On 26/03/13 09:02, Adam Murdoch wrote:
On 26/03/2013, at 7:34 PM, Adam Murdoch <[email protected]
<mailto:[email protected]>> wrote:
On 22/03/2013, at 12:32 PM, Marcin Erdmann <[email protected]
<mailto:[email protected]>> wrote:
- I have no idea on how to achieve skipping dependencies of a
finalising task when the finalising task should be skipped - it gets
really tricky as those dependencies might be made mandatory by making
the finalising task mandatory after being added to the graph…
Add them all as finalisers of the target task when you're traversing
the graph down from a finaliser task - that is, a dependency of a
finaliser of task A is itself a finaliser of A.
Actually, this doesn't work for shared dependencies, as there's also
an ordering constraint on a finaliser. We can tweak it so that we
split finaliser relationships up into 2 parts: the ordering part which
is the same as a soft dependency, and the triggering part that
switches the task state to `must run`. The dependencies of a finaliser
don't pick up the ordering part, but they do pick up the triggering part.
- Modify DefaultTaskExecutionPlan to go into "finaliser task mode
only" after a task failure not handled by failureHandler; if in that
mode then getNextReadyAndMatching() would use a Spec that only
allows finaliser tasks, and (I don't know how to achieve that)
finaliser task dependencies
I would split the `ready` state into several: `should run` and `must
run` and `should not run`. The entry tasks and any task with an
incoming hard dependency will be in `should run`. A finaliser task
starts off as `should not run` and is switched to `must run` once any
of the tasks that it finalises is executed. Before the first failure,
the execution plan runs tasks in `should run` and `must run` state.
Following a failure, it runs only tasks in `must run` state and
continues until there are no such tasks left.
This should be easy to achieve. The only trick here will be how to put
all finaliser dependencies that are not in the graph already to `should
not run` and to mark all finaliser dependencies as finalisers. With that
it should be easy to mark all tasks that are finalisers for the executed
task that did any work from `should not run` to `must run`. Currently if
there is a failure all tasks in `ready` state are set to `skipped`, the
only change will be to now skip all in `should run` and `should not run`
states.
