Functionality of #[serde(with = "serde_bytes")] for deserialization of &[u8] about bytes

[erickt]

Some serialization formats, like cbor, have a more efficient encoding for an array of u8s compared to an array of arbitrary types. Formats like bincode go even further and don’t even encode a tag describing what the next item is, so the deserializer doesn’t know the type of the next type. Other formats, like json, do not.

to handle this, serde Deserialize impls need to hint to the Deserializer what we expect the next type to be. The Deserializer could then use this to decide to parse the next chunk in a special way for an untagged format like bincode, or fall back to the default. serde_bytes is just a convenient way to provide this hint.

as to why serde doesn’t do this automatically, we can’t because rust doesn’t support specialization yet, so we need to default to the generic array deserialization, and provide serde_bytes to override the default.

[Xiretza]

I get that that's the idea, but I can't find the point in the serde_bytes code where this actually happens for deserializing a &[u8]. It seems to just forward straight to the default serde implementation.

[Xiretza]

I have a feeling this might just be a simple bug, I've opened #31.

[erickt]

Actually, my previous comment was wrong, and I don't think we need #31. I think what might be happening here is some confusion over reusing the same name (especially since it confused me). First, there actually is an impl in serde explicitly for &'a [u8]. I thought we implemented that for a generic &[T] but I was wrong. Second, the impl in serde_bytes::de for &'a [u8] is implementing serde_bytes::Deserialize, not serde::Deserialize. This is just wiring up the generic impls so it can be used with serde_bytes::deserialize, which takes a serde_bytes::Deserialize, not a serde::Deserialize.

[dtolnay]

#[serde(with = "serde_bytes")] on &[u8] isn't for deserialization, it is for serialization. On other types it optimizes both serialization and deserialization. The deserialize-related code in this crate related to &[u8] is just to make that attribute work without getting a compile error if you also derive Deserialize.

[Xiretza]

Ah, I see! I thought the visit_borrowed_{bytes,str}() methods on BytesVisitor were the optimization, but if deserialization of Bytes is actually the same as that of &[u8] in plain serde, I'll just document that.

[Xiretza]

Does that also mean that the entire BytesVisitor is redundant and that impl serde::Deserialize for Bytes could just forward to serde::Deserialize for &[u8]?