ipld
Remove the legacy codec system
Serde is significantly more usable, and supporting two codec systems is very confusing and a waste of time.
Agreed. Though I'd like to get close to feature parity, so that people using the legacy codec can smoothly upgrade.
Heya, @vmx asked me to comment on our usage of the codec approach. To be honest, my first thought was the same as yours @Stebalien but there are two things that I enjoy in the codec approach:
- Codecs know how to extract links
- We mostly use
dag-cbor, but in some cases we also encode asraw
I would love for there to be one true way, and if serializing data as blocks was as simple as deriving Serialize/Deserialize in most cases, that would be awesome :+1:
I agree that we definitely need a way to cleanly extract links. It shouldn't be too difficult to re-implement that with serde by just throwing away everything we don't care about, but the current method is more efficient.
I recently learned about the DeserializeSeed trait. Maybe this could be used to collect links while deserializing?
That's not even necessary, really. We can just deserialize into a "links collector" that:
- Collects links.
- Ignores everything else.
Ah right, that should be possible too. When working with many small blocks, the current API allows to collect all links into a single container (without allocating a links container for each block), that's why I thought of DeserializeSeed.
Maybe there's two features at play even?
- Only collect links without deserializing everything
- Deserialize into a struct and collect all links, in a single pass
The latter might not matter much in practice though.
Only collect links without deserializing everything
Deserialize into a struct and collect all links, in a single pass
The latter might not matter much in practice though.
Agreed. I'd start with the former one and add the latter only if it turns out to be useful. I prefer keeping APIs small and targeted at needs, rather then adding theoretically useful things.