Return iterators rather than vectors when deserializing in Rust backend

[teh-cmc]

TODO:

• Support for end-to-end iterators (dyn)
• Integrate fallible iterators
• Generate fully typed associated iterator types

---

The deserialization routines for components and datatypes in the Rust backend currently return full-on Vecs, while there is in fact nothing preventing us from returning lazy iterators.

Current:

pub trait Datatype {
    // [...]

    fn from_arrow_opt(data: &dyn ::arrow2::array::Array) -> Vec<Option<Self>>
    where
        Self: Sized;

    // [...]
}

pub trait Component {
    // [...]

    fn from_arrow_opt(data: &dyn ::arrow2::array::Array) -> Vec<Option<Self>>
    where
        Self: Sized;

    // [...]
}

impl trait is still unstable when used in a return position in traits, but it wouldn't be too much work to add an associated type, especially now that we have GATs! e.g.:

pub trait Datatype {
    type Iterator<T> = /* ... */

    // [...]

    fn from_arrow_opt(data: &dyn ::arrow2::array::Array) -> Self::Iterator<Option<Self>>
    where
        Self: Sized;

    // [...]
}

pub trait Component {
    // [...]

    fn from_arrow_opt(data: &dyn ::arrow2::array::Array) -> Self::Iterator<Option<Self>>
    where
        Self: Sized;

    // [...]
}

Of course we could also settle for a boxed trait object (Box<dyn Iterator<Item = Self>>) instead, but why do a half arsed job when it's all codegen'd anyway 🙃

[teh-cmc]

Update: we now know for a fact that Box<dyn Iterator> is not viable.

[jleibs]

This will be a blocker for 0.9 but doesn't need to block landing into main once we've cut 0.8.

[teh-cmc]

Also, we need to take care of fallibility, e.g. by using something like https://docs.rs/fallible-iterator/latest/fallible_iterator/index.html

[emilk]

Probably related: just py-build --release && ./examples/python/open_photogrammetry_format/main.py is ~100x slower than it should be (10s/frame)

[emilk]

If this helps as much as we hope, we can then revert the central commit of #2950 (the caching added in 4e6df7e)

[nikolausWest]

Let's discuss if we even need this at all. Most of the performance problems have already been fixed.

[emilk]

I think we should do this instead:

• Arrow deser: Add zero-copy happy path #3109