You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@airflow.apache.org by "Siddharth Anand (JIRA)" <ji...@apache.org> on 2016/06/01 23:24:59 UTC
[jira] [Commented] (AIRFLOW-191) Database connection leak on
Postgresql backend
[ https://issues.apache.org/jira/browse/AIRFLOW-191?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15311344#comment-15311344 ]
Siddharth Anand commented on AIRFLOW-191:
-----------------------------------------
Hi Sergei, [~criccomini]
I am running the Local Executor and I see the "idle" running COMMIT (and a few ROLLBACK) and I see the "idle in transaction" running the heartbeat query you are seeing.
{code}
Current query for procpid 121145:
SELECT job.id AS job_id, job.dag_id AS job_dag_id, job.state AS job_state, job.job_type AS job_job_type, job.start_date AS job_start_date, job.end_date AS job_end_date, job.latest_heartbeat AS job_latest_heartbeat, job.executor_class AS job_executor_class, job.hostname AS job_hostname, job.unixname AS job_unixname
FROM job
WHERE job.id = 166658
LIMIT 1
{code}
{code}
postgres=# select state, count(*)
from pg_stat_activity
where datname = 'airflow'
group by 1
;
state | count
---------------------+-------
idle in transaction | 18
idle | 22
(2 rows)
{code}
> Database connection leak on Postgresql backend
> ----------------------------------------------
>
> Key: AIRFLOW-191
> URL: https://issues.apache.org/jira/browse/AIRFLOW-191
> Project: Apache Airflow
> Issue Type: Bug
> Components: executor
> Affects Versions: Airflow 1.7.1.2
> Reporter: Sergei Iakhnin
>
> I raised this issue on github several months ago and there was even a PR but it never maid it into mainline. Basically, workers tend to hang onto DB connections in Postgres for recording heartbeat.
> I'm running a cluster with 115 workers, each with 8 slots. My Postgres DB is configured to allow 1000 simultaneous connections. I should effectively be able to run 920 tasks at the same time, but am actually limited to only about 450-480 because of idle transactions from workers hanging on to DB connections.
> If I run the following query
> select count(*),state, client_hostname from pg_stat_activity group by state, client_hostname
> These are the results:
> count state client_hostname
> 1 active (null)
> 1 idle localhost
> 451 idle in transaction (null)
> 446 idle (null)
> 1 active localhost
> The idle connections are all trying to run COMMIT
> The "idle in transaction" connections are all trying to run
> SELECT job.id AS job_id, job.dag_id AS job_dag_id, job.state AS job_state, job.job_type AS job_job_type, job.start_date AS job_start_date, job.end_date AS job_end_date, job.latest_heartbeat AS job_latest_heartbeat, job.executor_class AS job_executor_class, job.hostname AS job_hostname, job.unixname AS job_unixname
> FROM job
> WHERE job.id = 213823
> LIMIT 1
> with differing job.ids of course.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)