[jira] [Created] (MESOS-7166) port_mapping isolator: netns and veth are not GC-ed

["Pierre Cheynier (JIRA)" <ji...@apache.org>]

Pierre Cheynier created MESOS-7166:
--------------------------------------

             Summary: port_mapping isolator: netns and veth are not GC-ed
                 Key: MESOS-7166
                 URL: https://issues.apache.org/jira/browse/MESOS-7166
             Project: Mesos
          Issue Type: Bug
          Components: isolation, network
            Reporter: Pierre Cheynier


By testing port_mapping isolator during a few days in a preproduction environment where a lot of container starts, sometime fails, are destroyed continuously, I faced this issue: some agents host still have their network configuration, meaning that the netns, the veth interfaces, the tc rules are there.

Here is my setup.
* Cent OS 7.2
* LTS Kernel, 4.4.21 at that time
* libnl 3.2.28
* mesos 1.0.2 compiled using:
{noformat}
./configure \
    CFLAGS="%{optflags}" \
    CXXFLAGS="%{optflags}" \
    --disable-silent-rules \
    --prefix=%{_prefix} \
    --bindir=%{_bindir} \
    --libdir=%{_libdir} \
    --includedir=%{_includedir} \
    --disable-python \
    --disable-python-dependency-install \
    --enable-libevent \
    --enable-ssl \
    --enable-optimize \
    --with-network-isolator
{noformat}

I have logs that apparently says that at some point the container was considered as orphan (maybe due to an operation on the host, like an agent configuration update).

{noformat}
Feb 24 13:21:30 mesos-slave049-par mesos-slave[48375]: I0224 13:21:30.421066 48395 containerizer.cpp:690] Removing orphan container a8e05a03-7499-4566-bcba-53d8bf204e5f
Feb 24 13:21:30 mesos-slave049-par mesos-slave[48375]: I0224 13:21:30.421968 48395 linux_launcher.cpp:349] Using pid namespace to destroy container a8e05a03-7499-4566-bcba-53d8bf204e5f
{noformat}

It would be nice if the agent isolator could handle the cleaning of that.



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