Assigning to myself for further investigation.

So:
1) open attached file
2) add split viewer
3) connect image output to split viewe instead of the regular viewer
-> viewer node disappears?

From what I can remember yes.

I _do_ believe that while we are at it, we should likely permit _multiple_ viewers available for selection.

It would likely alleviate this bug (as the bug clearly is a tracking of viewport issue) and it will lay a solid foundation for color management and multiple differing viewers.

I know this might seem illogical at the moment, but it makes very good sense, especially when comparing looks and such.

Sorry Nathan - the viewer nodes remain in the node view, but the UV / Image editor pane doesn't display the output no matter what one does.

While the Viewer Node is selectable in the UV / Image editor pane, it doesn't appear to be connected to the proper output.

Another file attached that has lost the viewer node forever.

No amount of adding a new viewer node and attempting to get a display output is working.

Tested against current SVN.

It appears that viewer nodes are _completely_ broken currently. I can't get any output on recent SVNs whatsoever, so whatever the issue is, it is now looming large.

The recent regression appears induced by blendix's commit at r32122


------------------------------------------------------------------------
r32122 | blendix | 2010-09-25 09:35:02 -0700 (Sat, 25 Sep 2010) | 3 lines

Fix #23690: threading crash with compositing nodes and scopes view.


It should be noted that whatever happens here is likely extremely similar to the original bug report.

As an additional request, would it be possible to list all viewer nodes in the data list box so that we can pick between say, a split viewer node and a single viewer node? Or multiple viewer nodes in multiple different window panes?

It was broken in for a few hours but fixed again in revision 32127.

Developer note: a temporary viewer image is created as the compositing thread starts and removed again later. This is not a very reliable system an may well cause issues like this, but it's not entirely clear to me how.

Fixed another bug where viewer + split viewer would crash, don't know if that has some effect on this bug. I haven't been able to reproduce the problem from scratch yet.

I realize this is a borderline feature request Brecht, but I am wondering how difficult it would be to simply keep all of the viewer nodes around like viewer-001 etc. until the node is destroyed? This would allow us to have multiple viewer nodes - which is dearly needed for comparing work against say, three or four different looks simultaneously.

Blender's nodal compositor would seem to be the only compositor that doesn't permit this, in addition to breaking the similar semantics of the other nodes.

This is a feature request and it's not simple to implement. I think I've found the bug now in svn, it is related to the other bug where both a viewer and split viewer could output to the same image. I'm not 100% sure because reproducing this is quite unreliable since it's a threading issue, but it seems to be fixed.

Still breaking as of current SVN 33578.

Attaching a blend file that is has manifested in. Connected a viewer node and nothing happened. Delete the viewer node and re-added and it still is happening.

Assigning to self to tackle!

Fix is in svn now!