config: detect unsatisfiable tasks

wxtim commented 2 years ago

Edit:

Have generalised the issue based on the discussion below.

Cylc 7 used to identify unsatisfiable tasks, this information was provided in get_graph_raw which allowed > graph clients to represent these dependencies in a special manner.

We have lost this feature, we should try to re-implement it and:

Put the information back into get_graph_raw.

Consider additional outlets for the info e.g. a warning on config load?

Describe the bug

The way Cylc Graph displays dependencies on tasks that don't exist has changed:

Cylc 7 - non-existant task appeared, but grey. Now - non-existant task not plotted at all.

Steps to reproduce the bug

This workflow

[scheduler]
    UTC mode = True
    allow implicit tasks = True
[scheduling]
    initial cycle point = 20000101T00Z
    [[graph]]
        +PT6H/PT6H = """
            forecast[-PT6H] => forecast
        """

when graphed by cylc graph . 2000 20000101T12Z should display the following "ghost" dependency:

But doesn't.

Pull requests welcome! This is an Open Source project - please consider contributing a bug fix yourself (please read CONTRIBUTING.md before starting any work though).

Cylc Doc Change Required

May require re-write of exercise in cylc-doc/src/tutorial/scheduling/datetime-cycling

hjoliver commented 2 years ago

Interestingly, if you run that example it just shuts down without doing anything.

Under spawn-on-demand, we start up by auto-spawning any parentless tasks to start the workflow ... here there are no parentless tasks, but the parent is outside the recurrence bounds (but not outside the graph bounds), so nothing happens.

This is not incorrect, but would likely be confusing to non-experts. Extending pre-initial-ignore to individual recurrences would be one solution (I think there's an issue up for that somewhere...).

MetRonnie commented 2 years ago

Interestingly, if you run that example it just shuts down without doing anything.

Under spawn-on-demand, we start up by auto-spawning any parentless tasks to start the workflow ... here there are no parentless tasks, but the parent is outside the recurrence bounds (but not outside the graph bounds), so nothing happens.

This is not incorrect, but would likely be confusing to non-experts. Extending pre-initial-ignore to individual recurrences would be one solution (I think there's an issue up for that somewhere...).

This could prove rather confusing for users.

If the behaviour is to be left as is, could we warn about any instances of this during validation?

hjoliver commented 2 years ago

Extending pre-initial-ignore to individual recurrences would be one solution (I think there's an issue up for that somewhere...).

See #1936

Until then (or if that's complicated by not being able to universally ignore pre-recurrence dependencies) we should be able to get validation to see this.

hjoliver commented 2 years ago

Interestingly, if you run that example it just shuts down without doing anything.

BTW, Cylc 7 would start up and stay up, but stall immediately with unsatisfied prerequisites.