zstd is almost always a better choice than lz4 if you have a superscalar, vectorizing CPU. This is certainly the case for any datacenter, desktop, or modern laptop CPU out there (although you need to tune zstd accordingly). Even newer phone CPUs are probably better off with zstd. The only place where lz4 is truly appropriate is in embedded/very low power applications, and maybe for opportunistic compression in general purpose applications where it's not known how compressible the data will be. It's worth noting that both algorithms were created by the same person, Yann Collet.