How to handle Cow-like strings for partial zero-copy deserialization from JSON?

[silvanshade]

I'm trying to understand how to use zerovec for (partial) zero-copy deserialization of a data format that is specified as JSON.

Full zero-copy deserialization from the JSON is of course not possible for the usual reason of potentially needing to unescape certain strings when parsing as UTF-8.

The way I currently handle this is by deserializing all UTF-8 strings to Cow<str>, where everything is borrowed unless unescaping occurred.

As far as I understand, using zerovec here is still desirable because it would allow for more efficient deserialization and serialization from machine-friendly binary formats (which could be useful for caching data originally from the JSON, etc).

The main problem I'm running into while trying to use zerovec is that I don't see a straightforward way to represent these Cow-like strings within ZeroVec or VarZeroVec.

There is an instance Cow<str>: EncodeAsVarULE<str> which is used with VarZeroVecOwned::push but that doesn't seem very useful here because it would force me to make entire vectors owned just for a single owned string.

It almost seems like what I would want here is to have a ULE type which just encodes the string pointer representation as bytes, and use ZeroVec instead of VarZeroVec, but ULE has a Copy bound, which prevents me from making a type (like, e.g., compact_str::CompactString) that can impl Drop for the owned case to clean up the String remnants appropriately.

Maybe I'm thinking about this the wrong way or maybe zerovec just isn't appropriate here?

I suppose I could roll my own Cow-like vector types and just use zerofrom (and yoke) but I'd still miss out on more efficient deserialization.

Any advice on what to do for this scenario?

[silvanshade]

I figured out an alternate representation that avoids the problems I described above.

[Manishearth]

It almost seems like what I would want here is to have a ULE type which just encodes the string pointer representation as bytes, and use ZeroVec instead of VarZeroVec, but ULE has a Copy bound, which prevents me from making a type (like, e.g., compact_str::CompactString) that can impl Drop for the owned case to clean up the String remnants appropriately.

Yeah, this would absolutely not be safe or valid ULE.

But you don't need to. If you want to stick a vector inside a vector, use VarZeroVec<ZeroSlice>. You can write a guaranteed-valid-utf8 ZeroStrSlice VarULE type that wraps ZeroSlice<u8> if you want.