Run tasks concurrently or in parallel
Learn how to use task runners for concurrent, parallel or distributed execution of tasks.
Task runners are not required for task execution. Calling a task function directly executes the function in the main thread by default, blocking execution of its caller until the task completes.
To enable concurrent, parallel, or distributed execution of tasks, use the .submit()
or .map()
methods to submit tasks to a task runner.
The default task runner in Prefect is the ThreadPoolTaskRunner
,
which runs tasks concurrently in independent threads.
For truly parallel or distributed task execution, you must use one of the following task runners, which are available as extras of the prefect
library:
DaskTaskRunner
can run tasks usingdask.distributed
(installprefect[dask]
)RayTaskRunner
can run tasks using Ray (installprefect[ray]
)
Concurrency vs. parallelism
- Concurrency refers to a system that can do more than one thing simultaneously, but not at the exact same time. Think of concurrent execution as non-blocking: within the restrictions of resources available in the execution environment and data dependencies between tasks, execution of one task does not block execution of other tasks in a flow.
- Parallelism refers to a system that can do more than one thing at the exact same time. Within the restrictions of resources available, parallel execution can run tasks at the same time, such as for operations mapped across a dataset.
Configure a task runner
To configure a specific task runner, provide a task_runner
keyword argument to the parent flow:
The max_workers
parameter of the ThreadPoolTaskRunner
controls the number of threads that the task runner will use to execute tasks concurrently.
Submit tasks to a task runner
When you use .submit()
to submit a task to a task runner, the task runner creates a
PrefectFuture
for access to the state and
result of the task.
A PrefectFuture
is an object that provides:
- a reference to the result returned by the task
- a
State
indicating the current state of the task run
PrefectFuture
objects must be resolved explicitly before returning from a flow or task.
Dependencies between futures will be automatically resolved whenever their dependents are resolved.
This means that only terminal futures need to be resolved, either by:
- returning the terminal futures from your flow or task
- calling
.wait()
or.result()
on each terminal future - using one of the top level
wait
oras_completed
utilities to resolve terminal futures
Not doing so may leave your tasks in an unfinished state.
When you pass a future into a task, Prefect automatically waits for the “upstream” task (the one that the future references), to reach a final state before starting the downstream task.
This means that the downstream task won’t receive the PrefectFuture
you passed as an argument.
Instead, the downstream task receives the value that the upstream task returned.
For example:
If we run this, we see that we only had to wait for the final print_result
future as Prefect automatically resolved say_hello
to a string.
Access results from submitted tasks
You can access the result of a future explicitly with the .result()
method.
The .result()
method waits for the task to complete before returning the result to the caller.
If the task run fails, .result()
will raise the task run’s exception. Disable this behavior
with the raise_on_failure
option:
A few notes on .result()
.result()
is a blocking call. This means that calling.result()
will block the caller until the task run completes.- Only use
.result()
when you need to interact directly with the return value of your submitted task; for example, you should use.result()
if passing the return value to a standard Python function (not a Prefect task) but do not need to use.result()
if you are passing the value to another task (as these futures will be automatically resolved).
Creating state dependencies
You may also use the wait_for=[]
parameter
when calling a task by specifying upstream task dependencies. This enables you to control task execution
order for tasks that do not share data dependencies.
Mapping over iterables
Prefect provides a .map()
method that automatically submits a new task run for each element of its
input data.
This can be useful when submitting a lot of work to a task runner simultaneously.
wait_for
can also be used with .map()
Similar to the .submit()
method, the .map()
method accepts a wait_for
argument to establish state dependencies between mapped tasks that do not share data dependencies.
Simple mapping
Using the unmapped
annotation
Sometimes you may not want to map a task over a certain input value.
By default, non-iterable values will not be mapped over (so unmapped
is not required):
… but if your argument is an iterable type, wrap it with unmapped
to tell .map
to treat it
as static:
Bulk PrefectFuture
operations
When using .map
as in the above example, the result of the task is a list of futures.
You can wait for or retrieve the results from these futures with wait
or result
methods:
which is syntactic sugar for the corresponding list comprehension:
Nested mapped tasks
To model more complex concurrent workflows, you can map tasks within other tasks:
This pattern is useful when you need to:
- Process combinations of parameters concurrently
- Apply multiple transformations to multiple datasets
- Create a grid of operations where each cell is an independent task
Real-world applications
Mapped tasks are particularly valuable in common data science and ETL workflows such as:
- Machine learning model evaluation: Train multiple models on multiple datasets concurrently
- ETL pipelines: Process multiple data sources with multiple transformations
- API data enrichment: Enrich multiple records with data from multiple external services
For example, imagine you want to find the best training configuration for a series of datasets, and you want to process all datasets concurrently:
Use multiple task runners
Each flow can only have one task runner, but sometimes you may want a subset of your tasks to run using a different task runner than the one configured on the flow. In this case, you can create nested flows for tasks that need to use a different task runner.
For example, you can have a flow (in the example below called multiple_runner_flow
) that runs its tasks locally using the ThreadPoolTaskRunner
.
If you have some tasks that can run more efficiently in parallel on a Dask cluster, you can create a nested flow (such as dask_nested_flow
) to run those tasks using the DaskTaskRunner
.
Design considerations
When choosing how and when to achieve concurrency using task runners, consider the following:
-
Task granularity: The “proper” size for tasks depends on the nature and complexity of the work you’re doing, e.g. too many small tasks might create overhead - see Writing tasks for more.
-
Resource constraints: Be aware of environment limitations. Using
.map
can create many task instances very quickly, which might exceed your resource availability. -
Data transfer: Large data passed between tasks can impact performance. Consider passing references to external storage when dealing with large datasets.
-
Parallelism: For true parallelism (rather than just concurrency), consider using a specialized task runner like the
DaskTaskRunner
orRayTaskRunner
(or propose a new task runner type). -
Beware of unsafe global state: Use of concurrency or parallelism features like
.submit
and.map
must respect the underlying primitives to avoid unexpected behavior. For example, the defaultThreadPoolTaskRunner
runs each task in a separate thread, which means that any global state must be threadsafe. Similarly,DaskTaskRunner
andRayTaskRunner
are multi-process runners that require global state to be picklable.
Was this page helpful?