It's true that 128 is better to use than 120 but the main problem is not in the value but in the fact we never get
CurrentRunningXacts->subxid_overflow = suboverflowed; with value more than 120. This solves the problem but it doesn't seem the right way to fix the issue.
It seems to me a solution too.
Instead it's better to process suboverflowed state which is legit itself not resulting getting the crash on the Assert.
Of course it would be great to find the root of the problem.
So the discussion of "better" value doesn't seem related to the problem. It is for demonstration only.
IMHO, you could propose a patch, documenting this whole situation and proposing this workaround. I've been studying commits, and on several occasions problems have been fixed like this. But important is the documentation of the problem.
