[moonchild]

> It typically offers better compression ratios than LZ4 at a slight (depending on your data) cost in speed

Per the table at [0], zstd provides only a slight improvement in compression ratio, and in exchange is about half the speed of lz4.

They both have their place.

0. https://facebook.github.io/zstd/

[hyperpape]

That table shows zstd comparing poorly at its faster settings, but at slower settings, it offers a significantly better compression ratio, albeit 3x slower decompression.

[baybal2]

LZ4 has branchless decompression, and lower cache footprint, thus it can work on low end, and non-desktop CPUs equally well.

zstd, brotly, snappy were seemingly all made with high end x86 capabilities in mind.

[_vvhw]

I also appreciate LZ4's simplicity and tiny code footprint.

zstd is brilliant as well, but in terms of code base it's a whole other beast.

[makapuf]

Yes decompression on baremetal cortex m4 is a mere hundreds of bytes, you can decompress it from flash directly to its output buffer.

[Teknoman117]

I've used it in bootloaders that have slow transfer mechanisms (uart, i2c) to get whatever speedup I can for a few hundred bytes of binary.

[wolf550e]

Google snappy is same class as lzo and lz4, not same class as brotli and zstd.

[ignoramous]

Also see Daniel Reiter Horn's DivANS built at Dropbox: https://dropbox.tech/infrastructure/building-better-compress...