Add ScalarValue::Struct variant

[jonmmease]

Which issue does this PR close?

Closes #602

Rationale for this change

DataFusion already works well, in many cases, with Arrow StructArray columns. This PR adds a corresponding ScalarValue variant that can be used to represent a single element of a StructArray.

What changes are included in this PR?

This is currently a fairly minimal PR that adds a new ScalarValue::Struct variant, largely following the existing pattern for the ScalarValue::List variant. The Struct variant should support the same feature set as ScalarValue::List.

Let me know if there are uses of ScalarValue throughout the rest of DataFusion that should be updated to include struct support before these changes are merged. Thanks!

Are there any user-facing changes?

Yes, a new ScalarValue enumeration variant

[xudong963]

nitpick. Delete the blank line to keep the same as above.

[jonmmease]

Done in e25e500

[alamb]

Thank you @jonmmease -- this looks great and I am excited to extend DataFusion's type system support to include better handling of structs. @Igosuki and @houqp have been working to add the ability to select fields in https://github.com/apache/arrow-datafusion/pull/1006/files and it seems to me that this code sets the foundation for field access to struct types as well.

There appears to be a small rustfmt issue preventing CI from running on this PR: https://github.com/apache/arrow-datafusion/pull/1091/checks?check_run_id=3850128421

(I think it can be resolved by running cargo fmt and checking in the result)

Otherwise, if you could just add a test for a nested struct I think this PR is ready to go

[alamb]

You might be able to write this in a more functional style (and avoid mut) using unzip: https://doc.rust-lang.org/std/iter/trait.Iterator.html#method.unzip

Something like (untested)

let (fields, scalars) = value.iter()
      .for_each(|(name, scalar)| {
            (Field::new(name, scalar.get_datatype(), false),
            scalar.clone())
        })
      .unzip();

[alamb]

to be clear I don't think this change is needed for this PR, I just wanted to point it out -- it took me quite a while to get comfortable with the functional parts of rust -- at first I thought it was just style preference, but then the longer I work in rust I realize it is one of the key mechanisms rust uses to eliminate bounds checks during iteration

[jonmmease]

Ahh, I hadn't realized Rust had an unzip. Thanks for the suggestion. Done in 459c029

[alamb]

These are very nice and thorough tests.

The only thing I didn't see a test for was struct recursion -- like

A : {
  B: "foo"
  C: "bar"
}

I wonder if you could extend the A, B, C example below to include a nested struct as well?

[jonmmease]

Good idea. Done in 62ad373

[jonmmease]

Thanks for the feedback @alamb and @xudong963! I'll do a round of updates tomorrow

[jonmmease]

(I think it can be resolved by running cargo fmt and checking in the result)

Yeah, that took care of it. Thanks!

[alamb]

Looks great @jonmmease -- thank you! I just kicked off the CI checks and if those succeed I'll plan to merge this in

[jonmmease]

Thanks @alamb. The build failed on some clippy warnings that should be fixed by 9e6472d.

[alamb]

Looks great -- thanks again for the contribution @jonmmease !