Change default for `TokenStreamFactory.Feature.INTERN_FIELD_NAMES` to `false`

[cowtowncoder]

(note: 3.0 will split JsonFactory into TokenStreamFactory base for all codecs, and JsonFactory just or json -- features to be part of former)

Since there are many users concerned about downsides of interning JSON property field names (and similarly for other dataformats, as per #332, it may make sense to consider changing default setting for INTERN_FIELD_NAMES feature for Jackson 3.x.

As a background, the main measurable benefit of using String.intern() on field names (and ONLY once usually, first time a new symbol is encountered -- it is not an on-going cost) is that matching of JSON property name to Java logical property can usually be done with simple identity check.
This speeds up data-binding by ~10% for many cases (see jackson-benchmarks).

But doing this will not usually have benefits if:

1. Databinding is not used (unless caller makes some use of identity comparisons)
2. If number of symbols used is big (due to overhead by JVM for managing big number -- hundreds of thousands or millions -- of Strings off-heap)

To me it seems reasonable that for use case of ObjectMapper implicitly creating a JsonFactory (or similar factory for other formats) default should remain the same. But I have no objection to using different default to "direct" usage.
There are multiple ways to achieve this: one possibility would just be to make ObjectMapper explicitly set the default if it directly creates the factory; but not change setting if it is passed one.
And then JsonFactory base class can just default to intern()ing being disabled by default.

Other than this I am also open to possibility of changing defaults for databinding use case, if all major framework developers choose explicitly disabled settings as their baseline.
(authors/contributors of major frameworks like Spring Boot, Dropwizard, Restlet etc etc please send your notes to mailing list, or directly to developers -- I am not looking forward to "voting" here; but I will add condensed notes here for any input we receive).

[cowtowncoder]

I have worked on a performance optimization for field name handling, to try to avoid dual-lookup (first to map from byte sequence to String, then from String to property object), and an interesting side benefit here is that this should remove benefit of String.intern() at streaming level.
(it may or may not make sense within jackson-databind, but set of symbols there is bounded).

Assuming this development (addition of new variant of nextFieldName(), FieldNameMatcher abstraction) succeeds as it should (I have implemented it for majority of format modules) I will change the default setting for 3.0 so that interning is disabled by default for JsonFactory (and other format parser/generator factories).

So in the end I think there is a solution that works nicely for all use cases.

[IRus]

You can write own string pool to implement what String.intern() do. So than you may benefit from identity check, and doesn't harm overall JVM performance.

[cowtowncoder]

@IRus Jackson has (and has had) its canonicalizer(s) since first versions. This mostly helps in avoiding allocation overhead, but can not quite be used for identity checks for couple of reasons (separate one for byte- and char-backed sources, out of necessity; databind has no access; these are per JsonFactory, not global to avoid classloader problems and overflow).
It is true that technically it would be possible to implement global singleton (over per-factory one), but that is fortunately not necessary after changes. Same benefits (for databind at least) can be achieved by sort of per-nameset canonicalizer passed by databind (via deserializers) and used by parser.

[cowtowncoder]

After updating all codecs to implement nextFieldName(FieldNameMatcher), corresponding databind changes, intern()ing is now disabled by default.

And with other changes databind is typically 10-20% faster, despite not being able to use identity checks. So this ended up being win/win situation after all.

Thank you everyone who helped convince me this was worth the effort. :)