rudb_bind/

statement.rs

//! From an `Ast` to a `Bound`, which is a statement rather than a query.
//!
//! A `SELECT` binds to a [`Plan`] and nothing else, and that is why [`bind`](crate::bind) can hand
//! one back. `CREATE TABLE`, `DROP TABLE` and `INSERT` are not plans and are deliberately not being
//! made into plans. A `Node::CreateTable` would be a node with no columns, no rows, no cost and no
//! reason to be pushed past anything, which is to say a node the optimizer has to be told to leave
//! alone and the executor has to special case at the root. `spec/09-optimizer.md` section 9.1 says
//! every node in a plan produces rows, and a DDL statement does not, so it goes beside the plan and
//! not inside it.
//!
//! What each variant carries is the statement with every name and type already resolved, so the
//! thing that runs it does catalog calls and nothing else. An `INSERT` in particular arrives with
//! a plan whose output is exactly the target's columns in the target's order and the target's
//! types, with the casts and the nulls for unmentioned columns already in it, so appending is a
//! loop over chunks.

use rudb_catalog::{Catalog, Entry, QualifiedName, duplicate_check, same_name};
use rudb_common::bounds::End;
use rudb_common::{
    Bound as ColumnBound, Clustering, Error, Field, LogicalType, Result, Session, Stat, Value,
    Width,
};
use rudb_parse::ast::{self, Ast};
use rudb_parse::{NONE, deparse, parse_ast};
use rudb_plan::{Expr, ExprRef, Node, Plan, SortKey};

use crate::binder::Binder;
use crate::parameters::Parameters;

/// One statement, bound.
///
/// Not `#[non_exhaustive]`. A new variant here is a new kind of statement, and the compiler
/// pointing at every place that has to decide what to do with it is the whole value of the enum.
#[derive(Debug)]
pub enum Bound {
    /// A query, which is the only one of these that produces rows.
    Query(Plan),
    /// `CREATE TABLE`.
    CreateTable(CreateTable),
    /// `CREATE VIEW`.
    CreateView(CreateView),
    /// `DROP TABLE` or `DROP VIEW`.
    DropTable(DropTable),
    /// `INSERT INTO`.
    Insert(Insert),
    /// `SET name = value`, or `RESET name`, which is the same thing with no value.
    Setting(Setting),
    /// Flushes a persistent database snapshot.
    Checkpoint,
    /// `EXPLAIN` over a query, holding the plan of the query rather than the query.
    ///
    /// The same `Plan` a [`Bound::Query`] would have carried, bound the same way and by the same
    /// code. What makes it an explain is that the layer above optimizes it and prints it instead
    /// of running it, which is the point: a plan that was built differently because somebody asked
    /// to see it is not the plan that runs.
    ///
    /// With `analyze` set the layer above runs it as well and prints what happened on it. Still the
    /// same plan, for the same reason.
    ///
    /// With `statistics` set it prints what the planner knew as well, which is the use and the class
    /// behind every number in the plan. That one changes nothing about the plan or the run either.
    Explain { plan: Plan, analyze: bool, statistics: bool },
}

/// A bound `SET` or `RESET`.
///
/// The value is a [`Value`] rather than an expression, because every setting there is takes a
/// string or a number and nothing that runs one wants a plan. What a setting does with the value it
/// gets is the setting's own business and is decided a layer up, since the binder has no idea what
/// settings exist.
///
/// The narrow part of that is that the value has to already be a constant. `SET threads = 2 + 2` is
/// four in DuckDB and is refused here, because folding it needs the expression rewriter and the
/// rewriter is two layers above the binder. Nothing writes arithmetic in a `SET` and the refusal
/// says what it is, so this waits for a reason to move.
#[derive(Debug)]
pub struct Setting {
    /// The setting name, as written.
    pub name: String,
    /// The scope word, if one was written.
    pub scope: ast::Scope,
    /// The value, or `None` for a `RESET`.
    pub value: Option<Value>,
    /// Whether the statement was written as a bare `PRAGMA name`, which carries its value in it.
    pub pragma: bool,
}

/// A bound `CREATE TABLE`.
#[derive(Debug)]
pub struct CreateTable {
    /// The full name the table gets.
    pub name: QualifiedName,
    /// The columns, in order, with the types already resolved. For a `CREATE TABLE AS` these are
    /// the query's output types under whatever names the statement or the query gave them.
    pub columns: Vec<Field>,
    /// The query to fill it from, for a `CREATE TABLE AS`.
    pub source: Option<Plan>,
    /// Whether an existing table of that name is left alone rather than being an error.
    pub if_not_exists: bool,
    /// Whether an existing table of that name is dropped first.
    pub or_replace: bool,
}

/// A bound `CREATE VIEW`.
///
/// The body is the text that was written rather than the plan it bound to. It was bound once on the
/// way through here, which is what refuses a view over a table that is not there, and the plan that
/// came out of that is then thrown away, because a view follows the tables underneath it and a plan
/// cannot. See [`rudb_catalog::View`].
#[derive(Debug)]
pub struct CreateView {
    /// The full name the view gets.
    pub name: QualifiedName,
    /// The body, as written.
    pub sql: String,
    /// The whole statement written back out, which is what `duckdb_views()` reports as `sql`.
    ///
    /// Written here because this is the last place the tree is in reach. See
    /// [`rudb_catalog::View::statement`] for what the column is and why it is not the text.
    pub statement: String,
    /// The column names the statement gave, which rename a prefix of what the body produces.
    pub aliases: Vec<String>,
    /// Whether an existing entry of that name is left alone rather than being an error.
    pub if_not_exists: bool,
    /// Whether an existing entry of that name is dropped first.
    pub or_replace: bool,
    /// The columns binding the body produced, after the alias list was applied.
    ///
    /// Worked out here because this is where the body is bound, and carried to the catalog because
    /// that is where `duckdb_columns()` and `duckdb_views()` read it from. See the doc on
    /// `rudb_catalog::View` for why the catalog keeps a list it will have to refresh later.
    pub columns: Vec<Field>,
}

/// A bound `DROP TABLE` or `DROP VIEW`.
#[derive(Debug)]
pub struct DropTable {
    /// The tables or views to drop, already resolved. With `IF EXISTS` a name that does not resolve
    /// is not in here at all, which is what makes running this a sequence of drops that cannot
    /// fail for being missing. Dropping one of these as the wrong type still can, because `DROP
    /// TABLE IF EXISTS v` where `v` is a view is an error in DuckDB and was measured to be one.
    pub names: Vec<QualifiedName>,
    /// Which of the two the statement said it was dropping.
    pub kind: Entry,
}

/// A bound `INSERT`.
#[derive(Debug)]
pub struct Insert {
    /// The table to append to.
    pub name: QualifiedName,
    /// The rows to append. The output is the table's columns, in the table's order, with the
    /// table's types, so nothing between here and the append has a decision left to make.
    pub source: Plan,
}

/// Binds one parsed statement against a catalog.
///
/// # Errors
///
/// If the script does not hold exactly one statement, if a name does not resolve, if a type does
/// not work out, or if the statement uses something that is not bound yet.
pub fn bind_statement(ast: &Ast, catalog: &Catalog) -> Result<Bound> {
    bind_statement_with(ast, catalog, &Parameters::new(), &Session::new())
}

/// Binds one parsed statement against a catalog, with values for its parameters and its settings.
///
/// This is the prepared statement path. The statement is parsed once and bound once per set of
/// values, so a parameter is a constant by the time the plan exists and everything after the binder
/// sees an ordinary query. That is why there is no parameter in `rudb_plan::Expr`.
///
/// # Errors
///
/// Everything [`bind_statement`] reports, plus an error for a parameter that was given no value.
pub fn bind_statement_with(
    ast: &Ast,
    catalog: &Catalog,
    parameters: &Parameters,
    session: &Session,
) -> Result<Bound> {
    let statement = match ast.statements.as_slice() {
        [statement] => *statement,
        [] => return Err(Error::binder("no statement to bind")),
        _ => return Err(Error::not_implemented("a script of more than one statement")),
    };
    match statement {
        ast::Statement::Query(query) => {
            let mut binder = Binder::with(catalog, parameters, session);
            let (root, _) = binder.bind_query(ast, query)?;
            Ok(Bound::Query(finish(binder, root)?))
        }
        ast::Statement::CreateTable(index) => {
            create_table(ast, catalog, parameters, session, index)
        }
        ast::Statement::CreateView(index) => create_view(ast, catalog, parameters, session, index),
        ast::Statement::DropTable(index) => drop_table(ast, catalog, index),
        ast::Statement::Insert(index) => insert(ast, catalog, parameters, session, index),
        ast::Statement::Set(index) | ast::Statement::Reset(index) => {
            setting(ast, catalog, parameters, session, index)
        }
        ast::Statement::Checkpoint => Ok(Bound::Checkpoint),
        ast::Statement::Explain { query, analyze, statistics } => {
            let mut binder = Binder::with(catalog, parameters, session);
            let (root, _) = binder.bind_query(ast, query)?;
            Ok(Bound::Explain { plan: finish(binder, root)?, analyze, statistics })
        }
    }
}

/// Parses and binds one statement, which is the whole front end in one call.
///
/// # Errors
///
/// Anything the parser or the binder reports.
pub fn bind_statement_sql(sql: &str, catalog: &Catalog) -> Result<Bound> {
    let ast = parse_ast(sql)?;
    bind_statement(&ast, catalog)
}

/// Roots a binder's plan and checks it.
fn finish(binder: Binder<'_>, root: rudb_plan::NodeRef) -> Result<Plan> {
    let mut plan = binder.into_plan();
    plan.set_root(root);
    plan.validate()?;
    Ok(plan)
}

fn create_table(
    ast: &Ast,
    catalog: &Catalog,
    parameters: &Parameters,
    session: &Session,
    index: ast::CreateTableRef,
) -> Result<Bound> {
    let written = ast.create_table(index);
    let parts: Vec<&str> = ast.name(written.name).collect();
    let name = if written.temporary {
        catalog.resolve_for_create_temporary(&parts)?
    } else {
        catalog.resolve_for_create(&parts)?
    };
    let defs = ast.column_defs(written.columns);
    let (columns, source) = if written.query == NONE {
        let mut columns = Vec::with_capacity(defs.len());
        for def in defs {
            let text = ast.string(def.ty);
            if text.is_empty() {
                return Err(Error::binder(format!(
                    "Column \"{}\" was declared without a type",
                    ast.string(def.name)
                )));
            }
            let ty = LogicalType::parse(text)?;
            let column = ast.string(def.name);
            columns.push(if def.not_null {
                Field::required(column, ty)
            } else {
                Field::new(column, ty)
            });
        }
        (columns, None)
    } else {
        let mut binder = Binder::with(catalog, parameters, session);
        let (root, scope) = binder.bind_query(ast, written.query)?;
        if defs.len() > scope.len() {
            // DuckDB's sentence, typo and all. A column list shorter than the query is fine and
            // renames a prefix, so only this direction is an error.
            return Err(Error::binder("Target table has more colum names than query result."));
        }
        let mut columns = Vec::with_capacity(scope.len());
        for (at, column) in scope.columns.iter().enumerate() {
            let named = match defs.get(at) {
                Some(def) => ast.string(def.name).to_string(),
                None => column.name.clone(),
            };
            columns.push(Field::new(named, column.ty.clone()));
        }
        if defs.is_empty() {
            deduplicate(&mut columns);
        }
        (columns, Some(finish(binder, root)?))
    };
    duplicate_check(&columns)?;
    Ok(Bound::CreateTable(CreateTable {
        name,
        columns,
        source,
        if_not_exists: written.if_not_exists,
        or_replace: written.or_replace,
    }))
}

/// Renames the columns a query repeated, which is what makes `CREATE TABLE t AS SELECT 1 AS a, 2 AS
/// a` a table rather than an error.
///
/// A query is allowed to produce two columns of one name and `SELECT 1 AS a, 2 AS a` prints two
/// columns called `a`, so a statement that turns a query into a table has to decide what to do with
/// that, and DuckDB renames rather than refusing. The suffix is `_1`, then `_2`, counting up until
/// the name is free, so a query that already has an `a_1` in it pushes the renamed column to `a_2`
/// rather than colliding with it.
///
/// This only runs when the statement wrote no column list. With a list, even a short one, duckdb
/// v1.4.1 takes the names as they come and a repeat is an error, so `CREATE TABLE t (z) AS SELECT 1
/// AS a, 2 AS a` is a table of `z` and `a` and adding a third `a` to that query is a refusal.
fn deduplicate(columns: &mut [Field]) {
    for at in 0..columns.len() {
        let taken = |name: &str, upto: usize, columns: &[Field]| {
            columns[..upto].iter().any(|held| same_name(&held.name, name))
        };
        if !taken(&columns[at].name, at, columns) {
            continue;
        }
        let mut suffix = 1;
        let mut candidate = format!("{}_{suffix}", columns[at].name);
        while taken(&candidate, at, columns) {
            suffix += 1;
            candidate = format!("{}_{suffix}", columns[at].name);
        }
        columns[at].name = candidate;
    }
}

/// Binds a `CREATE VIEW`, which means binding the body and then throwing the plan away.
///
/// Throwing it away is the point. The body is bound here so that a view over a table that is not
/// there is refused now rather than at the first select, and so that the column list can be checked
/// against what the body actually produces. What the catalog keeps is the text, because a view
/// follows the tables underneath it and a plan is a photograph of the day it was built.
fn create_view(
    ast: &Ast,
    catalog: &Catalog,
    parameters: &Parameters,
    session: &Session,
    index: ast::CreateViewRef,
) -> Result<Bound> {
    let written = ast.create_view(index);
    let parts: Vec<&str> = ast.name(written.name).collect();
    let name = if written.temporary {
        catalog.resolve_for_create_temporary(&parts)?
    } else {
        catalog.resolve_for_create(&parts)?
    };
    let aliases: Vec<String> = ast.name(written.columns).map(str::to_string).collect();

    let mut binder = Binder::with(catalog, parameters, session);
    let (_, mut scope) = binder.bind_query(ast, written.query)?;
    if aliases.len() > scope.len() {
        return Err(Error::binder("More VIEW aliases than columns in query result"));
    }
    if !aliases.is_empty() {
        let written: Vec<&str> = aliases.iter().map(String::as_str).collect();
        scope.rename(&written, "unnamed_subquery")?;
    }

    Ok(Bound::CreateView(CreateView {
        name,
        sql: ast.string(written.sql).to_string(),
        statement: deparse::create_view(ast, index),
        aliases,
        if_not_exists: written.if_not_exists,
        or_replace: written.or_replace,
        columns: scope.fields(),
    }))
}

fn drop_table(ast: &Ast, catalog: &Catalog, index: ast::DropTableRef) -> Result<Bound> {
    let written = ast.drop_table(index);
    let kind = if written.view { Entry::View } else { Entry::Table };
    let mut names = Vec::new();
    for &name in ast.name_list(written.names) {
        let parts: Vec<&str> = ast.name(name).collect();
        // The statement said which of the two it meant, so a name that is not there is a missing
        // one of those and not a missing table.
        match catalog.resolve_as(&parts, kind) {
            Ok(resolved) => names.push(resolved),
            Err(error) if written.if_exists => drop(error),
            Err(error) => return Err(error),
        }
    }
    Ok(Bound::DropTable(DropTable { names, kind }))
}

/// Binds a `SET` or a `RESET`, which is resolving its value and nothing else.
///
/// The name is not checked here. The binder knows what tables exist and has no idea what settings
/// exist, since a setting is a knob on the engine rather than an entry in a catalog, and a version
/// of this that held the list would be the binder holding a copy of something it cannot enforce.
fn setting(
    ast: &Ast,
    catalog: &Catalog,
    parameters: &Parameters,
    session: &Session,
    index: ast::SettingRef,
) -> Result<Bound> {
    let written = ast.setting(index);
    let name = ast.string(written.name).to_string();
    let value = if written.value == NONE {
        None
    } else {
        let mut binder = Binder::with(catalog, parameters, session);
        let bound = binder.bind_setting_value(ast, written.value)?;
        let Expr::Constant(value) = *binder.plan().expr(bound) else {
            return Err(Error::not_implemented(format!(
                "a value for {name} that is not a constant"
            )));
        };
        Some(binder.plan().value(value).clone())
    };
    Ok(Bound::Setting(Setting { name, scope: written.scope, value, pragma: written.pragma }))
}

/// Sorts an insert's rows into the order the target table declared.
///
/// Returns the input unchanged when the statement supplies none of the declared columns, because
/// every one of them is then a constant null and sorting on a constant is a sort that buys nothing
/// and costs a pass. A statement that supplies some of them sorts on those: the declaration is
/// about the order the rows are written in, and the columns that are there still order them.
///
/// The leading key carries the width. `date_trunc('month', d)` and `d` sort the same rows into the
/// same fragments for any predicate a month wide or wider, and the difference is what happens
/// inside a month: bucketed, the second key orders the whole month, which is the key locality the
/// joins want and the reason the width is part of the declaration at all.
fn clustered(
    binder: &mut Binder<'_>,
    input: rudb_plan::NodeRef,
    scope: &crate::scope::Scope,
    clustering: &Clustering,
    targets: &[usize],
    fields: &[Field],
) -> Result<rudb_plan::NodeRef> {
    let mut keys: Vec<SortKey> = Vec::with_capacity(clustering.columns().len());
    for (at, &column) in clustering.columns().iter().enumerate() {
        let Some(from) = targets.iter().position(|&target| target == column as usize) else {
            continue;
        };
        let source = &scope.columns[from];
        let expr = binder.plan_mut().add_expr(Expr::Column(source.binding), source.ty.clone());
        // Cast to the column's own type before bucketing, since the source of a load is a file
        // whose date column can arrive as a timestamp and `date_trunc` gives back the type it was
        // handed. Sorting on a different type than the column stores would still be an order, but
        // it would not be the order the declaration names.
        let expr = binder.checked_cast_to(expr, &fields[column as usize].ty, false)?;
        let expr =
            if at == 0 { bucketed(binder, expr, clustering.width(), fields, column) } else { expr };
        keys.push(SortKey { expr, descending: false, nulls_first: false });
    }
    if keys.is_empty() {
        return Ok(input);
    }
    let keys = binder.plan_mut().add_sort_keys(&keys);
    Ok(binder.plan_mut().add_node(Node::Sort { input, keys }))
}

/// The declaration with an automatic width turned into the bucket the incoming rows ask for.
///
/// A declaration that named no width says the bucket should come from how many rows a partition
/// would hold, and this is the only place that number is in reach. The rows are the source's, not
/// the target's: a load into an empty table has a target with nothing to count, and the whole case
/// the rule exists for is the first load of a big table. So the count and the range come off the
/// source's own zones, which is the Parquet footer for a file and the directory for a table, and
/// both are already on the plan because the estimator wanted them.
///
/// Everything about this is best effort and that is by design. The three widths hold the same rows
/// and answer the same queries, so guessing wrong costs some pruning or some key locality and
/// cannot cost an answer. A source that is a join, a group by or a values list has no zones to read
/// and gets [`Width::DEFAULT`], which is what the fixed default was before the rule existed.
fn fitted(
    binder: &Binder<'_>,
    scope: &crate::scope::Scope,
    clustering: &Clustering,
    targets: &[usize],
) -> Clustering {
    if clustering.width() != Width::Auto {
        return clustering.clone();
    }
    let Some(from) = targets.iter().position(|&target| target == clustering.partition() as usize)
    else {
        return clustering.fitted(0, 0);
    };
    let source = &scope.columns[from];
    let Some(zones) = binder.plan().sole_zones() else {
        return clustering.fitted(0, 0);
    };
    // By name, and off whichever store the plan reads rather than off the one this column is bound
    // to. The binding points at the projection over the scan, since a load is a projection into the
    // target's types, and following a binding back through a projection is the optimizer's job. A
    // load reads one table or one file, so the store with bounds on it is the store the name is in.
    let Some(at) = zones.column(&source.name) else {
        return clustering.fitted(0, 0);
    };
    let rows = zones.surviving(&[]).unwrap_or(0);
    let days = span(&zones.extreme(at, End::Low), &zones.extreme(at, End::High)).unwrap_or(0);
    clustering.fitted(rows, days)
}

/// How many days a column covers, from the smallest and largest values in it.
///
/// `None` wherever the two do not make a span, which is a column that is entirely null, a store
/// that could not fold its parts into one answer, and a pair of bounds that are not the same shape.
/// All of them mean the same thing here, which is that there is nothing to divide the row count by.
fn span(low: &Stat<ColumnBound>, high: &Stat<ColumnBound>) -> Option<u64> {
    let (Stat::Known { value: low, .. }, Stat::Known { value: high, .. }) = (low, high) else {
        return None;
    };
    let days = match (low, high) {
        // A date is a day count already, which is the common case and the only exact one.
        (ColumnBound::Int(low), ColumnBound::Int(high)) => high.checked_sub(*low)?,
        // A timestamp is a count of seconds at whichever unit the column keeps, so the span is that
        // difference divided by a day's worth of them. A scale wide enough to overflow the divisor
        // is a column no calendar covers and falls out as no span at all.
        (
            ColumnBound::Scaled { unscaled: low, scale: at },
            ColumnBound::Scaled { unscaled: high, scale: to },
        ) if at == to => {
            let day = 86_400_i128.checked_mul(10_i128.checked_pow(u32::from(*at))?)?;
            high.checked_sub(*low)? / day
        }
        _ => return None,
    };
    u64::try_from(days).ok()
}

/// Wraps a sort key in the calendar bucket its declaration asked for.
fn bucketed(
    binder: &mut Binder<'_>,
    expr: ExprRef,
    width: Width,
    fields: &[Field],
    column: u32,
) -> ExprRef {
    if width == Width::Exact {
        return expr;
    }
    let unit = binder.plan_mut().add_value(Value::Varchar(width.to_string().to_lowercase()));
    let unit = binder.plan_mut().add_expr(Expr::Constant(unit), LogicalType::Varchar);
    let args = binder.plan_mut().add_expr_list(&[unit, expr]);
    let name = binder.plan_mut().intern("date_trunc");
    let ty = fields[column as usize].ty.clone();
    binder.plan_mut().add_expr(Expr::Function { name, args }, ty)
}

fn insert(
    ast: &Ast,
    catalog: &Catalog,
    parameters: &Parameters,
    session: &Session,
    index: ast::InsertRef,
) -> Result<Bound> {
    let written = ast.insert(index);
    let parts: Vec<&str> = ast.name(written.name).collect();
    let name = catalog.resolve(&parts)?;
    if catalog.entry(&name)? == Entry::View {
        // The binary's sentence, article and all. A view has no rows of its own to append to, and
        // an updatable view is a rule about rewriting the insert that neither database has.
        return Err(Error::catalog(format!("{} is not an table", name.table)));
    }
    let target = catalog.table(&name)?;
    let fields: Vec<Field> = target.columns().to_vec();
    let clustering = target.clustering().cloned();

    // Which table column each source column lands in. Without a column list that is the first n
    // columns in order, and with one it is whatever the list says, which is also the check that
    // the list names columns the table has and names none of them twice.
    let targets: Vec<usize> = if written.columns.is_empty() {
        (0..fields.len()).collect()
    } else {
        let mut targets = Vec::new();
        for column in ast.name(written.columns) {
            let at = fields.iter().position(|field| same_name(&field.name, column)).ok_or_else(
                || {
                    Error::binder(format!(
                        "Table \"{}\" does not have a column named \"{column}\"",
                        name.table
                    ))
                },
            )?;
            if targets.contains(&at) {
                return Err(Error::binder(format!(
                    "Column \"{column}\" is named twice in the same INSERT"
                )));
            }
            targets.push(at);
        }
        targets
    };

    let mut binder = Binder::with(catalog, parameters, session);
    let (root, scope) = binder.bind_query(ast, written.source)?;
    if scope.len() != targets.len() {
        return Err(Error::binder(format!(
            "Table \"{}\" has {} columns but {} values were supplied",
            name.table,
            targets.len(),
            scope.len()
        )));
    }

    // A table that declared what order its rows go in gets the sort here, under the projection
    // rather than over it, because a projection does not reorder rows and the bindings the sort
    // keys need are the ones the query just produced. This is the whole of the loader honouring
    // the declaration: the rows arrive at the writer in order and the per fragment ranges, which
    // are built from whatever order arrives, come out narrow instead of each covering the table.
    let root = match &clustering {
        None => root,
        Some(clustering) => {
            // The width is settled here and not on the table. A declaration that left the bucket to
            // the data is a standing instruction, so it stays on the table as one and every load
            // answers it with the rows that load is carrying. What the sort needs is an answer, and
            // that is what this is.
            let fitted = fitted(&binder, &scope, clustering, &targets);
            clustered(&mut binder, root, &scope, &fitted, &targets, &fields)?
        }
    };

    // The projection that makes the source look exactly like the table. Every column the statement
    // did not name becomes a null of the column's own type, so the append never has to know that a
    // column list was written at all.
    let mut exprs: Vec<ExprRef> = Vec::with_capacity(fields.len());
    let mut names = Vec::with_capacity(fields.len());
    for (at, field) in fields.iter().enumerate() {
        let expr = match targets.iter().position(|&target| target == at) {
            Some(from) => {
                let column = &scope.columns[from];
                let expr =
                    binder.plan_mut().add_expr(Expr::Column(column.binding), column.ty.clone());
                binder.checked_cast_to(expr, &field.ty, false)?
            }
            None => {
                // A typed null rather than `add_constant`, which would give it the null type and
                // make the column's type depend on whether a row happened to be inserted into it.
                let value = binder.plan_mut().add_value(Value::Null);
                binder.plan_mut().add_expr(Expr::Constant(value), field.ty.clone())
            }
        };
        exprs.push(expr);
        let interned = binder.plan_mut().intern(&field.name);
        names.push(interned);
    }
    let exprs = binder.plan_mut().add_expr_list(&exprs);
    let names = binder.plan_mut().add_name_list(&names);
    let index = binder.fresh_index();
    let root = binder.plan_mut().add_node(Node::Project { input: root, index, exprs, names });
    Ok(Bound::Insert(Insert { name, source: finish(binder, root)? }))
}