[
https://issues.apache.org/jira/browse/AIRFLOW-160?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15307352#comment-15307352
]
Paul Yang commented on AIRFLOW-160:
-----------------------------------
In the preliminary PR, the child process handles both the parsing of the DAG
definition file and scheduling of task instances. Because scheduling needs to
be done anyway, the periodic parsing of DAG definition files remains. We could
prioritize files (or force a refresh) based on an API call in addition though.
> Parse DAG files through child processes
> ---------------------------------------
>
> Key: AIRFLOW-160
> URL: https://issues.apache.org/jira/browse/AIRFLOW-160
> Project: Apache Airflow
> Issue Type: Improvement
> Components: scheduler
> Reporter: Paul Yang
> Assignee: Paul Yang
>
> Currently, the Airflow scheduler parses all user DAG files in the same
> process as the scheduler itself. We've seen issues in production where bad
> DAG files cause scheduler to fail. A simple example is if the user script
> calls `sys.exit(1)`, the scheduler will exit as well. We've also seen an
> unusual case where modules loaded by the user DAG affect operation of the
> scheduler. For better uptime, the scheduler should be resistant to these
> problematic user DAGs.
> The proposed solution is to parse and schedule user DAGs through child
> processes. This way, the main scheduler process is more isolated from bad
> DAGs. There's a side benefit as well - since parsing is distributed among
> multiple processes, it's possible to parse the DAG files more frequently,
> reducing the latency between when a DAG is modified and when the changes are
> picked up.
> Another issue right now is that all DAGs must be scheduled before any tasks
> are sent to the executor. This means that the frequency of task scheduling is
> limited by the slowest DAG to schedule. The changes needed for scheduling
> DAGs through child processes will also make it easy to decouple this process
> and allow tasks to be scheduled and sent to the executor in a more
> independent fashion. This way, overall scheduling won't be held back by a
> slow DAG.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
