I think tag_time_source should just not do anything if there is no time source node. So the check should be moved there.

In D8690#211894, @Sergey Sharybin (sergey) wrote:

This requires deeper investigation about which exact dependency graph is missing time source and why. They are not added conditionally, and are required for proper animation handling. So it is not obvious to me why dependency graph without time source is even considered valid.

If you open the file from T80035 [has three scenes Scene, Scene.001 and Scene.002 -- Scene.002 being the active]
Only Scene.002 depsgraph is fully built afaict [build_view_layer], this is correct, no?

But switching scenes, we are looping over all depsgraphs [even the ones of inactive scenes] and try to tag the timesource nodes... this is wrong, no?

Btw., you can easily reproduce, just add 2 new scenes in a default blender startup, save, reopen and switch to one of the other scenes --> Boom

Ok, so this makes it sound that the depsgraph is allocated but not yet built. Don't thin there is anything wrong with that, but with the current state (after Jacques's refactor) such situation will miss the animation update when the graph is built.

So think more proper would be to always allocate time source, in the Depsgraph's constructor, or something like this.

In D8690#211909, @Sergey Sharybin (sergey) wrote:

Ok, so this makes it sound that the depsgraph is allocated but not yet built. Don't thin there is anything wrong with that, but with the current state (after Jacques's refactor) such situation will miss the animation update when the graph is built.

Not totally sure if this covers all situations, but in build_view_layer(aka "when the graph is built") the TimeSourceNode gets added, so that should take care of animation update as well then?

Existence of node does not mean it will enforce evaluation unless is tagged. The initial tags after builds are done in deg_graph_on_visible_update, which is called from AbstractBuilderPipeline::build_step_finalize with time update set to false. So I think the time updates will be missing with the current state of master branch + this patch, and a more complete fix is needed.

P1598 will fix the crash and will make it so time updates are properly done.

Probably @Jacques Lucke (JacquesLucke) would want to double-check.

In D8690#211924, @Sergey Sharybin (sergey) wrote:

P1598 will fix the crash and will make it so time updates are properly done.

I think always adding the time node is a good idea. I think we can also remove the find_time_source method then, the time_source data member is public anyway.

In D8690#211935, @Jacques Lucke (JacquesLucke) wrote:

I think we can also remove the find_time_source method then, the time_source data member is public anyway.

Fine by me as well :) Just do it separately from this fix.

@Philipp Oeser (lichtwerk), do you mind double-checking the snippet form me and committing it? ;)