[C++][Dataset] Filter expressions should not require exact type match

[asfimport]

It's not trivial for users to be able to ensure that scalars are of identical type to the fields they relate to in Expressions. For one, FieldExpressions don't contain a type reference, so at the time when I construct field_ref("col1") > scalar(42), I don't know exactly what type col1 is to be able to ensure that scalar(42) matches. Even if it were available, I wouldn't be able to determine what type to make it if the expression were (field_ref("col1") + field_ref("col2")) > scalar(42).

We should allow CompareExpressions to cast the inputs as necessary. This should be among integer types and floating point types, and across integers and floats too. Likewise among date/timestamp types, and probably if comparing a string scalar against a date/timestamp column, the string should be parsed as a datetime. We also need to think about DictionaryTypes (though in practice this is moot until we have a comparison kernels that work on strings).

@fsaintjacques @bkietz

Reporter: Neal Richardson / @nealrichardson
Assignee: Ben Kietzman / @bkietz

PRs and other links:

• GitHub Pull Request #5813

_{Note: This issue was originally created as ARROW-7047. Please see the migration documentation for further details.}

[asfimport]

Francois Saint-Jacques / @fsaintjacques:
In typical databases, this is not the responsibility of physical operators (the dataset module in this case). It is assumed that inputs are properly typed/casted by who ever generates the physical plan. I would propose that we create an interface to bind/validate/cast an expression to a Schema, e.g.

/// \brief Bind an expression to a schema
///
/// Binding will try to align types of expressions and referenced fields. It will also checks that all references are valid.
Result<Expression> Expression::Bind(const Schema& schema, BindOptions options);

This utility could be used by high level languages and the planner (execution engine). This could be exposed as an option in the ScannerBuilder interface, or an optional parameter to the AddFilter.

[asfimport]

Ben Kietzman / @bkietz:
@fsaintjacques thanks, that seems reasonable. I'll add CastExpression and Bind can insert these everywhere a type mismatch is encountered.
It seems like a direct extension of Expression::Validate, is it worth keeping them both? Bind could also output the expression's evaluated type

FieldExpressions are currently not typed but it seems reasonable to add optional typing when the field's type is known.

[asfimport]

Neal Richardson / @nealrichardson:
Issue resolved by pull request 5813
#5813