airflow-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Bolke de Bruin (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (AIRFLOW-928) Same {task,execution_date} run multiple times in worker when using CeleryExecutor
Date Tue, 14 Mar 2017 19:30:42 GMT

    [ https://issues.apache.org/jira/browse/AIRFLOW-928?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15924861#comment-15924861
] 

Bolke de Bruin commented on AIRFLOW-928:
----------------------------------------

[~saguziel] have you seen this?

> Same {task,execution_date} run multiple times in worker when using CeleryExecutor
> ---------------------------------------------------------------------------------
>
>                 Key: AIRFLOW-928
>                 URL: https://issues.apache.org/jira/browse/AIRFLOW-928
>             Project: Apache Airflow
>          Issue Type: Bug
>          Components: celery
>    Affects Versions: Airflow 1.7.1.3
>         Environment: Docker
>            Reporter: Uri Shamay
>         Attachments: airflow.log, dag_runs.png, dummy_dag.py, processes.list, rabbitmq.queue,
scheduler.log, worker_2.log, worker.log
>
>
> Hi,
> When using with Airflow with CeleryExecutor, both RabbitMQ && Redis I tested,
I see that when workers are down, the scheduler run each period of time **append** to the
same key of {task,execution_date} in the broker, the same {task,execution_date}, what means
is that if workers are down/can't connect to broker for few hours, I got in the broker thousands
of same executions.
> In my scenario I have just one dummy dag to run with dag_concurrency of 4,
> I expected in that scenario that broker will hold just 4 messages, and the scheduler
shouldn't queuing another and another and another for same {task, execution_date}
> What happened is that when workers start to consume messages, they got thousands of tasks
for just 4 tasks, and when they trying to write to database for task_instances - there are
errors of integrity while such {task,execution_date} already exist.
> Note that in my test after let Airflow to consume works of just one dag without workers
for few hours, then I connect to the broker outside by custom client and retrieve the messages
- there was thousands of same {dag,execution_date}.
> Even if the case is that there are a lot of dag works on the same key that run just one
instance when poll thousands - it's still bad behavior, better to produce one message to the
queue, and if some timeout occurred (like visibility), to set the key - and not append to
it. 
> What happened is when workers are down for long time and have a lot of jobs that scheduled
each minute, when workers come back, they got thousands of same jobs => cause to the worker
to run the same dags a lot of times => a lot of wasted python runners => utilized all
celery worker threads/processes => starve all other jobs till he understood that need just
one instance from all same.
> Attached files:
> 1. airflow.log - this is the task log, you can see that few instances processes of same
{task,execution_date} write to the same log file.
> 2. worker.log - this is the worker log, you can see that worker trying to run same {task,execution_date}
multiple times + the errors from the database integrity that said that those tasks on those
dates already exists.
> 3. scheduler.log to show that scheduler decided to send again and again and again infinitely
the same {job,execution_date}
> 4. the dummy_dag.py of the test
> 5. rabbitmq.queue - show that after 5 minutes the broker queue contains 40 messages of
same 4 {job,execution_date}
> 6. dag_runs.png - show that there are only 4 jobs that need to be run, while there are
much more messages in the queue
> 7. processes.list - show that when start worker and doing: ps -ef | grep "airflow run",
it show that worker run multiple times same {job,execution_date}
> 8. worker_2.log - show that when worker started - the same {job,execution_date} keys
shown multiple times
> Thanks.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Mime
View raw message