[surajrmal]

Or compute. Compression isn't free, especially on power constrained devices. At scale power and compute also have real cost implications. Most data centers have long been using binary encoding formats such as protobuf to save on compute and network bandwidth. cbor is nice because it's self describing so you can still understand it without a schema, which is a nice property people like about json.

[8n4vidtmkvmk]

Doesn't capn proto win hands down on compute?

I haven't used it, but I thought that was the big claim.

[kentonv]

Not necessarily.

Cap'n Proto serialization can be a huge win in terms of compute if you are communicating using shared memory or reading huge mmaped files, especially if the reader only cares to read some random subset of the message but not the whole thing.

But in the common use case of sending messages over a network, Cap'n Proto probably isn't a huge difference. Pushing the message through a socket is still O(n), and the benefits of compression might outweigh the CPU cost. (Though at least with Cap'n Proto, you have the option to skip compression. Most formats have some amount of compression baked into the serialization itself.)

Note that benchmarks vary wildly depending on the use case and the type of data being sent, so it's not really possible to say "Well it's N% faster"... it really depends. Sometimes Protobuf wins! You have to test your use case. But most people don't have time to build their code both ways to compare.

I actually think Cap'n Proto's biggest wins are in the RPC system, not the serialization. But these wins are much harder to explain, because it's not about speed, but instead expressiveness. It's really hard to understand the benefits of using a more expressive language until you've really tried it.

(I'm the author of Cap'n Proto.)

[Zardoz84]

gzip, deflate, brotli ?