serde support

[anna-is-cute]

A feature that enables these Hash{Map,Set}s to be used with serde (like std::collections) would be awesome.

[hcpl]

See https://github.com/Amanieu/hashbrown/issues/9 as an alternative solution to this issue, because serde already provides:

• impl<K, V, H> Serialize for HashMap<K, V, H>
• impl<T, H> Serialize for HashSet<T, H>
• impl<'de, K, V, S> Deserialize<'de> for HashMap<K, V, S>
• impl<'de, T, S> Deserialize<'de> for HashSet<T, S>

So hashbrown only needs to expose its hasher for serde (and many other crates) to work!

Nevermind, hashbrown is a new implementation of HashMap and HashSet that is generic over any hasher as well. In that case I'll send a PR for serde support sometime today.

[clbarnes]

#14 was merged before the 0.1.5 release, but I still can't #[derive(Deserialize)] on structs containing hashbrown::HashMaps in the same way that I can hashmaps from std, fxhash, or fnv. Is this expected?

error[E0277]: the trait bound `hashbrown::HashMap<T, T>: serde::Deserialize<'_>` is not satisfied
  --> src/main.rs:32:10
   |
32 | #[derive(Deserialize)]
   |          ^^^^^^^^^^^ the trait `serde::Deserialize<'_>` is not implemented for `hashbrown::HashMap<T, T>`
   |
   = note: required by `serde::de::SeqAccess::next_element`

error[E0277]: the trait bound `hashbrown::HashMap<T, T>: serde::Deserialize<'_>` is not satisfied
  --> src/main.rs:32:10
   |
32 | #[derive(Deserialize)]
   |          ^^^^^^^^^^^ the trait `serde::Deserialize<'_>` is not implemented for `hashbrown::HashMap<T, T>`
   |
   = note: required by `serde::de::MapAccess::next_value`

[hcpl]

@clbarnes sorry for a very late response, but if your issue hasn't been resolved yet, could you please share code that produces these errors?

Is this expected?

The Deserialize impl for hashbrown::HashMap is a port from upstream implementation for std HashMap (https://docs.rs/serde/1.0.80/src/serde/de/impls.rs.html#1192-1236), so if your code worked for std, fxhash, or fnv then hashbrown should have worked out of the box too.

[clbarnes]

Sure thing!

https://github.com/clbarnes/hash_test/tree/master

As the readme says, that repo is in a state which builds and runs (demonstrating that std, fx and fnv all work); uncomment L51 on src/main.rs (which attempts to deserialize a JSON string into a hashbrown::HashMap) and it'll break.

Can't wait to use this library as soon as I can get the deserializing to work (my application is a port from JS and very lookup-heavy, hashbrown gives a 2x speedup over fx); thanks very much for your work.

[hcpl]

Oh, Serde support was made optional for hashbrown to reduce compile times for those who don't need it. So to make your code compile, just replace

hashbrown = "0.1.6"

with

hashbrown = { version = "0.1.6", features = ["serde"] }

The reason why Serde support worked for other cases is thanks to the combination of distinct factors:

• std provides two building blocks for creating hash maps: RandomState hasher implementation (which uses common Hasher/BuildHasher API) and HashMap implementation. Both are designed to be independently replaceable by third-party means.
• fnv and fxhash provide alternative hasher implementations (which use faster hashing algorithms). For convenience, they provide type aliases FnvHashMap<K, V> and FxHashMap<K, V> that use the standard HashMap implementation so that you didn't have to write HashMap<K, V, FnvBuildHasher> and HashMap<K, V, FxBuildHasher> every single time.
• hashbrown, on the other hand, provides an alternative HashMap implementation (which is faster than the standard one).
• serde's definition of Deserialize for the standard HashMap implementation is generic over any hasher. That's why your std_map, fx_map and fnv_map cases all worked. But the definition of Deserialize for hashbrown::HashMap is in hashbrown and needs to be manually switched on.

Hope the explanation is able to clear your confusion!

[clbarnes]

Got it, didn't think about the feature flag. Thanks!