We've also had great success using zstd with training. We dump a lot of JSON data into kafka, most of which has a similar schema, and training it easily gave a 2-3x reduction in size over lz4.