Qian Yu created AIRFLOW-5391:
--------------------------------
Summary: Clearing a task skipped by BranchPythonOperator will
cause the task to execute
Key: AIRFLOW-5391
URL: https://issues.apache.org/jira/browse/AIRFLOW-5391
Project: Apache Airflow
Issue Type: Bug
Components: operators
Affects Versions: 1.10.4
Reporter: Qian Yu
I tried this on 1.10.3 and 1.10.4, both have this issue:
E.g. in this example from the doc, branch_a executed, branch_false was skipped
because of branching condition. However if someone Clear branch_false, it'll
cause branch_false to
execute.!https://airflow.apache.org/_images/branch_good.png!
This behaviour is understandable given how BranchPythonOperator is implemented.
BranchPythonOperator does not store its decision anywhere. It skips its own
downstream tasks in the branch at runtime. So there's currently no way for
branch_false to know it should be skipped without rerunning the branching task.
This is obviously counter-intuitive from the user's perspective. In this
example, users would not expect branch_false to execute when they clear it
because the branching task should have skipped it.
There are a few ways to improve this:
Option 1): Make downstream tasks skipped by BranchPythonOperator not clearable
without also clearing the upstream BranchPythonOperator. In this example, if
someone clears branch_false without clearing branching, branch_false should not
execute.
Option 2): Make BranchPythonOperator store the result of its skip condition
somewhere. Make downstream tasks check for this stored decision and skip
themselves if they should have been skipped by the condition. This probably
means the decision of BranchPythonOperator needs to be stored in the db.
[kevcampb|https://blog.diffractive.io/author/kevcampb/] attempted a workaround
and on this blog. And he acknowledged his workaround is not perfect and a
better permanent fix is needed:
[https://blog.diffractive.io/2018/08/07/replacement-shortcircuitoperator-for-airflow/]
--
This message was sent by Atlassian Jira
(v8.3.2#803003)
