airflow-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Xinbin Huang <bin.huan...@gmail.com>
Subject Re: [AIP-34] Rewrite SubDagOperator
Date Thu, 18 Jun 2020 00:04:06 GMT
Thank you, Max, Kaxil, and everyone's feedback!

I have rethought about the concept of subdag and task groups. I think the
better way to approach this is to entirely remove subdag and introduce the
concept of TaskGroup, which is a container of tasks along with their
dependencies *without execution/scheduling logic as a DAG*. The only
purpose of it is to group a list of tasks, but you still need to add it to
a DAG for execution.

Here is a small code snippet.

```
class TaskGroup:
    """
    A TaskGroup contains a group of tasks.

    If default_args is missing, it will take default args from the DAG.
    """
    def __init__(self, group_id, default_args):
        pass


"""
You can add tasks to a task group similar to adding tasks to a DAG

This can be declared in a separate file from the dag file
"""
download_group = TaskGroup(group_id='download', default_args=default_args)
download_group.add_task(task1)
task2.dag = download_group

with download_group:
    task3 = DummyOperator(task_id='task3')

    [task, task2] >> task3


"""Add it to a DAG for execution"""
with DAG(dag_id='start_download_dag', default_args=default_args,
schedule_interval='@daily', ...) as dag:
    start = DummyOperator(task_id='start')
    start >> download_group
    # this is equivalent to
    # start >> [task, task2] >> task3
```

With this, we can still reuse a group of tasks and set dependencies between
them; it avoids the boilerplate code from using SubDagOperator, and we can
declare dependencies as `task >> task_group >> task`.

User migration wise, we can introduce it before Airflow 2.0 and allow
gradual transition. Then we can decide if we still want to keep the
SubDagOperator or simply remove it.

Any thoughts?

Cheers,
Bin


On Wed, Jun 17, 2020 at 7:37 AM Maxime Beauchemin <
maximebeauchemin@gmail.com> wrote:

> +1, proposal looks good.
>
> The original intention was really to have tasks groups and a zoom-in/out in
> the UI. The original reasoning was to reuse the DAG object since it is a
> group of tasks, but as highlighted here it does create underlying
> confusions since a DAG is much more than just a group of tasks.
>
> Max
>
> On Mon, Jun 15, 2020 at 2:43 AM Poornima Joshi <joshipoornima06@gmail.com>
> wrote:
>
> > Thank you for your email.
> >
> > On Sat, Jun 13, 2020 at 12:18 AM Xinbin Huang <bin.huangxb@gmail.com>
> > wrote:
> >
> > > > >   - *Unpack SubDags during dag parsing*: This rewrites the
> > > *DagBag.bag_dag*
> > > > >   method to unpack subdag while parsing, and it will give a flat
> > > > > structure at
> > > > >   the task level
> > > >
> > > > The serialized_dag representation already does this I think. At least
> > if
> > > > I've understood your idea here correctly.
> > >
> > > I am not sure about serialized_dag representation, but at least it will
> > > still keep the subdag entry in the DAG table? In my proposal as also in
> > the
> > > draft PR, the idea is to *extract the tasks from the subdag and add
> them
> > > back to the root_dag. *So the runtime DAG graph will look exactly the
> > > same as without subdag but with metadata attached to those sections.
> > These
> > > metadata will be later on used to render in the UI. So after parsing (
> > > *DagBag.process_file()*), it will just output the *root_dag *instead of
> > *root_dag +
> > > subdag + subdag + nested subdag* etc.
> > >
> > >    - e.g. section-1-* will have metadata current_group=section-1,
> > >    parent_group=<the-root-dag-id> (welcome for naming suggestions), the
> > >    reason for parent_group is that we can have nested group and still
> be
> > >    able to capture the dependency.
> > >
> > > Runtime DAG:
> > > [image: image.png]
> > >
> > > While at the UI, what we see would be something like this by utilizing
> > the
> > > metadata, and then we can expand or zoom into in some way.
> > > [image: image.png]
> > >
> > > The benefits I can see is that:
> > > 1. We don't need to deal with the extra complexity of SubDag for
> > execution
> > > and scheduling. It will be the same as not using SubDag.
> > > 2. Still have the benefits of modularized and reusable dag code and
> > > declare dependencies between them. And with the new SubDagOperator (see
> > AIP
> > > or draft PR), we can use the same dag_factory function for generating 1
> > > dag, a lot of dynamic dags, or used for SubDag (in this case, it will
> > just
> > > extract all underlying tasks and append to the root dag).
> > >
> > >    - Then it gets to the idea of replacing subdag with a simpler
> concept
> > >    by Ash:  the proposed change basically drains out the contents of a
> > SubDag
> > >    and becomes more like ExtractSubdagTasksAndAppendToRootdagOperator
> > (forgive
> > >    me about the crazy name..). In this case, it is still necessary to
> > keep the
> > >    concept of subdag as it is nothing more than a name?
> > >
> > > That's why the TaskGroup idea comes up. Thanks Chris Palmer for helping
> > > conceptualize the functionality of TaskGroup, I will just paste it
> here.
> > >
> > > >   - Tasks can be added to a TaskGroup
> > > >  - You *can* have dependencies between Tasks in the same TaskGroup,
> but
> > > >   *cannot* have dependencies between a Task in a TaskGroup and
> either a
> > > >   Task in a different TaskGroup or a Task not in any group
> > > >   - You *can* have dependencies between a TaskGroup and either other
> > > >   TaskGroups or Tasks not in any group
> > > >   - The UI will by default render a TaskGroup as a single "object",
> but
> > > >   which you expand or zoom into in some way
> > > >   - You'd need some way to determine what the "status" of a TaskGroup
> > was
> > > >   at least for UI display purposes
> > >
> > > I agree with Chris:
> > > - From the backend's view (scheduler & executor), I think TaskGroup
> > should
> > > be ignored during execution. (unless we decide to implement some
> metadata
> > > operations that allows start/stop a group of tasks etc.)
> > > - From the UI's View, it should be able to pick up the individual
> tasks'
> > > status and then determine the TaskGroup's status
> > >
> > > Bin
> > >
> > > On Fri, Jun 12, 2020 at 10:28 AM Daniel Imberman <
> > > daniel.imberman@gmail.com> wrote:
> > >
> > >> I hadn’t thought about using the `>>` operator to tie dags together
> but
> > I
> > >> think that sounds pretty great! I wonder if we could essentially write
> > in
> > >> the ability to set dependencies to all starter-tasks for that DAG.
> > >>
> > >> I’m personally ok with SubDag being a mostly UI concept. It doesn’t
> need
> > >> to execute separately, you’re just adding more tasks to the queue that
> > will
> > >> be executed when there are resources available.
> > >>
> > >> via Newton Mail [
> > >>
> >
> https://cloudmagic.com/k/d/mailapp?ct=dx&cv=10.0.50&pv=10.14.6&source=email_footer_2
> > >> ]
> > >> On Fri, Jun 12, 2020 at 9:45 AM, Chris Palmer <chris@crpalmer.com>
> > wrote:
> > >> I agree that SubDAGs are an overly complex abstraction. I think what
> is
> > >> needed/useful is a TaskGroup concept. On a high level I think you want
> > >> this
> > >> functionality:
> > >>
> > >> - Tasks can be added to a TaskGroup
> > >> - You *can* have dependencies between Tasks in the same TaskGroup, but
> > >> *cannot* have dependencies between a Task in a TaskGroup and either a
> > >> Task in a different TaskGroup or a Task not in any group
> > >> - You *can* have dependencies between a TaskGroup and either other
> > >> TaskGroups or Tasks not in any group
> > >> - The UI will by default render a TaskGroup as a single "object", but
> > >> which you expand or zoom into in some way
> > >> - You'd need some way to determine what the "status" of a TaskGroup
> was
> > >> at least for UI display purposes
> > >>
> > >> Not sure if it would need to be a top level object with its own
> database
> > >> table and model or just another attribute on tasks. I think you could
> > >> build
> > >> it in a way such that from the schedulers point of view a DAG with
> > >> TaskGroups doesn't get treated any differently. So it really just
> > becomes
> > >> a
> > >> shortcut for setting dependencies between sets of Tasks, and allows
> the
> > UI
> > >> to simplify the render of the DAG structure.
> > >>
> > >> Chris
> > >>
> > >> On Fri, Jun 12, 2020 at 12:12 PM Dan Davydov
> > <ddavydov@twitter.com.invalid
> > >> >
> > >> wrote:
> > >>
> > >> > Agree with James (and think it's actually the more important issue
> to
> > >> fix),
> > >> > but I am still convinced Ash' idea is the right way forward (just
it
> > >> might
> > >> > require a bit more work to deprecate than adding visual grouping in
> > the
> > >> > UI).
> > >> >
> > >> > There was a previous thread about this FYI with more context on why
> > >> subdags
> > >> > are bad and potential solutions:
> > >> > https://www.mail-archive.com/dev@airflow.apache.org/msg01202.html
> . A
> > >> > solution I outline there to Jame's problem is e.g. enabling the >>
> > >> operator
> > >> > for Airflow operators to work with DAGs as well. I see this being
> > >> separate
> > >> > from Ash' solution for DAG grouping in the UI but one of the two
> items
> > >> > required to replace all existing subdag functionality.
> > >> >
> > >> > I've been working with subdags for 3 years and they are always a
> giant
> > >> pain
> > >> > to use. They are a constant source of user confusion and breakages
> > >> during
> > >> > upgrades. Would love to see them gone :).
> > >> >
> > >> > On Fri, Jun 12, 2020 at 11:11 AM James Coder <jcoder01@gmail.com>
> > >> wrote:
> > >> >
> > >> > > I'm not sure I totally agree it's just a UI concept. I use the
> > subdag
> > >> > > operator to simplify dependencies too. If you have a group of
> tasks
> > >> that
> > >> > > need to finish before another group of tasks start, using a subdag
> > is
> > >> a
> > >> > > pretty quick way to set those dependencies and I think also make
> it
> > >> > easier
> > >> > > to follow the dag code.
> > >> > >
> > >> > > On Fri, Jun 12, 2020 at 9:53 AM Kyle Hamlin <hamlin.kn@gmail.com>
> > >> wrote:
> > >> > >
> > >> > > > I second Ash’s grouping concept.
> > >> > > >
> > >> > > > On Fri, Jun 12, 2020 at 5:10 AM Ash Berlin-Taylor <
> ash@apache.org
> > >
> > >> > > wrote:
> > >> > > >
> > >> > > > > Question:
> > >> > > > >
> > >> > > > > Do we even need the SubDagOperator anymore?
> > >> > > > >
> > >> > > > > Would removing it entirely and just replacing it with
a UI
> > >> grouping
> > >> > > > > concept be conceptually simpler, less to get wrong,
and closer
> > to
> > >> > what
> > >> > > > > users actually want to achieve with subdags?
> > >> > > > >
> > >> > > > > With your proposed change, tasks in subdags could start
> running
> > in
> > >> > > > > parallel (a good change) -- so should we not also just
> > _enitrely_
> > >> > > remove
> > >> > > > > the concept of a sub dag and replace it with something
> simpler.
> > >> > > > >
> > >> > > > > Problems with subdags (I think. I haven't used them
> extensively
> > so
> > >> > may
> > >> > > > > be wrong on some of these):
> > >> > > > > - They need their own dag_id, but it has(?) to be of
the form
> > >> > > > > `parent_dag_id.subdag_id`.
> > >> > > > > - They need their own schedule_interval, but it has
to match
> the
> > >> > parent
> > >> > > > dag
> > >> > > > > - Sub dags can be paused on their own. (Does it make
sense to
> do
> > >> > this?
> > >> > > > > Pausing just a sub dag would mean the sub dag would
never
> > >> execute, so
> > >> > > > > the SubDagOperator would fail too.
> > >> > > > > - You had to choose the executor to operator a subdag
with --
> > >> always
> > >> > a
> > >> > > > > bit of a kludge.
> > >> > > > >
> > >> > > > > Thoughts?
> > >> > > > >
> > >> > > > > -ash
> > >> > > > >
> > >> > > > > On Jun 12 2020, at 12:01 pm, Ash Berlin-Taylor <
> ash@apache.org>
> > >> > wrote:
> > >> > > > >
> > >> > > > > > Workon sub-dags is much needed, I'm excited to
see how this
> > >> > > progresses.
> > >> > > > > >
> > >> > > > > >
> > >> > > > > >> - *Unpack SubDags during dag parsing*: This
rewrites the
> > >> > > > > *DagBag.bag_dag*
> > >> > > > > >> method to unpack subdag while parsing, and
it will give a
> > flat
> > >> > > > > >> structure at
> > >> > > > > >> the task level
> > >> > > > > >
> > >> > > > > > The serialized_dag representation already does
this I think.
> > At
> > >> > least
> > >> > > > if
> > >> > > > > > I've understood your idea here correctly.
> > >> > > > > >
> > >> > > > > > -ash
> > >> > > > > >
> > >> > > > > >
> > >> > > > > > On Jun 12 2020, at 9:51 am, Xinbin Huang <
> > bin.huangxb@gmail.com
> > >> >
> > >> > > > wrote:
> > >> > > > > >
> > >> > > > > >> Hi everyone,
> > >> > > > > >>
> > >> > > > > >> Sending a message to everyone and collect
feedback on the
> > >> AIP-34
> > >> > on
> > >> > > > > >> rewriting SubDagOperator. This was previously
briefly
> > >> mentioned in
> > >> > > the
> > >> > > > > >> discussion about what needs to be done for
Airflow 2.0, and
> > >> one of
> > >> > > the
> > >> > > > > >> ideas is to make SubDagOperator attach tasks
back to the
> root
> > >> DAG.
> > >> > > > > >>
> > >> > > > > >> This AIP-34 focuses on solving SubDagOperator
related
> issues
> > by
> > >> > > > > reattaching
> > >> > > > > >> all tasks back to the root dag while respecting
> dependencies
> > >> > during
> > >> > > > > >> parsing. The original grouping effect on the
UI will be
> > >> achieved
> > >> > > > through
> > >> > > > > >> grouping related tasks by metadata.
> > >> > > > > >>
> > >> > > > > >> This also makes the dag_factory function more
reusable
> > because
> > >> you
> > >> > > > don't
> > >> > > > > >> need to have parent_dag_name and child_dag_name
in the
> > function
> > >> > > > > signature
> > >> > > > > >> anymore.
> > >> > > > > >>
> > >> > > > > >> Changes proposed:
> > >> > > > > >>
> > >> > > > > >> - *Unpack SubDags during dag parsing*: This
rewrites the
> > >> > > > > *DagBag.bag_dag*
> > >> > > > > >> method to unpack subdag while parsing, and
it will give a
> > flat
> > >> > > > > >> structure at
> > >> > > > > >> the task level
> > >> > > > > >> - *Simplify SubDagOperator*: The new SubDagOperator
acts
> > like a
> > >> > > > > >> container and most of the original methods
are removed. The
> > >> > > > > >> signature is
> > >> > > > > >> also changed to *subdag_factory *with *subdag_args
*and
> > >> > > > > *subdag_kwargs*.
> > >> > > > > >> This is similar to the PythonOperator signature.
> > >> > > > > >> - *Add a TaskGroup model and add current_group
&
> parent_group
> > >> > > > > attributes
> > >> > > > > >> to BaseOperator*: This metadata is used to
group tasks for
> > >> > > > > >> rendering at
> > >> > > > > >> UI level. It may potentially extend further
to group
> > arbitrary
> > >> > > tasks
> > >> > > > > >> outside the context of subdag to allow group-level
> operations
> > >> > > (i.e.
> > >> > > > > >> stop/trigger a group of task within the dag)
> > >> > > > > >> - *Webserver UI for SubDag*: Proposed UI modification
to
> > allow
> > >> > > > > >> (un)collapse a group of tasks for a flat structure
to pair
> > with
> > >> > > the
> > >> > > > > first
> > >> > > > > >> change instead of the original hierarchical
structure.
> > >> > > > > >>
> > >> > > > > >>
> > >> > > > > >> Please see related documents and PRs for details:
> > >> > > > > >> AIP:
> > >> > > > > >>
> > >> > > > >
> > >> > > >
> > >> > >
> > >> >
> > >>
> >
> https://cwiki.apache.org/confluence/display/AIRFLOW/AIP-34+Rewrite+SubDagOperator
> > >> > > > > >>
> > >> > > > > >> Original Issue:
> > https://github.com/apache/airflow/issues/8078
> > >> > > > > >> Draft PR: https://github.com/apache/airflow/pull/9243
> > >> > > > > >>
> > >> > > > > >> Please let me know if there are any aspects
that you
> > >> > agree/disagree
> > >> > > > > >> with or
> > >> > > > > >> need more clarification (especially the third
change
> > regarding
> > >> > > > > TaskGroup).
> > >> > > > > >> Any comments are welcome and I am looking
forward to it!
> > >> > > > > >>
> > >> > > > > >> Cheers
> > >> > > > > >> Bin
> > >> > > > > >>
> > >> > > > >
> > >> > > > --
> > >> > > > Kyle Hamlin
> > >> > > >
> > >> > >
> > >> >
> > >
> > >
> >
> > --
> > Thanks & Regards
> >  Poornima
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message