Curious, if you compared zstd against other options? Go's native zstd is "famous" to be less optimal, comparing to its C implementation. Could S2 (aka "better snappy") provide different trade-off for resource usage/speed?