Hi Joris, all,
We are still at the default timeout values for those that you linked. In the meantime, since the community pushed us to look at other things besides evading firewall timeouts, we have disabled snapshot/backups on the VMs and this has resolved the issue for the past 24 hours on the control group that we have disabled, which has been the best behavior that we have ever seen. There was a very close correlation between snapshot creation and mesos-slave process restart (within minutes) that got us to this point. Apparently, the snapshot creation and quiesce of the filesystem cause enough disruption to trigger the default timeouts within mesos. We are fine with this solution because Mesos has enabled us to have a more heterogeneous fleet of servers and backups aren't needed on these hosts. Mesos for the win, there. Thanks to everyone that has contributed on this thread! It was a fun exercise for me, in the code. It was also useful to hear feedback from the list on places to look, eventually pushing me to a solution. -Jeremy ________________________________ From: Joris Van Remoortere <jo...@mesosphere.io> Sent: Wednesday, November 11, 2015 12:56 AM To: user@mesos.apache.org Subject: Re: Mesos and Zookeeper TCP keepalive Hi Jeremy, Can you read the description of these<https://github.com/apache/mesos/blob/249bc26306574d9db0527c04b7a83a1f1e75f71b/src/master/flags.cpp#L393-L422> parameters on the master, and possibly share your values for these flags? It seems from the re-registration attempt on the agent, that the master has already treated the agent as "failed", and so will tell it to shut down on any re-registration attempt. I'm curious if there is a conflict (or too narrow of a time gap) of timeouts in your environment to allow re-registration by the agent after the agent notices it needs to re-establish the connection. — Joris Van Remoortere Mesosphere On Tue, Nov 10, 2015 at 5:02 AM, Jeremy Olexa <jol...@spscommerce.com<mailto:jol...@spscommerce.com>> wrote: Hi Tommy, Erik, all, You are correct in your assumption that I'm trying to solve for a one hour session expire time on a firewall. For some more background info, our master cluster is in datacenter X, the slaves in X will stay "up" for days and days. The slaves in a different datacenter, Y, connected to that master cluster will stay "up" for about a few days and restart. The master cluster is healthy, with a stable leader for months (no flapping), same for the ZK "leader". There are about 35 slaves in datacenter Y. Maybe the firewall session timer is a red herring because the slave restart is seemingly random (the slave with the highest uptime is 6 days, but a handful only have uptime of a day) I've started debugging this awhile ago, and the gist of the logs is here: https://gist.github.com/jolexa/1a80e26a4b017846d083 I've posted this back in October seeking help and Benjamin suggested network issues in both directions, so I thought firewall. Thanks for any hints, Jeremy ________________________________ From: tommy xiao <xia...@gmail.com<mailto:xia...@gmail.com>> Sent: Tuesday, November 10, 2015 3:07 AM To: user@mesos.apache.org<mailto:user@mesos.apache.org> Subject: Re: Mesos and Zookeeper TCP keepalive same here , same question with Erik. could you please input more background info, thanks 2015-11-10 15:56 GMT+08:00 Erik Weathers <eweath...@groupon.com<mailto:eweath...@groupon.com>>: It would really help if you (Jeremy) explained the *actual* problem you are facing. I'm *guessing* that it's a firewall timing out the sessions because there isn't activity on them for whatever the timeout of the firewall is? It seems likely to be unreasonably short, given that mesos has constant activity between master and slave/agent/whatever-it-is-being-called-nowadays-but-not-really-yet-maybe-someday-for-reals. - Erik On Mon, Nov 9, 2015 at 10:00 PM, Jojy Varghese <j...@mesosphere.io<mailto:j...@mesosphere.io>> wrote: Hi Jeremy Its great that you are making progress but I doubt if this is what you intend to achieve since network failures are a valid state in distributed systems. If you think there is a special case you are trying to solve, I suggest proposing a design document for review. For ZK client code, I would suggest asking the zookeeper mailing list. thanks -Jojy On Nov 9, 2015, at 7:56 PM, Jeremy Olexa <jol...@spscommerce.com<mailto:jol...@spscommerce.com>> wrote: Alright, great, I'm making some progress, I did a simple copy/paste modification and recompiled mesos. The keepalive timer is set from slave to master so this is an improvement for me. I didn't test the other direction yet - https://gist.github.com/jolexa/ee9e152aa7045c558e02 - I'd like to file an enhancement request for this since it seems like an improvement for other people as well, after some real world testing I'm having some harder time figuring out the zk client code. I started by modifying build/3rdparty/zookeeper-3.4.5/src/c/zookeeper.c but either a) my change wasn't correct or b) I'm modifying a wrong file, since I just assumed using the c client. Is this the correct place? Thanks much, Jeremy ________________________________ From: Jojy Varghese <j...@mesosphere.io<mailto:j...@mesosphere.io>> Sent: Monday, November 9, 2015 2:09 PM To: user@mesos.apache.org<mailto:user@mesos.apache.org> Subject: Re: Mesos and Zookeeper TCP keepalive Hi Jeremy The “network” code is at "3rdparty/libprocess/include/process/network.hpp” , "3rdparty/libprocess/src/poll_socket.hpp/cpp”. thanks jojy On Nov 9, 2015, at 6:54 AM, Jeremy Olexa <jol...@spscommerce.com<mailto:jol...@spscommerce.com>> wrote: Hi all, Jojy, That is correct, but more specifically a keepalive timer from slave to master and slave to zookeeper. Can you send a link to the portion of the code that builds the socket/connection? Is there any reason to not set the SO_KEEPALIVE option in your opinion? hasodent, I'm not looking for keepalive between zk quorum members, like the ZOOKEEPER JIRA is referencing. Thanks, Jeremy ________________________________ From: Jojy Varghese <j...@mesosphere.io<mailto:j...@mesosphere.io>> Sent: Sunday, November 8, 2015 8:37 PM To: user@mesos.apache.org<mailto:user@mesos.apache.org> Subject: Re: Mesos and Zookeeper TCP keepalive Hi Jeremy Are you trying to establish a keepalive timer between mesos master and mesos slave? If so, I don’t believe its possible today as SO_KEEPALIVE option is not set on an accepting socket. -Jojy On Nov 8, 2015, at 8:43 AM, haosdent <haosd...@gmail.com<mailto:haosd...@gmail.com>> wrote: I think keepalive option should be set in Zookeeper, not in Mesos. See this related issue in Zookeeper. https://issues.apache.org/jira/browse/ZOOKEEPER-2246?focusedCommentId=14724085&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14724085 On Sun, Nov 8, 2015 at 4:47 AM, Jeremy Olexa <jol...@spscommerce.com<mailto:jol...@spscommerce.com>> wrote: Hello all, We have been fighting some network/session disconnection issues between datacenters and I'm curious if there is anyway to enable tcp keepalive on the zookeeper/mesos sockets? If there was a way, then the sysctl tcp kernel settings would be used. I believe keepalive has to be enabled by the software which is opening the connection. (That is my understanding anyway) Here is what I see via netstat --timers -tn: tcp 0 0 172.18.1.1:55842<http://172.18.1.1:55842/> 10.10.1.1:2181<http://10.10.1.1:2181/> ESTABLISHED off (0.00/0/0) tcp 0 0 172.18.1.1:49702 10.10.1.1:5050 ESTABLISHED off (0.00/0/0) Where 172 is the mesos-slave network and 10 is the mesos-master network. The "off" keyword means that keepalive's are not being sent. I've trolled through JIRA, git, etc and cannot easily determine if this is expected behavior or should be an enhancement request. Any ideas? Thanks much! -Jeremy -- Best Regards, Haosdent Huang -- Deshi Xiao Twitter: xds2000 E-mail: xiaods(AT)gmail.com<http://gmail.com>