回复: [PROPOSAL] Add Python API implementation of workflows-as-code

[zhang junfan <ju...@outlook.com>]

Good job, thanks focusing on multi-lang support.

Minor discussion.

  1.  Could you please provide some spark/flink process examples?
  2.  I'm confused with workflow-as-code, you means it just define the DAG and workflow parameters? Could we combine workflow and user task code(like spark/flink programs)?

________________________________
发件人: Jiajie Zhong <zh...@hotmail.com>
发送时间: 2021年9月28日 11:42
收件人: dev@dolphinscheduler.apache.org <de...@dolphinscheduler.apache.org>
主题: [PROPOSAL] Add Python API implementation of workflows-as-code

Hey guys,

    Apache DolphinScheduler is a good tool for workflow scheduler, it’s easy-to-extend,
distributed and have nice UI to create and maintain workflow. Our workflow only support
define in UI, which is easy to use and user friendly, it’s good but could be batter by
adding extend API and make workflow could define as code or yaml file. And consider yaml
file it’s hard to maintain manually I think it better to use code to define it, aka workflows-as-code.

    When workflow definitions as code, we could easy to modify some configure and do
some batch change for it. It’s could more easy to define similar task by loop statement,
and it give ability adding unittest for workflow too. I hope Apache DolphinScheduler could
combine the benefit of define by code and by UI, so I raise proposal for adding
workflows-as-code to Apache DolphinScheduler.

    Actually, I already start it by adding POC PR[1]. In this PR, I adding Python API give
user define workflow by Python code. This feature use *Py4J* connect Java and Python,
which mean I never add any new database model and infra to Apache DolphinScheduler,
I just reuse layer service in dolphinscheduler-api package to create workflow. And we could
consider Python API just another interface for Apache DolphinScheduler, just like our UI, it
allow we define and maintain workflow follow their rule.

    Here it’s an tutorial workflow definitions by Python API, which you could find it in PR file[2]

```python
from pydolphinscheduler.core.process_definition import ProcessDefinition
from pydolphinscheduler.tasks.shell import Shell

with ProcessDefinition(name="tutorial") as pd:
    task_parent = Shell(name="task_parent", command="echo hello pydolphinscheduler")
    task_child_one = Shell(name="task_child_one", command="echo 'child one'")
    task_child_two = Shell(name="task_child_two", command="echo 'child two'")
    task_union = Shell(name="task_union", command="echo union")

    task_group = [task_child_one, task_child_two]
    task_parent.set_downstream(task_group)

    task_union << task_group

    pd.run()
```

    In tutorial, we define a new ProcessDefinition named ‘tutorial’ using python context,
and then we add four Shell tasks to ‘tutorial’, just five line we could create one process
definition with four tasks.
    Beside process definition and tasks, another think we have to
add to workflow it’s task dependent, we add function `set_downstream` and `set_upstream`
to describe task dependent. At the same time, we overwrite bit operator and add a shortcut
`>>` and  `<<` to do it.
   After dependent set, we done our workflow definition, but all definition are in Python API
side, which mean it not persist to Apache DolphinScheduler database, and it could not runs
by Apache DolphinScheduler until declare `pd.submit()` or directly run it by `pd.run()`


[1]: https://github.com/apache/dolphinscheduler/pull/6269 <https://github.com/apache/dolphinscheduler/pull/6269>
[2]: https://github.com/apache/dolphinscheduler/pull/6269/files#diff-5561fec6b57cc611bee2b0d8f030965d76bdd202801d9f8a1e2e74c21769bc41 <https://github.com/apache/dolphinscheduler/pull/6269/files#diff-5561fec6b57cc611bee2b0d8f030965d76bdd202801d9f8a1e2e74c21769bc41>


Best Wish
― Jiajie

Re: ������: [PROPOSAL] Add Python API implementation of workflows-as-code

[Jiajie Zhong <zh...@apache.org>]

Hi junfan,

>   1.  Could you please provide some spark/flink process examples?

In this POC, we would just create an example for Shell task, to make sure all basic component are good to work. We would support all tasks in DolphinScheduler UI laster, and all of the tasks would have examples, but sorry that should not support for this PR. If you interesting in spark/flink tasks, please join us develop, I will update all we have to do in DSIP issue[1], your could pick one feature you interesting to do it.

>   2.  I'm confused with workflow-as-code, you means it just define the DAG and workflow parameters? Could we combine workflow and user task code(like spark/flink programs)?

Sorry I not quit understand what's your `user task code` mean. But it's just define workflow, including task type, name, parameters, dependent etc. It would pass all things you have to define in UI to DolphinScheduler, and create ProcessDefinition and TaskDefinition so on.

But if you have another idea about it, please share it to me

[1]: https://github.com/apache/dolphinscheduler/issues/6407

Best Wish
― Jiajie