View Source

A zombie is a running job that fails authentication when communicating with the ecflow_server

How are zombies created ?

There are wide variety of reasons why a zombie is created.

The most common causes are due to user action:

The node tree is deleted, replaced or reloaded whilst jobs are running
A task is rerun, whilst in a submitted or active state
A job is forced to new state, i.e. complete

More rarer causes might be:

ecf script errors, where we have multiple calls to init and complete child command s
The child command s in the ecf script are placed in the background. In this case order in which the child command contact the server, may be indeterminate.
Load leveller submitting a job twice
Server crash and recovered check point file is out of date
Machine crash

How can zombie’s be handled ?

The default behaviour for init, complete, abort and wait child commands, is to block the job, and for event, label, meter to fob. (from version 4.0.4, previously all zombie, child commands, blocked)

The child command continues attempting to contact the ecflow_server.

This is done for period of 24 hours. (This period is configurable see ECF_TIMEOUT on ecflow_client).

The jobs can also configured, so that if the server denies the communication, then

the child command can be set to fail immediately. (See ECF_DENIED on ecflow_client)

ecflow_ui provides a tab which lists all the zombies and the actions that can be taken.

The zombies tab is shown, in the info panel when the server node( i.e top most) is selected.

The actions include:

Terminate:
The child command is asked to fail.
Depending on your scripts,this may cause the abort child command to be called.
Which again will be flagged as a zombie.
Fob:
Allow the job to continue. The child command completes and hence no longer blocks the job.
Great care should be taken when this action is chosen.
If we have two jobs running, they may cause data corruption.
Even when we have a single job, issues can arise.
i.e. if the associated command was an event child command, then the
event would not be set. If this event was used in a trigger expression,it would never evaluate.
Delete:
Remove the zombie from the server. The job will continue blocking, hence
when the child command next contacts the ecflow_server, the zombie will re-appear.
If the job is killed manually, then this option can be used.
Rescue:
Adopt the zombie and update the node tree.
The ECF_PASS on the zombie is copied over to the task, so that the next child command will continue as normal.
This should only be used when the user is sure there are no additional jobs.
Kill:
Applies the kill command (ECF_KILL_CMD ) using the process id stored on the zombie.
If the script has correct signal trapping, this should end up calling abort.
Note: path zombies will need to be killed manually.

Of the four action above, only Rescue will allow child command to change the state of the node tree.

What to do

Create a zombie by starting a task, and setting it to complete immediately via ecflow_ui
Inspect the log file, it will show you how the zombie has arisen.
Inspect the zombie tab in ecflow_ui (select the host node, then select the zombies tab)
Experiment with the different actions on the zombie