On Mon, Apr 11, 2016, at 10:52 AM, Jakub Libosvar wrote: > On 04/11/2016 06:41 PM, Clark Boylan wrote: > > On Mon, Apr 11, 2016, at 03:07 AM, Jakub Libosvar wrote: > >> Hi, > >> > >> recently we hit an issue in Neutron with tests getting stuck [1]. As a > >> side effect we discovered logs are not collected properly which makes it > >> hard to find the root cause. The reason of missing logs is that we send > >> SIGKILL to whatever gate hook is running when we hit the global timeout > >> per gate job [2]. This gives no time to running process to perform any > >> post-processing. In post_gate_hook function in Neutron, we collect logs > >> from /tmp directory, compress them and move them to /opt/stack/logs to > >> make them exposed. > >> > >> I have in mind two solutions to which I'd like to get feedback before > >> sending patches. > >> > >> 1) In Neutron, we execute tests in post_gate_hook (dunno why). But even > >> if we would have moved test execution into gate_hook and tests get stuck > >> then the post_gate_hook won't be triggered [3]. So the solution I > >> propose here is to terminate gate_hook N minutes before global timeout > >> and still execute post_gate_hook (with timeout) as post-processing > >> routine. > >> > >> 2) Second proposal is to let timeout wrapped commands know they are > >> about to be killed. We can send let's say SIGTERM instead of SIGKILL and > >> after certain amount of time, send SIGKILL. Example: We send SIGTERM 3 > >> minutes before global timeout, letting these 3 minutes to 'command' to > >> handle the SIGTERM signal. > >> > >> timeout -s 15 -k 3 $((REMAINING_TIME-3))m bash -c "command" > >> > >> With the 2nd approach we can trap the signal that kills running test > >> suite and collects logs with same functions we currently have. > >> > >> > >> I would personally go with second option but I want to hear if anybody > >> has a better idea about post processing in gate jobs or if there is > >> already a tool we can use to collect logs. > >> > >> Thanks, > >> Kuba > > > > Devstack gate already does a "soft" timeout [0] then proceeds to cleanup > > (part of which is collecting logs) [1], then Jenkins does the "hard" > > timeout [2]. Why aren't we collecting the required log files as part of > > the existing cleanup? > This existing cleanup doesn't support hooks. Neutron tests produce a lot > of logs by default stored in /tmp/dsvm-<job_name> so we need to compress > and move them to /opt/stack/logs in order to get them collected by [1].
My suggestion would be to stop writing these log files to /tmp and instead write them to the log dir where they will be automagically compressed and collected. > > > > > [0] > > https://git.openstack.org/cgit/openstack-infra/devstack-gate/tree/devstack-vm-gate-wrap.sh#n569 > > [1] > > https://git.openstack.org/cgit/openstack-infra/devstack-gate/tree/devstack-vm-gate-wrap.sh#n594 > > [2] > > https://git.openstack.org/cgit/openstack-infra/project-config/tree/jenkins/jobs/devstack-gate.yaml#n325 > > > > Clark > > > > __________________________________________________________________________ > > OpenStack Development Mailing List (not for usage questions) > > Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe > > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev > > > > > __________________________________________________________________________ > OpenStack Development Mailing List (not for usage questions) > Unsubscribe: > openstack-dev-requ...@lists.openstack.org?subject:unsubscribe > http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev