When zombie s arise they can be handled manually by ecflow_ui. (See Zombie) or via the command line interface:
ecflow_client –-zombie_get # This will list all the task/jobs the server thinks are zombies.
ecflow_client --zombie_kill=<task-path> # Ask the server kill the zombie process. Use ECF_KILL_CMD
ecflow_client –-zombie_fail=<task-path> # Ask the zombie to fail. This may result in another zombie because abort child command in the job, will be called.
ecflow_client –-zombie_fob=<task-path> # Used to unblock the child, allows the job to proceed. However this will only work for zombies where the password does not match.
ecflow_client –-zombie_adopt=<task-path> # Copies the password stored on the zombie onto the task. Allows the job to proceed, and update the state in the server
# ( i.e. due to init,complete,abort).
# It is up to the user, to ensure that the zombie has been dealt with before doing this.
ecflow_client –-zombie_remove=<task-path> # Remove the zombie representation in the server. Typically this is done, when we are sure we have handled the zombie.
# The zombie will re-appear next time it communicates with server, if this is not the case.
ecflow_client –-zombie_block =<task-path> # Ask the jobs to block at the child command in the job. Prevents the job from proceeding.
# (This is the default behaviour for the init, complete and abort child commands)
|
Sometimes we may want the job to proceed but "ecflow_client –zombie_adopt=<task-path>" does not work. i.e. we have the case where zombies password matches, but the process id (ECF_RID) are different.
ecflow_client –zombie_adopt=<task-path>, will not allow this, due to the potential for data corruption.
In this case the normal behaviour would be kill both process, and re-queue the task.
In the extreme, we can by pass the authentication. (i.e. allowing the request to be handled by the server).
This should ONLY be done when you are sure the zombie has been killed, and you don’t want to re-queue the job.
> ecflow_client --alter=add variable ECF_PASS FREE /path/to/task |
This is also available from the GUI. Select the task. RMB->Special-> Free password.
After the job has completed, be sure to delete this variable. Otherwise if zombies arise again, there is a considerable risk of data corruption.
It is also possible to ask ecflow_server to make the same response in an automated fashion. How ever very careful consideration should be made before doing this. Otherwise it could mask a serious underlying problem.
The automated response can be defined statically using python and text interface or dynamically (add/remove) via alter.:
python interface( See ecflow.ZombieAttr)
text interface ( See Definition file Grammar)
zombie ::= "zombie" >> `zombie_type` >> ":" >> !(`client_side_action` | `server_side_action`) >> ":" >> *`child` >> ":" >> !`zombie_life_time`
zombie_type ::= "user" | "ecf" | "path" | "ecf_pid" | "ecf_passwd" | "ecf_pid_passwd"
child ::= "init" | "event" | "meter" | "label" | "wait" | "abort" | "complete"
client_side_action ::= "fob" | "fail" | "block"
server_side_action ::= "adopt" | "delete | "kill"
zombie_life_time ::= unsigned integer( default: user(300), ecf(3600), path(900) ), the server poll timer runs every 60 seconds, hence this is the effective minimum value
Where:ecf_pid - PID miss-match, password matches. Job scheduled twice. Check submitter
ecf_pid_passwd - Both PID and password miss-match. Re-queue & submit of active job?
ecf_passwd - Password miss-match, PID matches, system has re-cycled PID or hacked job file?
ecf - Two init commands or task complete or aborted but receives another child cmd
ecf_user - Created by user action
ecf_path - Task not found. Nodes replaced whilst jobs were running
The zombie attribute is inherited in the same manner as Variable inheritance.
Example: For tasks under suite “s1” add a zombie attribute, such that child label commands(i.e.. ecflow_client –label) never blocks the job: (not strictly needed as this is the default behaviour)
python
s1 = ecflow.Suite('s1')
child_list = [ ChildCmdType.label ]
zombie_attr = ZombieAttr(ZombieType.ecf, child_list, ZombieUserActionType.fob, 300)
s1.add_zombie(zombie_attr)
text
suite s1
zombie ecf:fob:label:Example: For tasks under suite “s1” add a zombie attribute, such that job that issues the child commands( event, meter, label) never blocks: (not strictly needed as this is the default behaviour)
python
s1 = ecflow.Suite('s1')
child_list = [ ChildCmdType.label, ChildCmdType.event, ChildCmdType.meter ]
zombie_attr = ZombieAttr(ZombieType.ecf, child_list, ZombieUserActionType.fob, 300)
s1.add_zombie(zombie_attr)
text
suite s1
zombie ecf:fob:label,event,meter:Example: For all tasks under family “critical”, if any zombies arise then fail the job( i.e. the zombies process will exit with a failure):
python
with ecflow.Suite('s1') as s1:
with s1.add_family("critical") as crit :
child_list = [] # empty child list means apply to all child commands
for zombie_type in (ZombieType.ecf,ZombieType.path,ZombieType.user,ZombieType.ecf_pid,ZombieType.ecf_passwd,ZombieType.ecf_pid_passwd):
crit.add_zombie(ZombieAttr(zombie_type, child_list, ZombieUserActionType.fail, 300)) text
suite s1
family critical
zombie ecf:fail::
zombie path:fail::
zombie user:fail::
zombie ecf_pid:fail::
zombie ecf_passwd:fail::
zombie ecf_pid_passwd:fail::ecflow_client --alter=add zombie "ecf_pid:fail::" /s1
ecflow_client --alter=add zombie "ecf_passwd:fail::" /s1
ecflow_client --alter=add zombie "ecf_pid_passwd:fail::" /s1
Here are some further example of using --alter:
You can only add one zombie attribute of each time(ecf,path,user).
To delete a zombie attribute, please use one of:
Here are some more examples:
ecflow_client --alter=add zombie "ecf:kill:init,complete:" /suiteZ
ecflow_client --alter=add zombie "user:kill::" /suiteZ
ecflow_client --alter=add zombie "ecf:adopt:complete:" /suiteZ