The file Watchdog-20200823-2033.dmp might tell a developer what went wrong,
I think you should add it to this case as well (note i'm not a blender developer, i work on other software).

Hm not sure what have caused it, i was thinking it related to me upgrading the blender versions for the project a few times.
Another cause might be that i tried to merge in a medical human model from another blend file (that didnt work well).
Eventually though i copied the collection over ( copy past from and into the outliner with 2 running blender programs), which works a lot easier.
..Which makes me wonder then might it had to do with the merger attempt. (the files where huge in size).
..That might also explain the two scenes, perhaps. (i wasnt aware of it).

Ok i got a file, its happening again here is a screen shot{F8603304} after this one i upload a file of it (with everything removed). but i leave a cube.
On an additional node the renders did got more white as well.

its our not yet finished marketing material cannot share it, though if it happens again i'l destory as much of the file as posible (leaving or adding only a cube or so) and i'll post it.
I also updated server and client version to 2.9 of today so (i hope) that might fix it as well,
I did it as i saw multiple people had issues with 2.9. (and this feels a bit as if some old version thing got in).

I created animation in 2.83 on windows, on my windows laptop,
I render on the server on cpu. in the drawing in compositor the intel denoiser is used.
But it wasn't recognised as the denoiser filter node and was shown as an invalid red node.
Which is strange since both have the intel denoiser
So i removed the unrecognized one on our linux render serer and then added it again after which it worked.

Ok i can confirm its combined problem with windows monitor scaling,
Windows uses a recomended scale here ( note that I also wear glasses, my eyesight isnt perfect anymore ) so If i disable it, it becomes too small for my eyes.
I have to set my laptop on 125% and a philips monitor on 100%, and those seam the default for both.

I tested it with blender 2.9 as well still the same problem in that file.
So you were able to connect or alter the node setup in both node views, i'm not.

In above file, started mainwindows 3dview+nodes side by side.
Duplicated screen (didnt happen then),
Then turned the seccond monitor (not main) nodes into 3dview
Then i vertically split left side of the (not main) and turned the upper in nodes view.
Then i couldnt connect nodes in the main windows

ehm isnt that because GPU uses larger tiles, and thus has a difefrent analysis of the image as a small tile cpu evalution ?

Hm i added a mesh light outside the room and i could see trough it as well... no idea original version didnt have a mesh light outside.

I have a workaround just using normal glass so its no prio to mee, just raising the bug to notify the people working on principled BSDF, will do some more testing on original file as well.

the orginal blend file is over 700MB and contains a product in its beta from which i cannot share.
However maybe i dould reduce the blend remove everything except the glass would that be ok to you ?.

Yesterday was working with blender 2.83 and texture disapeared partly (like loosing 2 hours on texture that i worked on 3 hours) , should have been packed in the project upon save. blender recovery didn return the picture either, knowing its experimental but without bug reports this will stay problematic (using blender several years).
I know alpha is a risk but i dont mind it, just reporting it.

In edit mode vertex align is equally to set pivot to active verticle then do Scale (press x,y or z) and zero.
And that works with customs axis'es too.
The align tool itself is for objects, so its not in edit mode, not a bug.
Clarify if you think it should work differently.

It was already reported in a few cases ( https://developer.blender.org/T72644 ), i'm not sure how to join cases though

@Marco Schlumpp (marco) (nacioss) Simulations should depend on normals, particles should never be inside objects, think of collision detection.
With single thin faces with wrong normals how would one else know if we're inside or outside of it.
Essentially per frame displacement it (or subrframe sampling) collisions should be calculated, if a particle endsup inside reflect its internal distance over normal angle.

Me another user did some additional testing and likes to add some notes :

I had the problem as well with latest builds, i found a manual fix that i think would be easy to code.
Before baking set the domain origin to the world center.
Then in edit mode adjust size and position of where you want your domain to be.

On a side note, the problem is in adaptive domain handling here.
Did some tests with the file, turning it of made it work in latest win build version.