Manual

When zombie s arise they can be handled manually by ecflowview. (See Zombie) or via the command line interface:

• ecflow_client –zombie_get
  This will list all the zombies in the server.
• 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.

Automated

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"
  child              ::=  "init" | "event" | "meter" | "label" | "wait" | "abort" | "complete"
  client_side_action ::=  "fob" | "fail" | "block"
  server_side_action ::=  "adopt" | "delete"
  zombie_life_time   ::=  unsigned integer  ( default:  user(300), ecf(3600), path(900)  )

• --alter command(dynamic)
       ecflow_client --alter add zombie <zombie-attribute>  <path>
       ecflow_client --later delete zombie < ecf | path | user>  <path>
  However note, the effect will only be seen, when the child command, makes the next attempt to communicate with the server.

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 from release 4.0.5 onwards)

• 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:

• alter
       ecflow_client --alter add zombie "ecf:fob:label:"  /s1

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 from release 4.0.5 onwards)

• 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:

• alter
       ecflow_client --alter add zombie "ecf:fob:label,event,meter:"  /s1

Example: For all tasks under family “critical”, if any zombies arise then fail the job:

• 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
        crit.add_zombie(ZombieAttr(ZombieType.ecf, child_list, ZombieUserActionType.fail, 300))
        crit.add_zombie(ZombieAttr(ZombieType.path, child_list, ZombieUserActionType.fail, 300))
        crit.add_zombie(ZombieAttr(ZombieType.user, child_list, ZombieUserActionType.fail, 300))
  

• text

     suite s1
       family critical
         zombie ecf:fail::
         zombie path:fail::
         zombie user:fail::

• alter
      ecflow_client --alter add zombie "ecf:fail::"    /s1
      ecflow_client --alter add zombie "path:fail::"  /s1
      ecflow_client --alter add zombie "user:fail::"  /s1

Here are some further example of using --alter:

• ecflow_client --alter add zombie "ecf:fob::"   /suiteX          # fob (init,event, meter, label,abort, complete) child commands.  This prevents zombies from blocking the script. Use with great care.
• ecflow_client --alter add zombie "ecf:fail::"   /suiteY          # fail the script straight away for any child command, in the job file.

You can only add one zombie attribute of each time(ecf,path,user).

To delete a zombie attribute, please use one of:

• ecflow_client --alter delete zombie ecf     /suiteX
• ecflow_client --alter delete zombie path   /suiteX
• ecflow_client --alter delete zombie user  /suiteX

Here are some more examples:

• Add a zombie attribute, that kills the zombie process automatically when  a init/complete child is recieved by the server. This will use whatever is defined for ECF_KILL_CMD

       ecflow_client --alter add zombie "ecf:kill:init,complete:" /suiteZ

• Add a zombie automatically kills zombies process, created out of user action.

             ecflow_client --alter add zombie "user:kill::" /suiteZ

• Add a zombie attribute that adopts all child complete zombies.

       ecflow_client --alter add zombie "ecf:adopt:complete:" /suiteZ

Semi-Automated

Sometimes zombies can arise for more obscure reasons. i.e The job sends a --init message to the server, meanwhile the server is busy(i.e processing jobs), when finally the server makes the task active, and sends a message back to the client/job, the ecflow_client has timed out. This causes the ecflow_client to send the same message again. However this time the server treats the child command as a zombie, since the task is already active. Hence we get these false zombies.

These scenario's are very rare, but tends to happen, for the following situations:

• High disk latencies  (i.e.  Check pointing  or job processing take to long. Typically happens when using virtual machines, with non local data)
• very large scripts ( i.e. in the megabytes),  can inflate the server memory, and cause job processing to take longer.
• Extremely large definitions, which are requested by many users, via the GUI. ( The download size, can be reduced, by only requesting the suite you are interested in)
• Very busy machine and/or not enough memory available. ( i.e. server is competing for the resources)
• Server is overloaded. ( this can  be visualised if you have gnuplot installed, and available on $PATH,  i.e invoke  ecflow_client --server_load=<path to the log file> )

To diagnose these cases, we need to look at the log file. Typically you will see two or more child commands (--init/complete), where the second will then be treated as a zombie.

To get round these issue you can add a variable ECF_NONSTRICT_ZOMBIES, which will reduce these false zombies.

       ecflow_client --alter add variable ECF_NONSTRICT_ZOMBIES 1 /              # adds the variable to the root/server level, and hence affect all suites on the server

       ecflow_client --alter add variable ECF_NONSTRICT_ZOMBIES 1 /suiteX      # adds the variable at the suite level,, and hence only affects this suite.