[
https://issues.apache.org/jira/browse/MESOS-5332?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15279791#comment-15279791
]
Stephan Erb commented on MESOS-5332:
------------------------------------
I was able to assemble a reproducing example (using Aurora master and Mesos
0.27.2): https://gist.github.com/StephanErb/5798b0d87c11473fb0ec147272ea0288
Summary of events:
* An iptables firewall is terminating idle TCP connections after the iptables
default of 5 days (reduced to 60 seconds in the example above).
* Mesos does not detect broken, half-open TCP connections that occur when
connections are terminated by iptables.
* Mesos tries to use the old, broken TCP connection when answering the agent
reconnect request. The message therefore never makes it to the agent.
* The agent ends up killing the executor because it does not receive a reply
for its reconnect request.
I'd conclude that there are several areas that need improvement:
* *Firewalling*: We have to fix our inhouse iptables firewall scripts so that
it does not apply connection tracking to local connections.
* *KeepAlive*: Mesos has to enable TCP keepalives in libprocess. As detailed
here (http://tldp.org/HOWTO/html_single/TCP-Keepalive-HOWTO/) this has two
advantages:
** Detection of dead peers in case there has been a hard crash, an unplugged
network cable, or if a firewall has silently dropped a connection.
** Prevention of disconnection due to network inactivity.
* *Unlink* By the addition of an {{unlink}} function libprocess could handle
exit events more gracefully by making sure it always creates a new connection
when talking to a new process.
> TASK_LOST on slave restart potentially due to executor race condition
> ---------------------------------------------------------------------
>
> Key: MESOS-5332
> URL: https://issues.apache.org/jira/browse/MESOS-5332
> Project: Mesos
> Issue Type: Bug
> Components: libprocess, slave
> Affects Versions: 0.26.0
> Environment: Mesos 0.26
> Aurora 0.13
> Reporter: Stephan Erb
> Attachments: executor-logs.tar.gz, executor-stderr.log,
> executor-stderrV2.log, mesos-slave.log
>
>
> When restarting the Mesos agent binary, tasks can end up as LOST. We lose
> from 20% to 50% of all tasks. They are killed by the Mesos agent via:
> {code}
> I0505 08:42:06.781318 21738 slave.cpp:2702] Cleaning up un-reregistered
> executors
> I0505 08:42:06.781366 21738 slave.cpp:2720] Killing un-reregistered executor
> 'thermos-nobody-devel-service-28854-0-6a88d62e-656
> 4-4e33-b0bb-1d8039d97afc' of framework
> 20151001-085346-58917130-5050-37976-0000 at executor(1)@10.X.X.X:40541
> I0505 08:42:06.781446 21738 slave.cpp:2720] Killing un-reregistered executor
> 'thermos-nobody-devel-service-23839-0-1d2cd0e6-699
> 4-4cba-a9df-3dfc1552667f' of framework
> 20151001-085346-58917130-5050-37976-0000 at executor(1)@10.X.X.X:35757
> I0505 08:42:06.781466 21738 slave.cpp:2720] Killing un-reregistered executor
> 'thermos-nobody-devel-service-29970-0-478a7291-d070-4aa8
> -af21-6fda889f750c' of framework 20151001-085346-58917130-5050-37976-0000 at
> executor(1)@10.X.X.X:51463
> ...
> I0505 08:42:06.781558 21738 slave.cpp:4230] Finished recovery
> {code}
> We have verified that the tasks and their executors are killed by the agent
> during startup. When stopping the agent using supervisorctl stop, the
> executors are still running (verified via {{ps aux}}). They are only killed
> once the agent tries to reregister.
> The issue is hard to reproduce:
> * When restarting the agent binary multiple times, tasks are only lost for
> the first restart.
> * It is much more likely to occur if the agent binary has been running for a
> longer period of time (> 7 days)
> * It tends to be more likely if the host has many cores (30-40) and thus many
> libprocess workers.
> Mesos is correctly sticking to the 2 seconds wait time before killing
> un-reregistered executors. The failed executors receive the reregistration
> request, but it seems like they fail to send a reply.
> A successful reregistration (not leading to LOST):
> {code}
> I0505 08:41:59.581231 21664 exec.cpp:456] Slave exited, but framework has
> checkpointing enabled. Waiting 15mins to reconnect with slave
> 20160118-141153-92471562-5050-6270-S17
> I0505 08:42:04.780591 21665 exec.cpp:256] Received reconnect request from
> slave 20160118-141153-92471562-5050-6270-S17
> I0505 08:42:04.785297 21676 exec.cpp:233] Executor re-registered on slave
> 20160118-141153-92471562-5050-6270-S17
> I0505 08:42:04.788579 21676 exec.cpp:245] Executor::reregistered took
> 1.492339ms
> {code}
> A failed one:
> {code}
> I0505 08:42:04.779677 2389 exec.cpp:256] Received reconnect request from
> slave 20160118-141153-92471562-5050-6270-S17
> E0505 08:42:05.481374 2408 process.cpp:1911] Failed to shutdown socket with
> fd 11: Transport endpoint is not connected
> I0505 08:42:05.481374 2395 exec.cpp:456] Slave exited, but framework has
> checkpointing enabled. Waiting 15mins to reconnect with slave
> 20160118-141153-92471562-5050-6270-S17
> {code}
> All task ending up in LOST have an output similar to the one posted above,
> i.e. messages seem to be received in a wrong order.
> Anyone an idea what might be going on here?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
