postrust-graphql 2.1.0

//! Boolean expression input types.
//!
//! Hasura's filter argument is `where: <table>_bool_exp` -- a generated input
//! object with one field per column, one per relationship, and `_and`, `_or`
//! and `_not` for structure:
//!
//! ```graphql
//! author(where: {_and: [{id: {_gt: 2}},
//!                       {articles: {title: {_ilike: "%rust%"}}}]}) { id }
//! ```
//!
//! The shape matters more than it looks. A `JSON` argument accepts the same
//! text, but nothing downstream can read it: a client cannot generate types
//! from it, an editor cannot complete it, and a typo in an operator name
//! reaches the server instead of being refused by the client. Every case in
//! Hasura's introspection corpus turns on these types existing, and so does
//! every `graphql-codegen` run a migrating client has in its build.
//!
//! Two naming rules, both taken from what a client has already compiled
//! against rather than from taste:
//!
//! * The table's expression is `<table>_bool_exp`, named for the table
//!   exactly as it is spelled in PostgreSQL. It appears in the client's own
//!   source wherever a query declares a variable.
//! * A column's expression is named for the column's scalar type, so every
//!   `text` column in the schema shares one `String_comparison_exp`.

use crate::schema::object::TableObjectType;
use crate::schema::relationship::RelationshipField;
use crate::types::GraphQLType;
use async_graphql::dynamic::{InputObject, InputValue, TypeRef};
use std::collections::{HashMap, HashSet};

/// Comparisons every type supports.
const UNIVERSAL: &[&str] = &["_eq", "_neq", "_gt", "_gte", "_lt", "_lte"];

/// Comparisons taking a list of the column's own type.
///
/// The items are non-null: a null in `_in` is not a value the column could
/// equal -- `= NULL` is never true -- so Hasura types the list `[Int!]` and
/// refuses one rather than answering nothing.
const LIST_VALUED: &[&str] = &["_in", "_nin"];

/// Pattern matching, for text-shaped columns only.
const TEXT: &[&str] = &[
    "_like",
    "_nlike",
    "_ilike",
    "_nilike",
    "_similar",
    "_nsimilar",
    "_regex",
    "_iregex",
    "_nregex",
    "_niregex",
];

/// Containment and key tests, for `json` and `jsonb` columns only.
const JSON: &[&str] = &["_contains", "_contained_in"];

/// Tree comparisons, for `ltree` columns.
///
/// A label path is compared by where it sits in the tree rather than by
/// ordering: `_ancestor` asks whether this path is above another, `_matches`
/// whether it satisfies a query pattern. The `_any` forms take a list and ask
/// the same question of each.
const LTREE: &[(&str, &str)] = &[
    ("_ancestor", "ltree"),
    ("_ancestor_any", "ltree"),
    ("_descendant", "ltree"),
    ("_descendant_any", "ltree"),
    ("_matches", "lquery"),
    ("_matches_any", "lquery"),
    ("_matches_fulltext", "ltxtquery"),
];

/// The operator each tree comparison becomes, and what its operand is cast to.
///
/// `?` is PostgreSQL's "any of these" for label paths, and is the same symbol
/// jsonb uses for key existence -- different types, one spelling, which is why
/// the cast is not optional.
pub fn ltree_operator(comparison: &str) -> Option<(&'static str, &'static str)> {
    Some(match comparison {
        "_ancestor" => ("@>", "ltree"),
        "_ancestor_any" => ("@>", "ltree[]"),
        "_descendant" => ("<@", "ltree"),
        "_descendant_any" => ("<@", "ltree[]"),
        "_matches" => ("~", "lquery"),
        "_matches_any" => ("?", "lquery[]"),
        "_matches_fulltext" => ("@", "ltxtquery"),
        _ => return None,
    })
}

/// Spatial relations taking one other shape, for `geometry` columns.
///
/// PostGIS names them `ST_Contains` and so on; Hasura spells each in the same
/// lower-cased shape as every other comparison, and a client sends what Hasura
/// spelled.
const GEOMETRY: &[&str] = &[
    "_st_contains",
    "_st_crosses",
    "_st_equals",
    "_st_intersects",
    "_st_3d_intersects",
    "_st_overlaps",
    "_st_touches",
    "_st_within",
];

/// The PostGIS function behind each of those, and behind the ones that take
/// more than a shape.
pub fn postgis_function(operator: &str) -> Option<&'static str> {
    Some(match operator {
        "_st_contains" => "ST_Contains",
        "_st_crosses" => "ST_Crosses",
        "_st_equals" => "ST_Equals",
        "_st_intersects" => "ST_Intersects",
        "_st_3d_intersects" => "ST_3DIntersects",
        "_st_overlaps" => "ST_Overlaps",
        "_st_touches" => "ST_Touches",
        "_st_within" => "ST_Within",
        "_st_d_within" => "ST_DWithin",
        "_st_3d_d_within" => "ST_3DDWithin",
        "_st_intersects_rast" => "ST_Intersects",
        "_st_intersects_geom_nband" => "ST_Intersects",
        "_st_intersects_nband_geom" => "ST_Intersects",
        _ => return None,
    })
}

/// The name of the comparison input for a scalar.
pub fn comparison_type_name(scalar: &str) -> String {
    format!("{}_comparison_exp", scalar)
}

/// Whether a filter has no way to call this relationship.
///
/// A computed relationship with a required argument of its own: the parent row
/// is not the whole call, and a boolean expression is not a place a client can
/// write the rest. Every relationship reached by a foreign key answers false,
/// having no arguments at all.
fn uncallable_from_a_filter(relationship: &RelationshipField) -> bool {
    relationship
        .arguments
        .iter()
        .any(|(_, _, required)| *required)
}

/// The name of a table's boolean expression input.
pub fn bool_exp_type_name(base_name: &str) -> String {
    format!("{}_bool_exp", base_name)
}

/// The name of a table's aggregate boolean expression input.
///
/// `where: {articles_aggregate: {count: {predicate: {_gt: 2}}}}` -- the
/// authors with more than two articles. The question is about the set of
/// related rows rather than about any one of them, which is why it is a type
/// of its own rather than another field of the child's filter.
pub fn aggregate_bool_exp_type_name(base_name: &str) -> String {
    format!("{}_aggregate_bool_exp", base_name)
}

/// The name of one aggregate's own input inside that.
pub fn aggregate_bool_exp_function(base_name: &str, function: &str) -> String {
    format!("{}_aggregate_bool_exp_{}", base_name, function)
}

/// The name of the enum naming the columns `bool_and` and `bool_or` may take.
///
/// Spelled as Hasura spells it, which is long because it says where it came
/// from: the select-column enum, narrowed to one aggregate of one table.
pub fn aggregate_bool_exp_columns(base_name: &str, function: &str) -> String {
    format!(
        "{}_select_column_{}_aggregate_bool_exp_{}_arguments_columns",
        base_name, base_name, function
    )
}

/// The aggregates a predicate may be written over.
///
/// `count` over any table; `bool_and` and `bool_or` only where there is a
/// boolean column for them to fold, since a GraphQL enum may not be empty.
pub const AGGREGATE_PREDICATES: [&str; 3] = ["bool_and", "bool_or", "count"];

/// Build the comparison input for one scalar type.
///
/// `_is_null` takes a Boolean whatever the column is, and the list-valued
/// comparisons take a list of the column's type. Everything else takes one
/// value of it.
fn comparison_input(scalar: &str) -> InputObject {
    let mut input = InputObject::new(comparison_type_name(scalar)).description(format!(
        "Comparisons against a {} column. All fields are combined with AND.",
        scalar
    ));

    for op in UNIVERSAL {
        input = input.field(InputValue::new(*op, TypeRef::named(scalar)));
    }
    for op in LIST_VALUED {
        input = input.field(InputValue::new(*op, TypeRef::named_nn_list(scalar)));
    }
    input = input.field(InputValue::new("_is_null", TypeRef::named("Boolean")));

    // `citext` is text that compares without regard to case. It is its own
    // GraphQL scalar because a client declaring `$name: citext!` is naming the
    // column's type, but every pattern question text answers, it answers.
    if scalar == "String" || scalar == "citext" {
        for op in TEXT {
            input = input.field(InputValue::new(*op, TypeRef::named(scalar)));
        }
    }
    // A shape is compared by how it lies against another shape, not by
    // ordering: `_gt` on a polygon means nothing, and PostGIS answers every
    // real question with a function instead.
    if scalar == "geometry" || scalar == "geography" {
        if scalar == "geometry" {
            for op in GEOMETRY {
                input = input.field(InputValue::new(*op, TypeRef::named("geometry")));
            }
        } else {
            input = input.field(InputValue::new(
                "_st_intersects",
                TypeRef::named("geography"),
            ));
        }
        input = input.field(InputValue::new(
            "_st_d_within",
            TypeRef::named(format!("st_d_within_{}_input", scalar)),
        ));
        // A geometry and a geography answer different questions about the same
        // shape -- one on a plane, one on a sphere -- so a client asks for the
        // other by casting the column rather than by keeping two of them.
        input = input.field(InputValue::new(
            "_cast",
            TypeRef::named(format!("{}_cast_exp", scalar)),
        ));
        if scalar == "geometry" {
            input = input.field(InputValue::new(
                "_st_3d_d_within",
                TypeRef::named("st_d_within_geometry_input"),
            ));
        }
    }
    if scalar == "raster" {
        input = input
            .field(InputValue::new(
                "_st_intersects_rast",
                TypeRef::named("raster"),
            ))
            .field(InputValue::new(
                "_st_intersects_geom_nband",
                TypeRef::named("st_intersects_geom_nband_input"),
            ))
            .field(InputValue::new(
                "_st_intersects_nband_geom",
                TypeRef::named("st_intersects_nband_geom_input"),
            ));
    }
    if scalar == "ltree" {
        for (comparison, operand) in LTREE {
            // The `_any` forms take a list of the same thing.
            input = input.field(InputValue::new(
                *comparison,
                match comparison.ends_with("_any") {
                    true => TypeRef::named_list(*operand),
                    false => TypeRef::named(*operand),
                },
            ));
        }
    }
    if scalar == "jsonb" || scalar == "json" {
        for op in JSON {
            input = input.field(InputValue::new(*op, TypeRef::named(scalar)));
        }
        input = input.field(InputValue::new("_has_key", TypeRef::named("String")));
        input = input.field(InputValue::new(
            "_has_keys_any",
            TypeRef::named_list("String"),
        ));
        input = input.field(InputValue::new(
            "_has_keys_all",
            TypeRef::named_list("String"),
        ));
        // A document has no `LIKE`, and the text it renders as does. Casting
        // is how a client asks the pattern question of a column whose type
        // cannot answer it.
        input = input.field(InputValue::new(
            "_cast",
            TypeRef::named(format!("{}_cast_exp", scalar)),
        ));
    }
    if scalar == "jsonb" {
        // The path language, which is a query over the document rather than a
        // comparison against one: `_exists` asks whether the path selects
        // anything, `_match` whether the predicate it ends in holds.
        input = input
            .field(InputValue::new(
                "_jsonb_path_exists",
                TypeRef::named("String"),
            ))
            .field(InputValue::new(
                "_jsonb_path_match",
                TypeRef::named("String"),
            ));
    }

    input
}

/// The scalar a column is compared against.
///
/// An array column is compared against its element type: `_contains` on a
/// `text[]` takes text, not a list of lists. Everything else compares as
/// itself, including the types that carry their own PostgreSQL name -- a
/// `geometry` column takes a `geometry`, and a client that declares
/// `$area: geometry!` is naming the type this produces.
fn scalar_for(graphql_type: &GraphQLType) -> String {
    match graphql_type {
        GraphQLType::List(inner) => scalar_for(inner),
        other => other.to_string(),
    }
}

/// Every input type the boolean expressions need, ready to register, and every
/// scalar those inputs name.
///
/// The second half is not a convenience. A comparison input names scalars its
/// own table may not have a column of -- a cast from a geometry names
/// `geography`, a raster comparison names `geometry` -- and a type the schema
/// mentions and never registers makes the whole schema unbuildable. Reading
/// the names off what was actually generated is the only way that cannot drift
/// from the truth; twice it was patched case by case, and twice it broke again
/// one operator later.
///
/// The inputs come back with the comparisons first and the table expressions
/// after, though registration order does not matter to the schema builder --
/// the types refer to each other by name and are resolved once the whole set
/// is present, which is also what makes a recursive `_and: [author_bool_exp!]`
/// expressible at all.
pub fn build_inputs(
    object_types: &HashMap<String, TableObjectType>,
    relationship_fields: &HashMap<String, Vec<RelationshipField>>,
) -> (Vec<InputObject>, HashSet<String>) {
    let mut scalars: HashSet<String> = HashSet::new();
    let mut inputs = Vec::new();

    for (type_name, object) in object_types {
        let local_type_name = object.local_name.as_str();
        let bool_exp = bool_exp_type_name(local_type_name);
        let mut input = InputObject::new(&bool_exp).description(format!(
            "Filter rows of {}. Fields are combined with AND unless _or says otherwise.",
            local_type_name
        ));

        input = input
            .field(InputValue::new("_and", TypeRef::named_nn_list(&bool_exp)))
            .field(InputValue::new("_or", TypeRef::named_nn_list(&bool_exp)))
            .field(InputValue::new("_not", TypeRef::named(&bool_exp)));

        // Two fields of one name abort the process rather than returning an
        // error, and a foreign key column named after its target -- the
        // ordinary `pizza.crust references crust` -- produces exactly that
        // clash. The column wins here for the same reason it wins in the
        // object type: it is the table's own data, and the two have to agree
        // about which fields exist.
        let mut taken: HashSet<String> = HashSet::new();

        for field in &object.fields {
            if !taken.insert(field.name.clone()) {
                continue;
            }
            let scalar = scalar_for(&field.graphql_type);
            input = input.field(InputValue::new(
                &field.name,
                TypeRef::named(comparison_type_name(&scalar)),
            ));
            scalars.insert(scalar);
        }

        // A relationship is filtered by filtering the rows at its other end:
        // `where: {articles: {…}}` keeps the authors that have a matching
        // article.
        //
        // Except one this server could not call. A computed relationship is a
        // function of the parent row, and `fetch_articles(search text,
        // author_row author)` also wants a search term -- which a selection
        // writes as `get_articles(args: {search: "..."})` and a filter has
        // nowhere to write at all. Filtering on it used to reach PostgreSQL as
        // `fetch_articles(author)` and come back `function
        // public.fetch_articles(author) does not exist`, an answer about a
        // call this server composed. Hasura's answer is that the field is not
        // there: `field 'get_articles' not found in type: 'author_bool_exp'`.
        //
        // An *optional* argument is not this. The call with the row alone is
        // the call the function is written to accept, so it stays filterable.
        for relationship in relationship_fields.get(type_name).into_iter().flatten() {
            if uncallable_from_a_filter(relationship) {
                continue;
            }
            if !taken.insert(relationship.name.clone()) {
                continue;
            }
            input = input.field(InputValue::new(
                &relationship.name,
                TypeRef::named(bool_exp_type_name(&relationship.target_local_name)),
            ));
        }

        // And by asking something about the whole set: `where:
        // {articles_aggregate: {count: {predicate: {_gt: 2}}}}` keeps the
        // authors with more than two articles, which no filter on one article
        // can express. Only a relationship to many -- there is nothing to
        // aggregate over one row.
        for relationship in relationship_fields.get(type_name).into_iter().flatten() {
            if !relationship.is_list || uncallable_from_a_filter(relationship) {
                continue;
            }
            let name = format!("{}_aggregate", relationship.name);
            if object.fields.iter().any(|field| field.name == name) {
                continue;
            }
            if !taken.insert(name.clone()) {
                continue;
            }
            input = input.field(InputValue::new(
                &name,
                TypeRef::named(aggregate_bool_exp_type_name(
                    &relationship.target_local_name,
                )),
            ));
            // `count`'s predicate is an `Int` comparison whether or not any
            // column is one, and a type the schema names and never registers
            // makes the whole schema unbuildable.
            scalars.insert("Int".to_string());
        }

        inputs.push(input);
    }

    // A cast from one shape names the other's comparison input, so a schema
    // with only one of them still needs both.
    if scalars.contains("geometry") {
        scalars.insert("geography".to_string());
    }
    if scalars.contains("geography") {
        scalars.insert("geometry".to_string());
    }

    // Every scalar named by anything generated here, which is more than the
    // scalars the columns are: a comparison may name another shape's type.
    let mut named: HashSet<String> = scalars.clone();
    if scalars.contains("raster") {
        named.insert("geometry".to_string());
    }
    // The tree comparisons name the query types beside the path type.
    if scalars.contains("ltree") {
        named.insert("lquery".to_string());
        named.insert("ltxtquery".to_string());
    }

    // A document casts to the text it renders as, so a schema carrying one
    // needs the text comparisons even where no column is text. Added after
    // `named`, because `String` is a built-in and registering it as a scalar
    // of our own is a duplicate the schema refuses to build with.
    if scalars.contains("jsonb") || scalars.contains("json") {
        scalars.insert("String".to_string());
    }

    let mut comparisons: Vec<InputObject> = scalars.iter().map(|s| comparison_input(s)).collect();

    // The comparisons that take more than one value need an input of their
    // own. Registered only where a column of that shape exists, so a schema
    // without PostGIS carries none of them.
    if scalars.contains("geometry") || scalars.contains("geography") {
        for shape in ["geometry", "geography"] {
            if !scalars.contains(shape) {
                continue;
            }
            let mut within = InputObject::new(format!("st_d_within_{}_input", shape))
                .description("Within a distance of another shape.")
                .field(InputValue::new(
                    "distance",
                    TypeRef::named_nn(TypeRef::FLOAT),
                ))
                .field(InputValue::new("from", TypeRef::named_nn(shape)));
            // On a sphere the answer depends on which sphere. A plane has no
            // such question, so only the geography form carries it.
            if shape == "geography" {
                within = within.field(
                    InputValue::new("use_spheroid", TypeRef::named(TypeRef::BOOLEAN))
                        .default_value(true),
                );
            }
            comparisons.push(within);
        }
    }
    // What each shape may be compared as.
    if scalars.contains("geometry") {
        comparisons.push(
            InputObject::new("geometry_cast_exp")
                .description("Compare a geometry column as another type.")
                .field(InputValue::new(
                    "geography",
                    TypeRef::named(comparison_type_name("geography")),
                )),
        );
    }
    if scalars.contains("geography") {
        comparisons.push(
            InputObject::new("geography_cast_exp")
                .description("Compare a geography column as another type.")
                .field(InputValue::new(
                    "geometry",
                    TypeRef::named(comparison_type_name("geometry")),
                )),
        );
    }
    for document in ["jsonb", "json"] {
        if scalars.contains(document) {
            comparisons.push(
                InputObject::new(format!("{}_cast_exp", document))
                    .description(format!("Compare a {} column as another type.", document))
                    .field(InputValue::new(
                        "String",
                        TypeRef::named(comparison_type_name("String")),
                    )),
            );
        }
    }
    if scalars.contains("raster") {
        comparisons.push(
            InputObject::new("st_intersects_geom_nband_input")
                .description("Intersecting a shape, optionally in one band.")
                .field(InputValue::new("geommin", TypeRef::named_nn("geometry")))
                .field(InputValue::new("nband", TypeRef::named(TypeRef::INT))),
        );
        comparisons.push(
            InputObject::new("st_intersects_nband_geom_input")
                .description("Intersecting a shape in a given band.")
                .field(InputValue::new("nband", TypeRef::named_nn(TypeRef::INT)))
                .field(InputValue::new("geommin", TypeRef::named_nn("geometry"))),
        );
    }
    comparisons.extend(inputs);
    (comparisons, named)
}

#[cfg(test)]
mod tests {
    use super::*;
    use pretty_assertions::assert_eq;

    #[test]
    fn a_table_expression_is_named_after_the_table() {
        assert_eq!(bool_exp_type_name("author"), "author_bool_exp");
    }

    #[test]
    fn a_comparison_is_named_after_the_scalar() {
        assert_eq!(comparison_type_name("String"), "String_comparison_exp");
    }

    #[test]
    fn an_array_column_is_compared_against_its_element_type() {
        let list = GraphQLType::List(Box::new(GraphQLType::String));
        assert_eq!(scalar_for(&list), "String");
    }

    #[test]
    fn a_type_with_its_own_name_compares_as_itself() {
        let custom = GraphQLType::Custom("geography".to_string());
        assert_eq!(scalar_for(&custom), "geography");
    }
}