_groups, ops_to_groups)
return ops_to_groups
def _get_groups(self, root_group):
"""Helper function to get all groups (not including ops) in a pipeline."""
def _get_groups_helper(group):
groups = [group]
for g in group.groups:
groups += _get_groups_helper(g)
return groups
return _get_groups_helper(root_group)
def _get_uncommon_ancestors(self, op_groups, op1, op2):
"""Helper function to get unique ancestors between two ops.
For example, op1's ancestor groups are [root, G1, G2, G3, op1], op2's ancestor groups are
[root, G1, G4, op2], then it returns a tuple ([G2, G3, op1], [G4, op2]).
"""
both_groups = [op_groups[op1.name], op_groups[op2.name]]
common_groups_len = sum(1 for x in zip(*both_groups) if x==(x[0],)*len(x))
group1 = op_groups[op1.name][common_groups_len:]
group2 = op_groups[op2.name][common_groups_len:]
return (group1, group2)
def _get_inputs_outputs(self, pipeline, root_group, op_groups):
"""Get inputs and outputs of each group and op.
Returns:
A tuple (inputs, outputs).
inputs and outputs are dicts with key being the group/op names and values being list of
tuples (param_name, producing_op_name). producing_op_name is the name of the op that
produces the param. If the param is a pipeline param (no producer op), then
producing_op_name is None.
"""
condition_params = self._get_condition_params_for_ops(root_group)
inputs = defaultdict(set)
outputs = defaultdict(set)
for op in pipeline.ops.values():
# op's inputs and all params used in conditions for that op are both considered.
for param in op.inputs + list(condition_params[op.name]):
# if the value is already provided (immediate value), then no need to expose
# it as input for its parent groups.
if param.value:
continue
full_name = self._param_full_name(param)
if param.op_name:
upstream_op = pipeline.ops[param.op_name]
upstream_groups, downstream_groups = self._get_uncommon_ancestors(
op_groups, upstream_op, op)
for i, g in enumerate(downstream_groups):
if i == 0:
# If it is the first uncommon downstream group, then the input comes from
# the first uncommon upstream group.
inputs[g].add((full_name, upstream_groups[0]))
else:
# If not the first downstream group, then the input is passed down from
# its ancestor groups so the upstream group is None.
inputs[g].add((full_name, None))
for i, g in enumerate(upstream_groups):
if i == len(upstream_groups) - 1:
# If last upstream group, it is an operator and output comes from container.
outputs[g].add((full_name, None))
else:
# If not last upstream group, output value comes from one of its child.
outputs[g].add((full_name, upstream_groups[i+1]))
else:
if not op.is_exit_handler:
for g in op_groups[op.name]:
inputs[g].add((full_name, None))
return inputs, outputs
def _get_condition_params_for_ops(self, root_group):
"""Get parameters referenced in conditions of ops."""
conditions = defaultdict(set)
def _get_condition_params_for_ops_helper(group, current_conditions_params):
new_current_conditions_params = current_conditions_params
if group.type == 'condition':
new_current_conditions_params = list(current_conditions_params)
if isinstance(group.condition.operand1, dsl.PipelineParam):
new_current_conditions_params.append(group.condition.operand1)
if isinstance(group.condition.operand2, dsl.PipelineParam):
new_current_conditions_params.append(group.condition.operand2)
for op in group.ops:
for param in new_current_conditions_params:
conditions[op.name].add(param)
for g in group.groups:
_get_condition_params_for_ops_helper(g, new_current_conditions_params)
_get_condition_params_for_ops_helper(root_group, [])
return conditions
def _get_dependencies(self, pipeline, root_group, op_groups):
"""Get dependent groups and ops for all ops and groups.
Returns:
A dict. Key is group/op name, value is a list of dependent groups/ops.
The dependencies are calculated in the following way: if op2 depends on op1,
and their ancestors are [root, G1, G2, op1] and [root, G1, G3, G4, op2],
then G3 is dependent on G2. Basically dependency only exists in the first uncommon
ancesters in their ancesters chain. Only sibling groups/ops can have dependencies.
"""
condition_params = self._get_condition_params_for_ops(root_group)
dependencies = defaultdict(set)
for op in pipeline.ops.values():
unstream_op_names = set()
for param in op.inputs + list(condition_params[op.name]):
if param.op_name:
unstream_op_names.add(param.op_name)
unstream_op_names |= set(op.dependent_op_names)
for op_name in unstream_op_names:
upstream_op = pipeline.ops[op_name]
upstream_groups, downstream_groups = self._get_uncommon_ancestors(
op_groups, upstream_op, op)
dependencies[downstream_groups[0]].add(upstream_groups[0])
return dependencies
def _create_condition(self, condition):
left = ('{{inputs.parameters.%s}}' % self._param_full_name(condition.operand1)
if isinstance(condition.operand1, dsl.PipelineParam)
else str(condition.operand1))
right = ('{{inputs.parameters.%s}}' % self._param_full_name(condition.operand2)
if isinstance(condition.operand2, dsl.PipelineParam)
else str(condition.operand2))
return ('%s == %s' % (left, right))
def _group_to_template(self, group, inputs, outputs, dependencies):
"""Generate template given an OpsGroup.
inputs, outputs, dependencies are all helper dicts.
"""
template = {'name': group.name}
# Generate inputs section.
if inputs.get(group.name, None):
template_inputs = [{'name': x[0]} for x in inputs[group.name]]
template_inputs.sort(key=lambda x: x['name'])
template['inputs'] = {
'parameters': template_inputs
}
# Generate outputs section.
if outputs.get(group.name, None):
template_outputs = []
for param_name, depentent_name in outputs[group.name]:
template_outputs.append({
'name': param_name,
'valueFrom': {
'parameter': '{{tasks.%s.outputs.parameters.%s}}' % (depentent_name, param_name)
}
})
template_outputs.sort(key=lambda x: x['name'])
template['outputs'] = {'parameters': template_outputs}
if group.type == 'condition':
# This is a workaround for the fact that argo does not support conditions in DAG mode.
# Basically, we insert an extra group that contains only the original group. The extra group
# operates in "step" mode where condition is supported.
only_child = group.groups[0]
step = {
'name': only_child.name,
'template': only_child.name,
}
if inputs.get(only_child.name, None):
arguments = []
for param_name, dependent_name in inputs[only_child.name]:
arguments.append({
'name': param_name,
'value': '{{inputs.parameters.%s}}' % param_name
})
arguments.sort(key=lambda x: x['name'])
step['arguments'] = {'parameters': arguments}
step['when'] = self._create_condition(group.condition)
template['steps'] = [[step]]
else:
# Generate tasks section.
tasks = []
for sub_group in group.groups + group.ops:
task = {
'name': sub_group.name,
'template': sub_group.name,
}
# Generate dependencies section for this task.
if dependencies.get(sub_group.name, None):
group_dependencies = list(dependencies[sub_group.name])
group_dependencies.sort()
task['dependencies'] = group_dependencies
# Generate arguments section for this task.
if inputs.get(sub_group.name, None):
arguments = []
for param_name, dependent_name in inputs[sub_group.name]:
if dependent_name:
# The value comes from an upstream sibling.
arguments.append({
'name': param_name,
'value': '{{tasks.%s.outputs.parameters.%s}}' % (dependent_name, param_name)
})
else:
# The value comes from its parent.
arguments.append({
'name': param_name,
'value': '{{inputs.parameters.%s}}' % param_name
})
arguments.sort(key=lambda x: x['name'])
task['arguments'] = {'parameters': arguments}
tasks.append(task)
tasks.sort(key=lambda x: x['name'])
template['dag'] = {'tasks': tasks}
return template
def _create_new_groups(self, root_group):
"""Create a copy of the input group, and insert extra groups for conditions."""
new_group = copy.deepcopy(root_group)
def _insert_group_for_condition_helper(group):
for i, g in enumerate(group.groups):
if g.type == 'condition':
child_condition_group = dsl.OpsGroup('condition-child', g.name + '-child')
child_condition_group.ops = g.ops
child_condition_group.groups = g.groups
g.groups = [child_condition_group]
g.ops = list()
_insert_group_for_condition_helper(child_condition_group)
else:
_insert_group_for_condition_helper(g)
_insert_group_for_condition_helper(new_group)
return new_group
def _create_templates(self, pipeline):
"""Create all groups and ops templates in the pipeline."""
# This is needed only because Argo does not support condition in DAG mode.
# Revisit when https://github.com/argoproj/argo/issues/921 is fixed.
new_root_group = self._create_new_groups(pipeline.groups[0])
op_groups = self._get_groups_for_ops(new_root_group)
inputs, outputs = self._get_inputs_outputs(pipeline, new_root_group, op_groups)
dependencies = self._get_dependencies(pipeline, new_root_group, op_groups)
groups = self._get_groups(new_root_group)
templates = []
for g in groups:
templates.append(self._group_to_template(g, inputs, outputs, dependencies))
for op in pipeline.ops.values():
templates.append(self._op_to_template(op))
return templates
def _create_pipeline_workflow(self, args, pipeline):
"""Create workflow for the pipeline."""
input_params = []
for arg in args:
param = {'name': arg.name}
if arg.value is not None:
param['value'] = str(arg.value)
input_params.append(param)
templates = self._create_templates(pipeline)
templates.sort(key=lambda x: x['name'])
exit_handler = None
if pipeline.groups[0].groups:
first_group = pipeline.groups[0].groups[0]
if first_group.type == 'exit_handler':
exit_handler = first_group.exit_op
workflow = {
'apiVersion': 'argoproj.io/v1alpha1',
'kind': 'Workflow',
'metadata': {'generateName': pipeline.name + '-'},
'spec': {
'entrypoint': pipeline.name,
'templates': templates,
'arguments': {'parameters': input_params},
'serviceAccountName': 'pipeline-runner'
}
}
if exit_handler:
workflow['spec']['onExit'] = exit_handler.name
return workflow
def _validate_args(self, argspec):
if argspec.defaults:
for value in argspec.defaults:
if not issubclass(type(value), dsl.PipelineParam):
raise ValueError(
'Default values of argument has to be type dsl.PipelineParam or its child.')
def _validate_exit_handler(self, pipeline):
"""Makes sure there is only one global exit handler.
Note this is a temporary workaround until argo supports local exit handler.
"""
def _validate_exit_handler_helper(group, exiting_op_names, handler_exists):
if group.type == 'exit_handler':
if handler_exists or len(exiting_op_names) > 1:
raise ValueError('Only one global exit_handler is allowed and all ops need to be included.')
handler_exists = True
if group.ops:
exiting_op_names.extend([x.name for x in group.ops])
for g in group.groups:
_validate_exit_handler_helper(g, exiting_op_names, handler_exists)
return _validate_exit_handler_helper(pipeline.groups[0], [], False)
def _compile(self, pipeline_func):
"""Compile the given pipeline function into workflow."""
argspec = inspect.getfullargspec(pipeline_func)
self._validate_args(argspec)
registered_pipeline_functions = dsl.Pipeline.get_pipeline_functions()
if pipeline_func not in registered_pipeline_functions:
raise ValueError('Please use a function with @dsl.pipeline decorator.')
pipeline_name, _ = dsl.Pipeline.get_pipeline_functions()[pipeline_func]
pipeline_name = self._sanitize_name(pipeline_name)
# Create the arg list with no default values and call pipeline function.
args_list = [dsl.PipelineParam(self._sanitize_name(arg_name))
for arg_name in argspec.args]
with dsl.Pipeline(pipeline_name) as p:
pipeline_func(*args_list)
# Remove when argo supports local exit handler.
self._validate_exit_handler(p)
# Fill in the default values.
args_list_with_defaults = [dsl.PipelineParam(self._sanitize_name(arg_name))
for arg_name in argspec.args]
if argspec.defaults:
for arg, default in zip(reversed(args_list_with_defaults), reversed(argspec.defaults)):
arg.value = default.value
workflow = self._create_pipeline_workflow(args_list_with_defaults, p)
return workflow
def compile(self, pipeline_func, package_path):
"""Compile the given pipeline function into workflow yaml.
Args:
pipeline_func: pipeline functions with @dsl.pipeline decorator.
package_path: the output workflow tar.gz file path. for example, "~/a.tar.gz"
"""
workflow = self._compile(pipeline_func)
yaml.Dumper.ignore_aliases = lambda *args : True
with tempfile.NamedTemporaryFile() as tmp:
with open(tmp.name, 'w') as fd:
yaml.dump(workflow, fd, default_flow_style=False)
with tarfile.open(package_path, "w:gz") as tar:
tar.add(tmp.name, arcname="pipeline.yaml")