[jvz]

Have you evaluated compression algorithms that support custom dictionaries, like zstd? You could generate a compression dictionary for each domain, or just for those above a certain size.