The dataset needs to be compressed with blosc2 first. The reader may not need to cache the decompressed chunks since blosc doesn't need to decompress the entire chunk (I think).