Ineffective use of CodecCache slows down queries

[packet23]

Profiling point queries using

var entity = ds.find(Entity.class).filter(eq("id", id)).first();

shows that ~19 % of CPU time is spent caching codecs. When the query is repeated on the same datastore instance,
the cache is effectively emptied.

Flame graph of point query execution, highlighting the stack traces which show ineffective caching:

Root cause seems to be an undesired side-effect of re-setting codec registry on every query:

• Datastore.find(Class<T>) calls DefaultQueryFactory.createQuery(Datastore, Class<T>, FindOptions, Document)
• which uses the MorphiaQuery(Datastore, Class<T>, FindOptions, Document) constructor to create a query
• the constructor calls Datastore.getCollection(Class<T>) to get a reference to the collection
• which in turn calls MongoCollectionImpl.withCodecRegistry(CodecRegistry)

Result of this call chain is that the MongoDB driver creates a wrapped copy of the CodecRegistry with empty cache.

Not resetting the CodecRegistry on Datastore.getCollection(Class<T>) but instead making sure Morphia's CodecRegistry is always registered at database-level when Datastore is created restores codec cache effectiveness.