Thursday, November 27, 2008

Compact, Declarative Serialization

A few posts back, I hinted at using parameterization as an alternative to metadata. I've just written a serialization interface using this technique, and a binary serializer/deserializer to demonstrate the inherent tradeoffs.

You can inspect the pair of interfaces required, and I implemented a binary serializer/deserializer pair as an example. ICompactSerializable is implemented for each serializable object. It's essentially a declarative method, which describes the sequential, internal structure of the object. It's simple and fast, since it provides native speed access to an object's fields without reflection, and no need for metadata.

Of course, the obvious downside is that clients must describe the internal structure themselves via ICompactSerializer, and refactoring must be careful about reordering the sequence of calls. The upshot is that serialization and deserialization is insanely fast as compared to ordinary reflection-driven serialization, the binary is far more compact, and clients have full control over versioning and schema upgrade without additional, complicated infrastructure (surrogates, surrogate selectors, binders, deserialization callbacks, etc.).

These serializers are very basic, but the principle is sound. My intention here is only to demonstrate that parameterization can often substitute for the typical approach of relying heavily on metadata and reflection. This is only one possible design of course, and other tradeoffs are possible.

For instance, each method in ICompactSerializer could also accept a string naming the field, which could make the Serialize call invariant to the ordering of fields, and thus more robust against refactoring; this brings the technique much closer to the rich structural information available via reflection, but without the heavy metadata infrastructure of the .NET framework.

The Serialize method of ICompactSerializable can also easily be derived by a compiler, just as the Haskell compiler can automatically derive some type class instances.

This application is so basic, that the user wouldn't even need to specify the field names manually, as the compiler could insert them all automatically. Consider how much metadata, and how much slow, reflection-based code can be replaced by fast compiler-derived methods using such techniques. Projects like NHibernate wouldn't need to generate code to efficiently get and set object properties, since full native speed methods are provided by the compiler.

No comments: