Could it be that the new server is low on disk space and that the FME Flow cleanup task is in aggressive clean up mode?

https://support.safe.com/hc/en-us/articles/25407763387277-FME-Flow-Troubleshooting-Cleanup-Tasks

Thanks for the tip ​@david_r but that should not be the issue.

The old workspace that does the same thing works 100% of the times.

But the new one that I’ve done some alternations with does not include the files(most often pictures).

One debugging step i’d consider would be to write to a defined location such as c:\temp. This will help narrow down if it is a ‘writing the file’ issue or a ‘location where the file is written’ issue

Thanks for the tip, @hkingsbury.

I’ve tried this now, but unfortunately it doesn’t work.
I think it's a strange bug in the Workbench.

It works when I activate the “old path” inside the embedded transformer — even though that path doesn’t lead to any output port.
If the part is active, everything works perfectly.
But if I deactivate it, it stops working.

So it seems like some kind of bug.
Any suggestions on how to solve it?

An approach that i’m not sure if it will work, but worth trying…..

Open a brand new, fresh workspace and copy the entire contents of the existing process into the new one. Save, publish to Flow and test

Are you using the data download service? Are the files missing from the ZIP?

I’ve seen some situations where is seems like the zip gets created before all data are actually available to be zipped. 

Perhaps enabling this “old path” changes the order in which things are processed.

Yes, we're using the download service, and it's always the same files that are missing — although I do receive the vector files. The workspace consists of two different processes: one that generates GeoJSON files and another that creates background maps.

I think you're onto something. When I tried removing the old path entirely instead of just disabling it, I got one image (it should have been four). So it's almost like the workspace misinterprets the process as completed even though it's not actually done.

You might be able to see clues of it in the log. You can try putting down a feature holder before the writer, that might help. 

There is also the option in the HTTPCaller to reduce the number of parallel requests. If you put this down to 1 (default is 25) then it may help. 

If you're able to somehow see this in the log or are able to share this with Safe then you should try and report it to them.