Mailing-List: contact commits-help@airflow.incubator.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@airflow.incubator.apache.org
Date: Fri, 16 Dec 2016 15:26:58 +0000 (UTC)
From: "Len Frodgers (JIRA)" <jira@apache.org>
To: commits@airflow.incubator.apache.org
Message-ID: <JIRA.13028692.1481901655000.539807.1481902018747@Atlassian.JIRA>
In-Reply-To: <JIRA.13028692.1481901655000@Atlassian.JIRA>
References: <JIRA.13028692.1481901655000@Atlassian.JIRA> <JIRA.13028692.1481901655822@arcas>
Subject: [jira] [Updated] (AIRFLOW-703) Xcom data cleared too soon
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit
archived-at: Fri, 16 Dec 2016 15:27:08 -0000


     [ https://issues.apache.org/jira/browse/AIRFLOW-703?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Len Frodgers updated AIRFLOW-703:
---------------------------------
    Attachment: xcom_bug_op2_logs.txt
                xcom_bug_op1_logs.txt

task instance logs attached

> Xcom data cleared too soon
> --------------------------
>
>                 Key: AIRFLOW-703
>                 URL: https://issues.apache.org/jira/browse/AIRFLOW-703
>             Project: Apache Airflow
>          Issue Type: Bug
>          Components: core, scheduler, xcom
>    Affects Versions: Airflow 2.0, Airflow 1.7.1.3
>         Environment: Tested using Dockerized Airflow setup with MySQL backend and Celery executor
>            Reporter: Len Frodgers
>              Labels: xcom
>         Attachments: xcom_bug.py, xcom_bug_op1_logs.txt, xcom_bug_op2_logs.txt
>
>
> Xcom data is cleared at the start of the `run` method of the `TaskInstance`, regardless of whether the TI is subsequently executed (e.g. if the TI has previously succeeded, it won't execute). This means that if a TI for a DagRun is run twice in close succession, the latter will correctly not execute (since the former TI succeeded or is still running), but WILL clear any xcoms set by the former TI. Therefore, any downstream tasks depending on these xcoms will fail.
> I noticed this bug when I changed num_runs of the scheduler from None to 10. It didn't happen every time, but probably 50% or so.
> However, I can reproduce this reliably and repeatably with the following test dag:
> [attached]
> To make op1 execute twice, I use the UI to run it twice while op2 is doing the `time.sleep`.
> Logs from running this:
> [attached]
> The fix seems straightforward: don't clear xcom unless the TI will actually execute. Will happily create a PR.


--
This message was sent by Atlassian JIRA
(v6.3.4#6332)