----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/53403/#review154671 -----------------------------------------------------------
Overall this looks good to me. I have no idea how to test this in practice other than to set up an e2e test that starts up mesos w/ pid isolation enabled, starting a job that double forks and then ensuring everything is torn down properly. src/main/python/apache/thermos/common/process_util.py (line 72) <https://reviews.apache.org/r/53403/#comment224299> Prefix this log statement so we have context. src/main/python/apache/thermos/core/helper.py (lines 236 - 240) <https://reviews.apache.org/r/53403/#comment224298> Something like: orphaned_pids = coordinator_pids - {c.pid for c in psutil.Process().children()} is maybe more pythonic (assuming I got my syntax right anyway...) src/main/python/apache/thermos/core/helper.py (line 242) <https://reviews.apache.org/r/53403/#comment224297> I think this would probably make sense as an info log. - Joshua Cohen On Nov. 2, 2016, 10:16 p.m., Zameer Manji wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/53403/ > ----------------------------------------------------------- > > (Updated Nov. 2, 2016, 10:16 p.m.) > > > Review request for Aurora, Joshua Cohen, Santhosh Kumar Shanmugham, and > Stephan Erb. > > > Bugs: AURORA-1808 > https://issues.apache.org/jira/browse/AURORA-1808 > > > Repository: aurora > > > Description > ------- > > This is a WIP patch showing a possible fix to AURORA-1808. > > # Problem > > Processes can deamonize and escape the supervision of a coordinator. Using > the Docker Containerizer or the Mesos Containerizer with pid isolation means > that the processes will be come reparented to the `sh` process that launches > the executor. For example: > ```` > root@aurora:/# ps xf > PID TTY STAT TIME COMMAND > 48 ? Ss 0:00 /bin/bash > 86 ? R+ 0:00 _ ps xf > 1 ? Ss 0:00 /bin/sh -c ${MESOS_SANDBOX=.}/thermos_executor.pex > --announcer-ensemble localhost:2181 --announcer-zookeeper-auth-config > /home/vagrant/aurora/examples/va > 5 ? Sl 0:02 python2.7 /mnt/mesos/sandbox/thermos_executor.pex > --announcer-ensemble localhost:2181 --announcer-zookeeper-auth-config > /home/vagrant/aurora/examples/vag > 23 ? S 0:00 _ /usr/local/bin/python2.7 > /mnt/mesos/sandbox/thermos_runner.pex > --task_id=www-data-devel-hello_docker_engine-0-bde5cdc7-8685-46fd-9078-4a86bd5be152 > -- > 29 ? Ss 0:00 _ /usr/local/bin/python2.7 > /mnt/mesos/sandbox/thermos_runner.pex > --task_id=www-data-devel-hello_docker_engine-0-bde5cdc7-8685-46fd-9078-4a86bd5be15 > 32 ? S 0:00 | _ /bin/bash -c while true; do > echo hello world sleep 10 done > 81 ? S 0:00 | _ sleep 10 > 31 ? Ss 0:00 _ /usr/local/bin/python2.7 > /mnt/mesos/sandbox/thermos_runner.pex > --task_id=www-data-devel-hello_docker_engine-0-bde5cdc7-8685-46fd-9078-4a86bd5be15 > 33 ? S 0:00 _ /bin/bash -c while true; do > echo hello world sleep 10 done > 82 ? S 0:00 _ sleep 10 > 47 ? S 0:00 python ./daemon.py > ```` > > # Solution > Ensure processes that escape the supervision of the coordinator reparent to > the runner who can send signals to them on task tear down. We do this by > using the `PR_SET_CHILD_SUBREAPER` flag of `prctl(2)`. > > After this change the process tree looks like: > ```` > root@aurora:/# ps xf > PID TTY STAT TIME COMMAND > 66 ? Ss 0:00 /bin/bash > 70 ? R+ 0:00 _ ps xf > 1 ? Ss 0:00 /bin/sh -c ${MESOS_SANDBOX=.}/thermos_executor.pex > --announcer-ensemble localhost:2181 --announcer-zookeeper-auth-config > /home/vagrant/aurora/examples/va > 5 ? Sl 0:02 python2.7 /mnt/mesos/sandbox/thermos_executor.pex > --announcer-ensemble localhost:2181 --announcer-zookeeper-auth-config > /home/vagrant/aurora/examples/vag > 23 ? S 0:00 _ /usr/local/bin/python2.7 > /mnt/mesos/sandbox/thermos_runner.pex > --task_id=www-data-devel-hello_docker_engine-0-721406db-00f5-4c0c-915e-1dbc5568b849 > -- > 33 ? Ss 0:00 _ /usr/local/bin/python2.7 > /mnt/mesos/sandbox/thermos_runner.pex > --task_id=www-data-devel-hello_docker_engine-0-721406db-00f5-4c0c-915e-1dbc5568b84 > 40 ? S 0:00 | _ /bin/bash -c while true; do > echo hello world sleep 10 done > 63 ? S 0:00 | _ sleep 10 > 36 ? Ss 0:00 _ /usr/local/bin/python2.7 > /mnt/mesos/sandbox/thermos_runner.pex > --task_id=www-data-devel-hello_docker_engine-0-721406db-00f5-4c0c-915e-1dbc5568b84 > 37 ? S 0:00 | _ /bin/bash -c while true; do > echo hello world sleep 10 done > 62 ? S 0:00 | _ sleep 10 > 55 ? S 0:00 _ python ./daemon.py > ```` > > Now the runner is aware of the reparented procesess can can tear it down > cleanly during teardown. > > Note that the man page for `prctl(2)` says that the processes that set > `PR_SET_CHILD_SUBREAPER` should reap children to get rid of zombies. It is > important to note tht the runner already does this in its run loop via > `TaskRunnerHelper.reap_children()`. This patch has the side effect of > ensuring it will reap all of the children launched via coordinators. > > > Diffs > ----- > > src/main/python/apache/thermos/common/process_util.py > abd2c0ef35858d13971319b0a7436ce2293824ce > src/main/python/apache/thermos/core/helper.py > 68855e1e54ba1cd4456e18a36fb237ce6a468c34 > src/main/python/apache/thermos/core/process.py > 3ec43e2719ef97026f399c4b2aa23002559b3153 > src/main/python/apache/thermos/core/runner.py > 7b9013d11f6ff4172b6b7bf56e62299b0d11c977 > > Diff: https://reviews.apache.org/r/53403/diff/ > > > Testing > ------- > > no automated tests yet. > > Validated behaviour with `ps` and `strace`. > > > Thanks, > > Zameer Manji > >
