In a typical datastore, we can have a rather large number of leaf nodes, whose set of possible values has provably low cardinality, most notably boolean and enumerations.

A typical example would be a L2FIB entry, which has a leaf action, which can have only two values. For 400K entries that translates to overhead of ~8MB in 400K objects.

Introducing a (pre-instantiated?) cache of such leaf nodes would bring this down to a single instance for each NodeIdentifier/value combination, thus lowering our memory overhead. Typical users of this are schema aware translators, such as XML/JSON parsers and DataObject/NormalizedNode codec.