dstandish commented on code in PR #22867: URL: https://github.com/apache/airflow/pull/22867#discussion_r846488139
########## docs/apache-airflow/concepts/dynamic-task-mapping.rst: ########## @@ -0,0 +1,256 @@ + .. Licensed to the Apache Software Foundation (ASF) under one + or more contributor license agreements. See the NOTICE file + distributed with this work for additional information + regarding copyright ownership. The ASF licenses this file + to you 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. + +==================== +Dynamic Task Mapping +==================== + +Dynamic Task Mapping allows a way for a workflow to create a number of tasks at runtime based upon current data, rather than the DAG author having to know in advance how many tasks would be needed. + +This can be thought of as defining your tasks in a for loop, but instead of having to have the DAG file fetch the data and do that itself, the scheduler can do this based on the output of a previous task. Right before a mapped task is executed the scheduler will create *n* copies of the task, one for each input. + +It is also possible to have a task operate on the collected output of a mapped task, commonly known as map and reduce. + +Simple mapping +============== + +In its simplest form you can map over a list defined directly in your DAG file using the ``expand()`` function instead of calling your task directly. + +.. code-block:: python + + from airflow import DAG + from airflow.decorators import task + + + with DAG(dag_id="simple_mapping", start_date=datetime(2022, 3, 4)): + + @task + def add_one(x: int): + return x + 1 + + @task + def sum_it(values): + total = sum(values) + print(f"Total was {total}") + + added_values = add_one.expand(x=[1, 2, 3]) + sum_it(added_values) + +This will show ``Total was 9`` in the task logs when executed. + +.. note:: A reduce task is not required. + + Although we show a "reduce" task here (``sum_it``) you don't have to have one, the mapped tasks will still be executed even if they have no downstream tasks. + +Repeated Mapping +================ + +The result of one mapped task can also be used as input to the next mapped task + +.. code-block:: python + + from airflow import DAG + from airflow.decorators import task + + + with DAG(dag_id="repeated_mapping", start_date=datetime(2022, 3, 4)): + + @task + def add_one(x: int): + return x + 1 + + first = add_one.expand(x=[1, 2, 3]) + second = add_one.expand(x=first) + +This would have a result of [3, 4, 5] + +Constant parameters +=================== + +As well as passing arguments that get expanded at run-time, it is possible to pass arguments that don't change – in order to clearly differentiate between the two kinds we use different functions, ``expand()`` for mapped arguments, and ``partial()`` for unmapped ones. + +For example: + +.. code-block:: python + + @task + def add(x: int, y: int): + return x + y + + + added_values = add.partial(y=10).expand(x=[1, 2, 3]) + # This results in add function being expanded to + # add(x=1,y=10) + # add(x=2,y=10) + # add(x=3,y=10) + +This would result in values of 11, 12, 13. + +This is also useful for passing things such as connection IDs, database table names, or bucket names to tasks. + +Mapping over multiple parameters +================================ + +As well as a single parameter it is possible to pass multiple parameters to expand. This will have the effect of creating a "cross product" effect, calling the mapped task with each combination of parameters. + +For example: + +.. code-block:: python + + @task + def add(x: int, y: int): + return x + y + + + added_values = add.expand(x=[2, 4, 8], y=[5, 10]) + # This results in the add function being called with + # add(x=2, y=5) + # add(x=2, y=10) + # add(x=4, y=5) + # add(x=4, y=10) + # add(x=8, y=5) + # add(x=8, y=10) + +This would result in the add task being called 6 times. Please note however that the order of expansion is not guaranteed. + +It is not possible to achieve an effect similar to Python's zip function with mapped arguments. + +Task-generated Mapping +====================== + +Up until now the examples we've shown could all be achieved with a ``for`` loop in the DAG file, but the real power of dynamic task mapping comes from being able to have a task generate the list to iterate over. For example: + +.. code-block:: python + + @task + def make_list(): + # This can also be from an API call, checking a database, -- almost anything you like, as long as the + # resulting list/dictionary can be stored in the current XCom backend. + return [1, 2, {"a": "b"}, "str"] + + + @task + def consumer(arg): + print(repr(arg)) + + + with DAG(dag_id="dynamic-map", start_date=datetime(2022, 4, 2)) as dag: + consumer.expand(arg=make_list()) + +The make_list runs as a normal task and must return a list or dict (see `What data types can be expanded?`_), and then the consumer task will be called four times, once with each value in the return of make_list. + +Mapping with non-TaskFlow operators +=================================== + +It is possible to use partial and expand with classic style operators as well. Some arguments are not mappable, such as task_id, queue, pool, and most other arguments to BaseOperator and have to be passed to ``partial()`` + + +.. code-block:: python + + BashOperator.partial(task_id="bash", do_xcom_push=False).expand( + bash_command=["echo 1", "echo 2"] + ) + +Mapping over result of classic operators +---------------------------------------- + +If you want to map over the result of a classic operator you will need to create an XComArg object manually: + +.. code-block:: python + + from airflow import XComArg + + task = MyOperator(task_id="source") + + downstream = MyOperator2.partial(task_id="consumer").expand(input=XComArg(task)) + +Putting it all together +======================= + +In this example you have a regular data delivery to an S3 bucket and want to apply the same processing to every file that arrives, no matter how many arrive each time. + +.. code-block:: python + + from datetime import datetime + + from airflow import DAG, XComArg + from airflow.decorators import task + from airflow.providers.amazon.aws.hooks.s3 import S3Hook + from airflow.providers.amazon.aws.operators.s3 import S3ListOperator + + + with DAG(dag_id="mapped_s3", start_date=datetime(2020, 4, 7)) as dag: + files = S3ListOperator( + task_id="get_input", + bucket="example-bucket", + prefix='incoming/provider_a/{{ data_interval_start.strftime("%Y-%m-%d") }}', + ) + + @task + def count_lines(aws_conn_id, bucket, file): + hook = S3Hook(aws_conn_id=aws_conn_id) + + return len(hook.read_key(file, bucket).splitlines()) + + @task + def total(lines): + return sum(lines) + + counts = count_lines.partial(aws_conn_id="aws_default", bucket=files.bucket).expand( + file=XComArg(files) + ) + total(lines=counts) + +What data types can be expanded? +================================ + +Currently it is only possible to map against a dict, a list, or one of those types stored in XCom as the result of a task. + +If an upstream task returns an unmappable type, the mapped task will fail at run-time with an ``UnmappableXComTypePushed`` exception. For instance, you can't have the upstream task return a plain string – it must be a list or a dict. + +Placing limits on mapped tasks +============================== + +There are two limits that you can place on a task: how many mapped task instances can be created as the result of expansion, and how many of the mapped task can run at once Review Comment: when i say "parent" i am referring to the mapped operator instance essentially.... which is itself never gonna be a task ... the tasks that get created are the instances that operato on each input. and that's what you mean by `how many of the mapped task can run at once` no? like the instances -- not the "parent". as to where, there are a lot of places where it's a little ambiguous (though not necessarily insufficiently clear). e.g. L24 and L26. L24 says "before a mapped task is executed the scheduler will create N copies of the task" so here you are referring to mapped task as "parent". and same on L26 it is mapped task as parent: "collective output of a mapped task" meaning the collection of mapped _TIs_. but when you refer "mapped task" in the line referenced above (L228) you are referring to the TIs and that's the ambiguity i'm talking about. not big deal just a suggestion. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
