Switch from YAML to JSON for build pipeline

[mbollmann]

YAML serialization (even with CDumper) is significantly slower than JSON serialization with msgspec in my testing (by a factor of at least 20); since Hugo also supports JSON for data files, we should probably switch the build pipeline to write JSON files instead.

[mjpost]

Interesting; why would this be the case? I have noticed that YAML support is often secondary whereas JSON is often native, and the YAML parsing library is more difficult to use, but I wouldn't have thought there would be speed orders of this magnitude.

[mjpost]

I added this to a milestone just so that it doesn't disappear into our massive issues list...

[mbollmann]

Interesting; why would this be the case? I have noticed that YAML support is often secondary whereas JSON is often native, and the YAML parsing library is more difficult to use, but I wouldn't have thought there would be speed orders of this magnitude.

I'm not sure about the reasons; I think the YAML format has more features at least, but maybe also more ambiguity?
In any case, this is consistent with benchmark results I could find on a quick search; for example:

• A write benchmark on a C++ library shows a factor of ~80 between JSON and YAML: https://github.com/getml/reflect-cpp/blob/main/benchmarks/img/licenses_write.png
• This Rust benchmark is for reading, but even here YAML is 738% slower than the JSON implementation: https://github.com/Canop/bench-config-deserializers?tab=readme-ov-file#results

[mjpost]

Interesting. I suppose JSON parsing has benefited from of ton of focus it's received, being at the heart of web browser data representations. Still, I would never have guessed this. But this is good news, in part because on #4146, I implemented a JSON representation for the issue and was going to switch to YAML, which requires loading an extra library and making a few code changes. I'll stick to JSON.

It would be fantastic to have the build speed brought down.

[akoehn]

I'm not sure about the reasons; I think the YAML format has more features at least, but maybe also more ambiguity?

YAML depends on intendation, so whitespaces have to be counted to create the correct structure. json does not rely on that and therefore probably does not need as much context. Additionally, json is simply everywhere, especially in server to server communication, whereas I have never seen huge yaml structures in production or as a data structure for data transfer.

There is a nice video going into detail what all goes into fast json parsing (a few years old already)

Switching to json sounds like a no-brainer to me, as human readability is not that important in our use case.