Nested parallelism

For exceptionally large or complicated workflows that can’t be adequately implemented as dynamic workflows or map tasks, it can be beneficial to have multiple levels of workflow parallelization.

This is useful for multiple reasons:

• Better code organization
• Better code reuse
• Better testing
• Better debugging
• Better monitoring, since each subworkflow can be run independently and monitored independently
• Better performance and scale, since each subworkflow is executed as a separate workflow and thus can be distributed among different propeller workers and shards. This allows for better parallelism and scale.

Nested dynamic workflows

You can use nested dynamic workflows to break down a large workflow into smaller workflows and then compose them together to form a hierarchy. In this example, a top-level workflow uses two levels of dynamic workflows to process a list through some simple addition tasks and then flatten the list again.

Example code

"""
A core workflow parallelized as six items with a chunk size of two will be structured as follows:

multi_wf -> level1 -> level2 -> core_wf -> step1 -> step2
                             -> core_wf -> step1 -> step2
                      level2 -> core_wf -> step1 -> step2
                             -> core_wf -> step1 -> step2
                      level2 -> core_wf -> step1 -> step2
                             -> core_wf -> step1 -> step2
"""

import union


@union.task
def step1(a: int) -> int:
    return a + 1


@union.task
def step2(a: int) -> int:
    return a + 2


@union.workflow
def core_wf(a: int) -> int:
    return step2(a=step1(a=a))


core_wf_lp = union.LaunchPlan.get_or_create(core_wf)


@union.dynamic
def level2(l: list[int]) -> list[int]:
    return [core_wf_lp(a=a) for a in l]


@union.task
def reduce(l: list[list[int]]) -> list[int]:
    f = []
    for i in l:
        f.extend(i)
    return f


@union.dynamic
def level1(l: list[int], chunk: int) -> list[int]:
    v = []
    for i in range(0, len(l), chunk):
        v.append(level2(l=l[i:i + chunk]))
    return reduce(l=v)


@union.workflow
def multi_wf(l: list[int], chunk: int) -> list[int]:
    return level1(l=l, chunk=chunk)

Overrides let you add additional arguments to the launch plan you are looping over in the dynamic. Here we add caching:

@union.task
def increment(num: int) -> int:
    return num + 1

@union.workflow
def child(num: int) -> int:
    return increment(num=num)

child_lp = union.LaunchPlan.get_or_create(child)

@union.dynamic
def spawn(n: int) -> list[int]:
    l = []
    for i in [1,2,3,4,5]:
        l.append(child_lp(num=i).with_overrides(cache=True, cache_version="1.0.0"))

    # you can also pass l to another task if you want
    return l

Mixed parallelism

This example is similar to nested dynamic workflows, but instead of using a dynamic workflow to parallelize a core workflow with serial tasks, we use a core workflow to call a map task, which processes both inputs in parallel. This workflow has one less layer of parallelism, so the outputs won’t be the same as those of the nested parallelization example, but it does still demonstrate how you can mix these different approaches to achieve concurrency.

Example code

"""

A core workflow parallelized as six items with a chunk size of two will be structured as follows:
multi_wf -> level1 -> level2 -> mappable
                             -> mappable
                      level2 -> mappable
                             -> mappable
                      level2 -> mappable
                             -> mappable
"""

import union


@union.task
def mappable(a: int) -> int:
    return a + 2


@union.workflow
def level2(l: list[int]) -> list[int]:
    return union.map(mappable)(a=l)


@union.task
def reduce(l: list[list[int]]) -> list[int]:
    f = []
    for i in l:
        f.extend(i)
    return f


@union.dynamic
def level1(l: list[int], chunk: int) -> list[int]:
    v = []
    for i in range(0, len(l), chunk):
        v.append(level2(l=l[i : i + chunk]))
    return reduce(l=v)


@union.workflow
def multi_wf(l: list[int], chunk: int) -> list[int]:
    return level1(l=l, chunk=chunk)

Design considerations

While you can nest even further if needed, or incorporate map tasks if your inputs are all the same type, the design of your workflow should be informed by the actual data you’re processing. For example, if you have a big library of music from which you’d like to extract the lyrics, the first level could loop through all the albums, and the second level could process each song.

If you’re just processing an enormous list of the same input, it’s best to keep your code simple and let the scheduler handle optimizing the execution. Additionally, unless you need dynamic workflow features like mixing and matching inputs and outputs, it’s usually most efficient to use a map task, which has the added benefit of keeping the UI clean.

You can also choose to limit the scale of parallel execution at a few levels. The max_parallelism attribute can be applied at the workflow level and will limit the number of parallel tasks being executed. (This is set to 25 by default.) Within map tasks, you can specify a concurrency argument, which will limit the number of mapped tasks that can run in parallel at any given time.