Mount Olympus in the Olympic National Park

Photo by bart hoevenaars

Sparked by a discussion I had with Steve, I started looking at mechanisms for safely passing data between tasks with the emphasis on neither side ever seeing changes made to the data. This led to (currently) three blog posts on my personal blog, this being work that is more of an exploration and not yet something planned for inclusion in Dream.

The first post “Sharing data without sharing data state” reviewed patterns for sharing, from cloning to immutability to serialization and finally introduced a pattern for freezing objects in .NET. I followed this up with an implementation of a “Freezer” to automatically attach IFreezable behavior to Data Transfer Objects (DTOs), AOP style. This work was described in “Freezing DTOs by proxy” and the project, Subzero,  can be found at github. The initial implementation used serialization for cloning objects as required when no Clone() method could be found on the DTO.

Next I decided to write a generic cloner for Subzero, based on my premise that serialization is expensive. This turned out to be a lesson in measuring and optimizing. My first pass was only faster than .NET’s built in serializer, BinaryFormatter, and being slower than BinaryFormatter can only be achieved by copying things to pager and typing it back in by hand. Other serializers blew my cloning out of the water. “When cloning isn’t faster than copy via serializer” describes the work of optimizing the cloner and comparisons to various serializers.

Dream 2.0 introduced the IKeyValueCache with an in-memory implementation and a plan for a Memcache wrapper. This cache uses serialization by default and let’s you easily register your favorite serializer. We’re using this cache in MindTouch Core for optimizing search performance and may use it for the MindTouch Platform data caching layer, which currently does explicit copying of entities in and out of the cache. So it would make sense to revisit IKeyValueCache and have it take advantage of IFreezable to avoid copying when possible and use Incubator.Clone when unavoidable to reduce all the manually maintained copy code.

Subzero may eventually make it into Dream, both for caching and potentially for Dream in-process REST calls, but we’ll have to see what makes sense. For right now Subzero is more an infrastructure experiment we’re evaluating for use.