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

442 lines
19 KiB
Python

# Copyright 2022 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.
"""Utility methods for compiler implementation that is IR-agnostic."""
import collections
from typing import Dict, List, Mapping, Set, Tuple, Union
from kfp.components import for_loop
from kfp.components import pipeline_channel
from kfp.components import pipeline_context
from kfp.components import pipeline_task
from kfp.components import tasks_group
GroupOrTaskType = Union[tasks_group.TasksGroup, pipeline_task.PipelineTask]
def get_all_groups(
root_group: tasks_group.TasksGroup,) -> List[tasks_group.TasksGroup]:
"""Gets all groups (not including tasks) in a pipeline.
Args:
root_group: The root group of a pipeline.
Returns:
A list of all groups in topological order (parent first).
"""
all_groups = []
def _get_all_groups_helper(
group: tasks_group.TasksGroup,
all_groups: List[tasks_group.TasksGroup],
):
all_groups.append(group)
for group in group.groups:
_get_all_groups_helper(group, all_groups)
_get_all_groups_helper(root_group, all_groups)
return all_groups
def get_parent_groups(
root_group: tasks_group.TasksGroup,
) -> Tuple[Mapping[str, List[GroupOrTaskType]], Mapping[str,
List[GroupOrTaskType]]]:
"""Get parent groups that contain the specified tasks.
Each pipeline has a root group. Each group has a list of tasks (leaf)
and groups.
This function traverse the tree and get ancestor groups for all tasks.
Args:
root_group: The root group of a pipeline.
Returns:
A tuple. The first item is a mapping of task names to parent groups,
and second item is a mapping of group names to parent groups.
A list of parent groups is a list of ancestor groups including the
task/group itself. The list is sorted in a way that the farthest
parent group is the first and task/group itself is the last.
"""
def _get_parent_groups_helper(
current_groups: List[tasks_group.TasksGroup],
tasks_to_groups: Dict[str, List[GroupOrTaskType]],
groups_to_groups: Dict[str, List[GroupOrTaskType]],
) -> None:
root_group = current_groups[-1]
for group in root_group.groups:
groups_to_groups[group.name] = [x.name for x in current_groups
] + [group.name]
current_groups.append(group)
_get_parent_groups_helper(
current_groups=current_groups,
tasks_to_groups=tasks_to_groups,
groups_to_groups=groups_to_groups,
)
del current_groups[-1]
for task in root_group.tasks:
tasks_to_groups[task.name] = [x.name for x in current_groups
] + [task.name]
tasks_to_groups = {}
groups_to_groups = {}
current_groups = [root_group]
_get_parent_groups_helper(
current_groups=current_groups,
tasks_to_groups=tasks_to_groups,
groups_to_groups=groups_to_groups,
)
return (tasks_to_groups, groups_to_groups)
# TODO: do we really need this?
def get_condition_channels_for_tasks(
root_group: tasks_group.TasksGroup,
) -> Mapping[str, Set[pipeline_channel.PipelineChannel]]:
"""Gets channels referenced in conditions of tasks' parents.
Args:
root_group: The root group of a pipeline.
Returns:
A mapping of task name to a set of pipeline channels appeared in its
parent dsl.Condition groups.
"""
conditions = collections.defaultdict(set)
def _get_condition_channels_for_tasks_helper(
group,
current_conditions_channels,
):
new_current_conditions_channels = current_conditions_channels
if isinstance(group, tasks_group.Condition):
new_current_conditions_channels = list(current_conditions_channels)
if isinstance(group.condition.left_operand,
pipeline_channel.PipelineChannel):
new_current_conditions_channels.append(
group.condition.left_operand)
if isinstance(group.condition.right_operand,
pipeline_channel.PipelineChannel):
new_current_conditions_channels.append(
group.condition.right_operand)
for task in group.tasks:
for channel in new_current_conditions_channels:
conditions[task.name].add(channel)
for group in group.groups:
_get_condition_channels_for_tasks_helper(
group, new_current_conditions_channels)
_get_condition_channels_for_tasks_helper(root_group, [])
return conditions
def get_inputs_for_all_groups(
pipeline: pipeline_context.Pipeline,
task_name_to_parent_groups: Mapping[str, List[GroupOrTaskType]],
group_name_to_parent_groups: Mapping[str, List[tasks_group.TasksGroup]],
condition_channels: Mapping[str,
Set[pipeline_channel.PipelineParameterChannel]],
name_to_for_loop_group: Mapping[str, tasks_group.ParallelFor],
) -> Mapping[str, List[Tuple[pipeline_channel.PipelineChannel, str]]]:
"""Get inputs and outputs of each group and op.
Args:
pipeline: The instantiated pipeline object.
task_name_to_parent_groups: The dict of task name to list of parent
groups.
group_name_to_parent_groups: The dict of group name to list of
parent groups.
condition_channels: The dict of task name to a set of pipeline
channels referenced by its parent condition groups.
name_to_for_loop_group: The dict of for loop group name to loop
group.
Returns:
A mapping with key being the group/task names and values being list
of tuples (channel, producing_task_name).
producing_task_name is the name of the task that produces the
channel. If the channel is a pipeline argument (no producer task),
then producing_task_name is None.
"""
inputs = collections.defaultdict(set)
for task in pipeline.tasks.values():
# task's inputs and all channels used in conditions for that task are
# considered.
task_condition_inputs = list(condition_channels[task.name])
for channel in task.channel_inputs + task_condition_inputs:
# If the value is already provided (immediate value), then no
# need to expose it as input for its parent groups.
if getattr(channel, 'value', None):
continue
# channels_to_add could be a list of PipelineChannels when loop
# args are involved. Given a nested loops example as follows:
#
# def my_pipeline(loop_parameter: list):
# with dsl.ParallelFor(loop_parameter) as item:
# with dsl.ParallelFor(item.p_a) as item_p_a:
# print_op(item_p_a.q_a)
#
# The print_op takes an input of
# {{channel:task=;name=loop_parameter-loop-item-subvar-p_a-loop-item-subvar-q_a;}}.
# Given this, we calculate the list of PipelineChannels potentially
# needed by across DAG levels as follows:
#
# [{{channel:task=;name=loop_parameter-loop-item-subvar-p_a-loop-item-subvar-q_a}},
# {{channel:task=;name=loop_parameter-loop-item-subvar-p_a-loop-item}},
# {{channel:task=;name=loop_parameter-loop-item-subvar-p_a}},
# {{channel:task=;name=loop_parameter-loop-item}},
# {{chaenel:task=;name=loop_parameter}}]
#
# For the above example, the first loop needs the input of
# {{channel:task=;name=loop_parameter}},
# the second loop needs the input of
# {{channel:task=;name=loop_parameter-loop-item}}
# and the print_op needs the input of
# {{channel:task=;name=loop_parameter-loop-item-subvar-p_a-loop-item}}
#
# When we traverse a DAG in a top-down direction, we add channels
# from the end, and pop it out when it's no longer needed by the
# sub-DAG.
# When we traverse a DAG in a bottom-up direction, we add
# channels from the front, and pop it out when it's no longer
# needed by the parent DAG.
channels_to_add = collections.deque()
channel_to_add = channel
while isinstance(channel_to_add, (
for_loop.LoopArgument,
for_loop.LoopArgumentVariable,
)):
channels_to_add.append(channel_to_add)
if isinstance(channel_to_add, for_loop.LoopArgumentVariable):
channel_to_add = channel_to_add.loop_argument
else:
channel_to_add = channel_to_add.items_or_pipeline_channel
if isinstance(channel_to_add, pipeline_channel.PipelineChannel):
channels_to_add.append(channel_to_add)
if channel.task_name:
# The PipelineChannel is produced by a task.
upstream_task = pipeline.tasks[channel.task_name]
upstream_groups, downstream_groups = (
_get_uncommon_ancestors(
task_name_to_parent_groups=task_name_to_parent_groups,
group_name_to_parent_groups=group_name_to_parent_groups,
task1=upstream_task,
task2=task,
))
for i, group_name 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.
producer_task = 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.
producer_task = None
inputs[group_name].add((channels_to_add[-1], producer_task))
if group_name in name_to_for_loop_group:
loop_group = name_to_for_loop_group[group_name]
# Pop out the last elements from channels_to_add if it
# is found in the current (loop) DAG. Downstreams
# would only need the more specific versions for it.
if channels_to_add[
-1].full_name in loop_group.loop_argument.full_name:
channels_to_add.pop()
if not channels_to_add:
break
else:
# The PipelineChannel is not produced by a task. It's either
# a top-level pipeline input, or a constant value to loop
# items.
# TODO: revisit if this is correct.
if getattr(task, 'is_exit_handler', False):
continue
# For PipelineChannel as a result of constant value used as
# loop items, we have to go from bottom-up because the
# PipelineChannel can be originated from the middle a DAG,
# which is not needed and visible to its parent DAG.
if isinstance(
channel,
(for_loop.LoopArgument, for_loop.LoopArgumentVariable
)) and channel.is_with_items_loop_argument:
for group_name in task_name_to_parent_groups[
task.name][::-1]:
inputs[group_name].add((channels_to_add[0], None))
if group_name in name_to_for_loop_group:
# for example:
# loop_group.loop_argument.name = 'loop-item-param-1'
# channel.name = 'loop-item-param-1-subvar-a'
loop_group = name_to_for_loop_group[group_name]
if channels_to_add[
0].full_name in loop_group.loop_argument.full_name:
channels_to_add.popleft()
if not channels_to_add:
break
else:
# For PipelineChannel from pipeline input, go top-down
# just like we do for PipelineChannel produced by a task.
for group_name in task_name_to_parent_groups[task.name]:
inputs[group_name].add((channels_to_add[-1], None))
if group_name in name_to_for_loop_group:
loop_group = name_to_for_loop_group[group_name]
if channels_to_add[
-1].full_name in loop_group.loop_argument.full_name:
channels_to_add.pop()
if not channels_to_add:
break
return inputs
def _get_uncommon_ancestors(
task_name_to_parent_groups: Mapping[str, List[GroupOrTaskType]],
group_name_to_parent_groups: Mapping[str, List[tasks_group.TasksGroup]],
task1: GroupOrTaskType,
task2: GroupOrTaskType,
) -> Tuple[List[GroupOrTaskType], List[GroupOrTaskType]]:
"""Gets the unique ancestors between two tasks.
For example, task1's ancestor groups are [root, G1, G2, G3, task1],
task2's ancestor groups are [root, G1, G4, task2], then it returns a
tuple ([G2, G3, task1], [G4, task2]).
Args:
task_name_to_parent_groups: The dict of task name to list of parent
groups.
group_name_tor_parent_groups: The dict of group name to list of
parent groups.
task1: One of the two tasks.
task2: The other task.
Returns:
A tuple which are lists of uncommon ancestors for each task.
"""
if task1.name in task_name_to_parent_groups:
task1_groups = task_name_to_parent_groups[task1.name]
elif task1.name in group_name_to_parent_groups:
task1_groups = group_name_to_parent_groups[task1.name]
else:
raise ValueError(task1.name + ' does not exist.')
if task2.name in task_name_to_parent_groups:
task2_groups = task_name_to_parent_groups[task2.name]
elif task2.name in group_name_to_parent_groups:
task2_groups = group_name_to_parent_groups[task2.name]
else:
raise ValueError(task2.name + ' does not exist.')
both_groups = [task1_groups, task2_groups]
common_groups_len = sum(
1 for x in zip(*both_groups) if x == (x[0],) * len(x))
group1 = task1_groups[common_groups_len:]
group2 = task2_groups[common_groups_len:]
return (group1, group2)
def get_dependencies(
pipeline: pipeline_context.Pipeline,
task_name_to_parent_groups: Mapping[str, List[GroupOrTaskType]],
group_name_to_parent_groups: Mapping[str, List[tasks_group.TasksGroup]],
group_name_to_group: Mapping[str, tasks_group.TasksGroup],
condition_channels: Dict[str, pipeline_channel.PipelineChannel],
) -> Mapping[str, List[GroupOrTaskType]]:
"""Gets dependent groups and tasks for all tasks and groups.
Args:
pipeline: The instantiated pipeline object.
task_name_to_parent_groups: The dict of task name to list of parent
groups.
group_name_to_parent_groups: The dict of group name to list of
parent groups.
group_name_to_group: The dict of group name to group.
condition_channels: The dict of task name to a set of pipeline
channels referenced by its parent condition groups.
Returns:
A Mapping where key is group/task name, value is a list of dependent
groups/tasks. The dependencies are calculated in the following way:
if task2 depends on task1, and their ancestors are
[root, G1, G2, task1] and [root, G1, G3, G4, task2], then G3 is
dependent on G2. Basically dependency only exists in the first
uncommon ancesters in their ancesters chain. Only sibling
groups/tasks can have dependencies.
Raises:
RuntimeError: if a task depends on a task inside a condition or loop
group.
"""
dependencies = collections.defaultdict(set)
for task in pipeline.tasks.values():
upstream_task_names = set()
task_condition_inputs = list(condition_channels[task.name])
for channel in task.channel_inputs + task_condition_inputs:
if channel.task_name:
upstream_task_names.add(channel.task_name)
upstream_task_names |= set(task.dependent_tasks)
for upstream_task_name in upstream_task_names:
# the dependent op could be either a BaseOp or an opsgroup
if upstream_task_name in pipeline.tasks:
upstream_task = pipeline.tasks[upstream_task_name]
elif upstream_task_name in group_name_to_group:
upstream_task = group_name_to_group[upstream_task_name]
else:
raise ValueError(
f'Compiler cannot find task: {upstream_task_name}.')
upstream_groups, downstream_groups = _get_uncommon_ancestors(
task_name_to_parent_groups=task_name_to_parent_groups,
group_name_to_parent_groups=group_name_to_parent_groups,
task1=upstream_task,
task2=task,
)
# a task cannot depend on a task created in a for loop group since individual PipelineTask variables are reassigned after each loop iteration
dependent_group = group_name_to_group.get(upstream_groups[0], None)
if isinstance(dependent_group,
(tasks_group.ParallelFor, tasks_group.Condition,
tasks_group.ExitHandler)):
raise RuntimeError(
f'Task {task.name} cannot dependent on any task inside'
f' the group: {upstream_groups[0]}.')
dependencies[downstream_groups[0]].add(upstream_groups[0])
return dependencies