[jira] [Commented] (YARN-2342) When killing a task, we don't always need to send a subsequent SIGKILL
[ https://issues.apache.org/jira/browse/YARN-2342?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14523773#comment-14523773 ] Junping Du commented on YARN-2342: -- Thanks [~aw] for moving the issue. I will take a look for LinuxContainerExecutor. > When killing a task, we don't always need to send a subsequent SIGKILL > -- > > Key: YARN-2342 > URL: https://issues.apache.org/jira/browse/YARN-2342 > Project: Hadoop YARN > Issue Type: Improvement >Reporter: Vinod Kumar Vavilapalli >Assignee: Junping Du > Labels: newbie > > In both TaskController/LinuxTaskController, while killing tasks, first a > SIGTERM and then a subsequent SIGKILL. We don't need to send the SIGKILL > always. It can be avoided when the SIGTERM command (kill pid for process or > kill -- -pid for session) returns a non-zero exit code, i.e. when the signal > is not sent successfully because process/process group doesn't exist. 'man 2 > kill' says exit code is non-zero only when process/process group is not alive > or invalid signal is specified or the process doesn't have permissions. The > last two don't happen in mapred code. -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (YARN-2342) When killing a task, we don't always need to send a subsequent SIGKILL
[ https://issues.apache.org/jira/browse/YARN-2342?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14072249#comment-14072249 ] Allen Wittenauer commented on YARN-2342: Moving this to YARN, as we need to check container executor. > When killing a task, we don't always need to send a subsequent SIGKILL > -- > > Key: YARN-2342 > URL: https://issues.apache.org/jira/browse/YARN-2342 > Project: Hadoop YARN > Issue Type: Improvement >Reporter: Vinod Kumar Vavilapalli > Labels: newbie > > In both TaskController/LinuxTaskController, while killing tasks, first a > SIGTERM and then a subsequent SIGKILL. We don't need to send the SIGKILL > always. It can be avoided when the SIGTERM command (kill pid for process or > kill -- -pid for session) returns a non-zero exit code, i.e. when the signal > is not sent successfully because process/process group doesn't exist. 'man 2 > kill' says exit code is non-zero only when process/process group is not alive > or invalid signal is specified or the process doesn't have permissions. The > last two don't happen in mapred code. -- This message was sent by Atlassian JIRA (v6.2#6252)
