`window_open` `Observable`s not notified

[lmshk]

I am creating a GLMakie window from a script and want to block that script until the window is closed (because otherwise the process exits). For this purpose, I attach a listener to the Figure's window_open Observable that notifies s a Base.Event when the value becomes false. This worked correctly before, but the listener stopped being called starting with Makie v0.18.0.

A simplified example looks like this:

using GLMakie
closed = Base.Event()
f = Figure()
display(f)
on(events(f).window_open) do is_open
    !is_open && notify(closed)
end
wait(closed)

When running this and manually closing the window, the script will remain blocked.

(I suspected that this is somehow related to #2306 but didn't confirm that because I couldn't figure out how to ]add both Makie and GLMakie at the specific commits and without error.)

After looking around a little, I noticed that there are multiple Events instances related to my plot, all of which have a window_open Observable, but I am confused about which of these I should use:

screen = begin
    f = Figure()
    display(f)
end

on(events(f).window_open) do _; @info "events(f)" end
on(screen.window_open) do _; @info "screen" end
on(events(f.scene).window_open) do _; @info "events(f.scene)" end

On Makie v0.17, when closing the window, the preceding code prints

[ Info: events(f.scene)
[ Info: events(f)
[ Info: events(f.scene)
[ Info: screen

while on v0.18 this prints just

[ Info: events(f.scene)

It seems like I can just use the scene events as a workaround, but the fact that this is triggered twice on v0.17 makes me think that I am doing something wrong. Am I meant to use it this way, or is there another Events instance I missed, or is this a bug?

[SimonDanisch]

Seems like a regression, I'll take a look...Meanwhile, you have wait(screen) for this...
so e.g.:

screen = display(fig)
wait(screen)

[SimonDanisch]

And if you use .window_open you should use the one from scene.events, since screen.window_open only exist as part of a work around and should be considered private...

[lmshk]

(Thanks for the quick response!)

waiting on the Screen is not quite enough in my case, because I am potentially rebuilding and replacing the whole Figure from a callback (when filter criteria change, which may change how the figure is structured). What I do is to keep track of the current window_open callback, and before displaying the new Figure calling off on it, then attaching a new callback after display. (See this post for my reasoning.) But this does seem kind of ugly, so maybe there's a better way to do it?

[SimonDanisch]

Can you explain, maybe with a very easy & short code example, why wait is not working in your case? I can't see the reason from the linked post...

[lmshk]

Maybe I'm missing something... in my setup, the display call happens from a callback (presumably called from some event loop), while the wait needs to happen on the main thread to prevent the process from exiting. So where I want to wait I don't have the Screen.

I'll paste the linked example here:

closed = Base.Event()
selection = Observable(...)  # this triggers and controls figure (re)construction

window_open_callback = Ref{Any}(nothing)
on(selection) do selected
    # filter data etc...
    figure = build_figure(data; selection)

    isnothing(window_open_callback[]) || off(window_open_callback[])
    display(figure)  # <- Screen available here
    window_open_callback[] = on(events(figure).window_open) do is_open
        !is_open && notify(closed)
    end
end

notify(selection)
wait(closed)  # <- Screen needed here

[lmshk]

(But my understanding is that doing that same thing on events(figure.scene).window_open should work both on v0.17 and v0.18, so this is fine for me.)

[SimonDanisch]

Can you do:

screen = GLMakie.Screen()
on(selection) do selected
    # filter data etc...
    figure = build_figure(data; selection)
    display(screen, figure)  # <- Screen available here
end
wait(screen)

?

[lmshk]

That works perfectly, thanks very much!

---

Regarding the original post, I just realized that the situation is more complex than I noticed at first. On v0.18:

f = Figure()
display(f)
on(f.scene.events.window_open) do is_open; @info "XXX" is_open end
on(f.scene.events.window_open) do is_open; @info "YYY" is_open end

Now:

• closing the window triggers only the "YYY" callback
• redisplay(f) triggers both callbacks
• closing the window again triggers only the "YYY" callback

(Of course now I don't need it anymore. 😉)

[lmshk]

Can you do:

screen = GLMakie.Screen()
on(selection) do selected
    # filter data etc...
    figure = build_figure(data; selection)
    display(screen, figure)  # <- Screen available here
end
wait(screen)

?

This worked correctly with v0.18.2, but stopped working with v0.18.3: Calling display the second time (with the first rebuilt figure) leaves the old figure on screen (now displaying both, overlapping), but resizing the window only resizes the newer figure. Calling empty!(screen) now closes the window. How should I replace the displayed figure without closing the window? (Should this be a separate issue?)

Example:

screen = GLMakie.Screen()

figure = Figure()
Label(figure[1, 1], "first")
display(screen, figure)

figure = Figure()
Label(figure[1, 1], "second")
display(screen, figure)  # does not clear old figure on v0.18.3

empty!(screen)  # closes window on v0.18.3

[SimonDanisch]

Oops... Pushed a fix for it bcfdd19
Will tag soon :)

[ffreyer]

This and #2419 don't work because

Makie.jl/src/display.jl

Lines 41 to 50 in f124a01

	deregister = on(events(scene).window_open, priority=typemax(Int)) do is_open
	# when screen closes, it should set the scene isopen event to false
	# so that's when we can remove the screen
	if !is_open
	delete_screen!(scene, screen)
	# deregister itself
	!isnothing(deregister) && off(deregister)
	end
	return Consume(false)
	end

removes the first callback from obs.listeners during the listener iteration in

https://github.com/JuliaGizmos/Observables.jl/blob/27b3fc61fed83a104cdc7b79a35f25f111558e02/src/Observables.jl#L168-L174

so it jumps to the third instead of second.