TypeAdapterFactory INTEGER_FACTORY will change INTEGER type to LONG type about gson HOT 2 OPEN

Yes can confirm this; it looks like this was indeed introduced by 3e3266c (by accident):
There is only JsonWriter.value(long), so the adapters for all other integral primitive types (byte, short, int) actually call value(long).

Sorry for that. Maybe it would make sense to (partially) revert that commit.

It seems AtomicInteger is affected as well, but it's adapter always behaved this way, that a Long instead of the AtomicInteger is stored in the JsonPrimitive. However, this might be a good thing since it is mutable and might cause confusing behavior if it was stored in JsonPrimitive.

from gson.

Ok, so the potential issues with the current implementation are:

• Conversion to Long might be unexpected (especially when implicitly using JsonTreeWriter as described in your issue here)
  Might depend on the use case whether that is a problem for user code.
• Conversion to Long instead of keeping for example the Byte or Integer might take up more memory
  This might be especially noticeable when you have a lot of JsonPrimitives containing them. But maybe that also depends on the JVM object layout for these boxed objects. Maybe a Long is not actually that much larger than an Integer, despite long being larger than int (?), but I haven't investigated that yet.

Reverting 3e3266c (or at least parts of it) would also revert narrowing numeric conversion, e.g. 1.5f -> (byte) 1, see #2156. Though that was originally reported by me, and I am not sure if / how many users were actually negatively affected by that behavior.
That could be solved again by performing the narrowing conversion, but then boxing as Number again and then call JsonWriter#value(Number) instead of JsonWriter#value(long), but performance-wise this will likely not be ideal.

Maybe JsonWriter#value(Number) could also be refactored to first check using a new isTrustedIntegralNumberType, to avoid redundant string.equals("-Infinity") checks in these cases. Though it would have to be verified if that actually makes a noticeable performance difference.

So am I currently not sure what the best approach here would be.
@eamonnmcmanus, what do you think?

from gson.