[Geotools-devel] Pickle

[Ian S. <Ian...@ar...>]

Hi,

Pickle now exceeds shapefile performance for both reading and writing.
The main method of PickleDataSource contains a shapefile vs. pickle
performance benchmark. The unit test "PickleTest", contains some
examples of usage and an example of limitations (see point 4 below). I
also committed some documentation for anyone interested.

There are several issues/ideas that the implementation process brought
to light:

1) AttributeType, and FeatureType should be constrained by the factory
pattern (Given that Feature follows this pattern). For example:
  Features is a class which provides the factory pattern.
  Instead of delegating Feature creation responsibility to a
FeatureFactory, make the FeatureType responsible.
  AttributeType att = Features.createAttribute(name,clazz);
  ...
  FeatureType schema = Features.createSchema(name,atts);
  Feature f = schema.createFeature(values);
  
2) There should be more allowable policy with AttributeTypes, including
and perhaps even specific subtypes to allow more meta data / validation
policy. Meta data can include units, editing policy, persistence
delegate, etc. Validation includes range validation, enumeration
validation, or for references to other Features, type validation, etc.

3) It would make sense to allow for FeatureCollections to be immutable,
enforce typed membership (i.e. Features must have same schema), etc.
Given the flexability and logic of the Collections framework, the
FeatureCollections framework should be modeled on this work. For
instance a FeatureCollection should be the base interface. There could
exist FeatureList and FeatureSet with the respective semantics.

4) The biggest challenge in the Feature API will be the use/management
of "shared" objects. If FeatureType A includes a FeatureType B attribute
and instance A1 and A2 both refer to the same instance of B1, this is
important information and should be preserved. Right now, the pickle
module ignores reference identities of objects and saves them as a
"value". So if every Feature contained a reference to the same Date
object, after writing and then reading back, each Feature now would
contain a unique date object. This is obviously bad in terms of memory
usage and is semantically lossy.

Food for IRC thought.

Ian