OpenZL takes a description of your data and builds from it a specialized compressor optimized for your specific format.

It provides a set of primitive codecs that can be composed in a DAG. Additionally, it allows for user-defined control flow to modify the DAG based on the data, at any point in the compression. OpenZL also provides a universal decompressor that can decompress anything produced by the compressor, independent of the compression DAG.

Conceptually very cool. Information theory says that the more one knows about a dataset’s format, the more one can compress it. If you describe the data’s encoding and overall shape using non-recursive enumerable languages, then you can safely use them as compression hints.

Instead of making twenty custom compressors for twenty different data formats, they’ve generalized it, and presumably decompression doesn’t require separate copies of the same hints used for compression.