razeri-plugin-contract 0.15.0

// Copyright (C) 2026 Gabriel Shea
// Licensed under either of the Apache License, Version 2.0 or the MIT
// license, at your option.

//! Storage rendering trait for partition, compression, and index strategies.
//!
//! Each dialect plugin implements [`StorageRenderer`] to translate the agnostic
//! [`StoragePlan`] into native SQL fragments. All methods have default no-op
//! implementations — dialects that lack a feature simply do not override.
//!
//! Two rules bind every implementation: callers must never branch on dialect
//! name, and implementations must never panic on unsupported features — emit a
//! storage warning event and produce no SQL instead.

use crate::data_type::DataType;
use crate::storage_strategy::StoragePlan;

/// **What a table is to the load that touches it.**
///
/// A storage renderer is asked the same questions about different tables, and
/// some answers depend on which kind it is: emptying one day of a table the
/// load is about to write keeps somewhere for that day, while emptying one day
/// of the landing table the load has just read can drop it.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum TableRole {
    /// A table a load writes: the target, or a table that holds its rows.
    #[default]
    Written,
    /// The landing table a load reads one day of, staged there before the
    /// load ran. Once the load has read the day, nothing reads it again.
    Landing,
    /// The identity table beside a written table, holding one row per key.
    Identity,
}

/// Lightweight reference to a target entity for storage rendering.
///
/// Carries only what storage rendering needs, so the contract does not depend
/// on Razeri's full design model.
#[derive(Debug, Clone, Copy)]
pub struct EntityRef<'a> {
    /// Optional schema/database namespace.
    pub schema: Option<&'a str>,
    /// Entity name (without schema qualifier).
    pub name: &'a str,
    /// **What this table is to the load**: one it writes, the landing table it
    /// reads a day of, or the identity table beside a written one. Identity
    /// tables generally do not partition, and a landing table's day is done
    /// with once its load has read it.
    pub role: TableRole,
    /// Column names of the target table, in declaration order. Required by
    /// dialects whose index DDL must enumerate columns explicitly — for
    /// example T-SQL `CREATE NONCLUSTERED COLUMNSTORE INDEX ... (col, ...)`,
    /// which fails with `Msg 35336` if the column list is missing. Empty when
    /// the caller does not have column information; renderers that need
    /// columns must skip emission and warn rather than emit invalid DDL.
    pub columns: &'a [&'a str],
    /// The declared type of the plan's partition column, when the caller knows
    /// it.
    ///
    /// **Some platforms spell a partition differently per column type, and
    /// spell it wrong without this.** `GoogleSQL` accepts
    /// `DATE(<timestamp_column>)` and `DATE(<datetime_column>)`, and it accepts
    /// a `DATE` column only *bare* — `PARTITION BY DATE(<date_column>)` is
    /// refused with the grammar quoted back (measured on the service). A
    /// renderer that cannot tell the three apart has to guess,
    /// and one of its guesses is invalid DDL.
    ///
    /// `None` when the caller has no type information; a renderer must then
    /// emit only what is valid for every type it could be, or emit nothing.
    pub partition_column_type: Option<&'a DataType>,
}

/// **The one slice a load was told to replace**, as a statement writes it.
///
/// The value is not known when the SQL is composed — it is a different day on
/// every run — so what travels here is the text that names it: the load
/// procedure's own parameter, already marked so each procedure grammar can
/// decorate it its own way. A renderer puts `value` where its statement takes
/// the slice and never invents a literal of its own.
#[derive(Debug, Clone, Copy)]
pub struct SliceRef<'a> {
    /// The time-slice column the value is compared against.
    pub column: &'a str,
    /// The text that names the value — the bound parameter, in the spelling
    /// composed statements use. Copy it through; do not re-spell it.
    pub value: &'a str,
}

/// **How one engine replaces a single slice of a table.**
///
/// A load asked to refresh one day has to clear that day before it writes it,
/// and engines divide into two answers. Most empty it with a statement of
/// their own that runs first — a partition truncated, a partition dropped, or
/// a row-level delete narrowed to the slice. A few fold the clearing into the
/// write, so there is nothing to run first and the load's own `INSERT` carries
/// a clause instead.
///
/// Answering at all is a claim that the engine documents a verb for this.
/// Where it does not, answer `None` and Razeri writes the portable form: a
/// delete filtered to the slice, then the insert.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum SliceReplacement {
    /// Statements that empty the named slice, run in order before the load
    /// writes it. Each must be complete and terminated the way this engine's
    /// procedure bodies expect.
    Emptied(Vec<String>),
    /// A clause the load's own `INSERT` carries, which replaces the named
    /// slice with the rows that insert writes — one statement, no separate
    /// clearing. Delta Lake's `REPLACE WHERE` is the worked case.
    WrittenByInsert(String),
}

/// **A second table the retention window governs**, handed to the retain
/// renderer alongside the partitioned one.
///
/// A design can ask a load to set a bad row aside instead of refusing the
/// whole day. Those rows go to a table standing beside the target, holding the
/// row as the load read it and the day it arrived in. That table is not
/// partitioned on any engine — the day it carries is the load's own column
/// rather than the design's time dimension — so dropping an old partition
/// never reaches it, and left alone it grows for as long as the warehouse
/// runs.
///
/// So it travels here instead. Delete from it by [`Self::slice_column`]
/// against the cutoff the partitions are judged by, in the same statement list
/// or procedure body, and an operator who runs retention once keeps both
/// tables inside one window. `None` means there is no such table and there is
/// nothing to add.
#[derive(Debug, Clone, Copy)]
pub struct RetainedCompanion<'a> {
    /// The table itself, addressed the way every other table here is, so it
    /// is qualified and quoted by whatever this renderer already uses.
    pub table: EntityRef<'a>,
    /// The column holding the day each row arrived in — compare the cutoff
    /// against this one, not against the design's time dimension, which this
    /// table does not carry.
    pub slice_column: &'a str,
}

/// **The call that makes a day's partitions**, as
/// [`StorageRenderer::partition_extend_call`] answers it.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ExtendCall {
    /// The procedure, named as this engine calls it: qualified with the
    /// table's schema where [`StorageRenderer::render_partition_extend_procedure`]
    /// created it there, and bare where it did not. Razeri writes this name
    /// into the call as it stands.
    pub procedure: String,
    /// What the call passes after the day, each already spelled: the forward
    /// window, where the engine gives a parameter no default. Empty where every
    /// parameter after the day has a default.
    pub after_the_day: Vec<String>,
    /// Where the call runs.
    pub runs: ExtendRuns,
}

/// **Where the call that makes a day's partitions runs.**
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ExtendRuns {
    /// Inside the load procedure, as its first statement: the call takes part
    /// in the load's transaction, so an operator calling the load by hand
    /// gets the day made as a scheduler does. A landing table the load reads
    /// rather than writes is still made before the load, by whatever stages
    /// into it.
    InsideTheLoad,
    /// Before the load's transaction opens, from `loads.sql` and the task
    /// list: the call commits the session, so inside the load it would end the
    /// load's transaction half way.
    BeforeTheLoad,
}

/// Renders agnostic [`StoragePlan`] data into platform-native SQL fragments.
///
/// All methods have default implementations that produce no SQL. Dialects
/// override only the methods their platform supports.
pub trait StorageRenderer {
    /// Render any standalone partition function DDL (e.g. T-SQL
    /// `CREATE PARTITION FUNCTION`). Returns a list of statements emitted
    /// before the `CREATE TABLE`.
    #[must_use]
    fn render_partition_function(
        &self,
        _entity: EntityRef<'_>,
        _plan: &StoragePlan,
    ) -> Vec<String> {
        Vec::new()
    }

    /// Render any standalone partition scheme DDL (e.g. T-SQL
    /// `CREATE PARTITION SCHEME`). Returns statements emitted after the
    /// partition function and before the `CREATE TABLE`.
    #[must_use]
    fn render_partition_scheme(&self, _entity: EntityRef<'_>, _plan: &StoragePlan) -> Vec<String> {
        Vec::new()
    }

    /// Render the inline storage clause appended directly to the
    /// `CREATE TABLE` statement (e.g. `PostgreSQL` `PARTITION BY RANGE (...)`,
    /// `ClickHouse` `PARTITION BY toYYYYMM(...) SETTINGS ...`).
    ///
    /// It is written after the dialect's own
    /// [`render_table_suffix`](crate::SqlRenderer::render_table_suffix), **and
    /// must not repeat anything that answers.** A dialect whose table suffix
    /// names an engine and a sorting key puts NEITHER here: a sorting key has
    /// to begin with the inline `PRIMARY KEY`, which this function is never
    /// handed and the table suffix is. A clause that spells its own produces
    /// two engine clauses in one statement — `Code: 62` on `ClickHouse` — or a
    /// key the engine rejects at `Code: 36`. Clustering intent reaches such a
    /// sorting key on the table instead, as its clustering columns.
    #[must_use]
    fn render_table_storage_clause(
        &self,
        _entity: EntityRef<'_>,
        _plan: &StoragePlan,
    ) -> Option<String> {
        None
    }

    /// Render statements emitted after the `CREATE TABLE` (e.g. `PostgreSQL`
    /// child partition `CREATE TABLE ... PARTITION OF ...`, secondary indexes,
    /// `CLUSTER BY`).
    #[must_use]
    fn render_post_create_storage(
        &self,
        _entity: EntityRef<'_>,
        _plan: &StoragePlan,
    ) -> Vec<String> {
        Vec::new()
    }

    /// Render the maintenance procedure that extends the partitioning window
    /// (creates the partitions covering a day and the days after it).
    ///
    /// **The procedure takes the day it should cover.** A warehouse loads the
    /// day a scheduler hands it, and that day is as often behind as ahead: a
    /// backfill runs last Tuesday today. A procedure that reads the clock
    /// instead can only ever cover a window around *now*, so the one day the
    /// caller actually asked about is the one day it may not make. Take the
    /// day as the first parameter, spelled the way the load takes it, and
    /// default it to today where the engine has defaults — the procedure then
    /// still answers the deploy-time call that names no day.
    ///
    /// The forward window is a second parameter where the engine allows one,
    /// so the same call covers the named day and the days a scheduler has not
    /// reached yet.
    #[must_use]
    fn render_partition_extend_procedure(
        &self,
        _entity: EntityRef<'_>,
        _plan: &StoragePlan,
    ) -> Option<String> {
        None
    }

    /// Render the maintenance procedure that retains the partitioning window
    /// (drops or archives the oldest partition per the retention policy).
    ///
    /// The day the window is measured back from is a parameter for the same
    /// reason the extend procedure's is, and defaults to today the same way.
    ///
    /// **`also` is a second table the same window governs**, handed over when
    /// there is one. Delete from it against the same cutoff, inside whatever
    /// this method returns, so that one call keeps both tables. See
    /// [`RetainedCompanion`] for why it cannot be partitioned along with the
    /// first.
    #[must_use]
    fn render_partition_retain_procedure(
        &self,
        _entity: EntityRef<'_>,
        _plan: &StoragePlan,
        _also: Option<RetainedCompanion<'_>>,
    ) -> Option<String> {
        None
    }

    /// **The call that makes a day's partitions before a load writes it**, and
    /// where that call runs — or `None` where this engine needs no such call.
    ///
    /// Engines divide on what happens to a row whose day falls outside every
    /// partition that has been declared. Some refuse it: `PostgreSQL`
    /// declarative partitioning answers *"no partition of relation … found for
    /// row"*. Others keep somewhere for it — a range ending in a catch-all, a
    /// partition function with no boundaries yet — and the row lands, but in
    /// the one partition every day shares, so a table stored by day is one
    /// partition until somebody opens its window. Either way the day the load
    /// was handed is the day to make, and the load is the only place that knows
    /// it.
    ///
    /// Answer with the procedure [`Self::render_partition_extend_procedure`]
    /// created, and only where it is rendered at all, and say where its call
    /// runs ([`ExtendRuns`]): inside the load, as its first statement, where
    /// the call takes part in the load's transaction; or before the load's
    /// transaction opens, where the call commits the session — `MySQL`'s
    /// `REORGANIZE PARTITION` does — and inside the load would end the load's
    /// transaction half way. Razeri writes the call with the same day the load
    /// is given.
    #[must_use]
    fn partition_extend_call(
        &self,
        _entity: EntityRef<'_>,
        _plan: &StoragePlan,
    ) -> Option<ExtendCall> {
        None
    }

    /// Render this engine's own way of replacing **one slice** of a table:
    /// the day, month or other value of the time-slice column that a load was
    /// told to refresh outright.
    ///
    /// The default is `None`, which is not a gap — it is the answer for every
    /// engine that documents no verb for this, and Razeri then writes the
    /// portable form: a `DELETE` filtered to the slice, followed by the
    /// insert. Snowflake, Redshift and `DuckDB` take that form because none of
    /// the three documents a single-slice replace; saying so is better than
    /// borrowing a neighbour's grammar and calling it theirs.
    ///
    /// **Answer with a partition-level verb only when one partition holds
    /// exactly one slice.** A table partitioned by month and loaded by day has
    /// thirty slices inside each partition, so truncating or dropping the
    /// partition that holds the named day would throw away twenty-nine days
    /// nobody asked about. Check the plan's granularity against the slice
    /// column's own grain ([`EntityRef::partition_column_type`]) and answer
    /// `None` when they do not agree — the portable delete is correct at every
    /// grain. A predicate-based verb, which names the slice rather than the
    /// partition, has no such limit.
    ///
    /// The slice arrives as text naming a bound parameter, not as a value:
    /// put it where your statement takes the slice, and where your engine
    /// needs a name rather than an expression, build the name from it the same
    /// way [`Self::render_partition_extend_procedure`] built it when the
    /// partition was created. Two spellings of one name is how a load comes to
    /// clear a partition that does not exist.
    #[must_use]
    fn render_slice_replacement(
        &self,
        _entity: EntityRef<'_>,
        _plan: &StoragePlan,
        _slice: SliceRef<'_>,
    ) -> Option<SliceReplacement> {
        None
    }

    /// **Whether the answer you just gave only runs inside a procedure**, so a
    /// caller knows where it may put it.
    ///
    /// The answers divide in two. Most are complete statements — an
    /// `ALTER TABLE … DROP PARTITION`, a `SET` and a `PREPARE` — and anything
    /// that runs statements runs them as they stand. Some are written in the
    /// procedural language the engine's procedure bodies are written in, and
    /// those are right inside the load Razeri generates and a syntax error
    /// anywhere else. Answer `true` and Razeri puts your text where it belongs
    /// and offers a plain delete to whoever wanted something to paste;
    /// answer `false` and your own text is what they are offered.
    ///
    /// Leave it alone and Razeri reads the shape of your answer instead: a
    /// statement that opens with `IF` and carries `END IF` is taken for a
    /// procedure body, and everything else for a statement. That reading is
    /// right about the engines it was written from and it cannot be right
    /// about every engine — a block spelled `IF EXISTS (…) BEGIN … END` runs
    /// on its own on one engine and inside a routine on the next, and no
    /// amount of looking at the text says which. So the engine says.
    ///
    /// Parameters:
    /// - `replacement`: the answer [`Self::render_slice_replacement`] gave for
    ///   this table. A clause the write carries goes wherever the write goes,
    ///   so `WrittenByInsert` is never a procedure body.
    #[must_use]
    fn slice_replacement_is_procedure_body(&self, replacement: &SliceReplacement) -> bool {
        let SliceReplacement::Emptied(statements) = replacement else {
            return false;
        };
        statements.iter().any(|statement| {
            let opens_a_conditional = statement
                .split_whitespace()
                .next()
                .is_some_and(|word| word.eq_ignore_ascii_case("IF"));
            opens_a_conditional && statement.to_ascii_uppercase().contains("END IF")
        })
    }
}

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

    struct DefaultRenderer;
    impl StorageRenderer for DefaultRenderer {}

    fn entity() -> EntityRef<'static> {
        EntityRef {
            schema: Some("core"),
            name: "events",
            role: TableRole::Written,
            columns: &[],
            partition_column_type: None,
        }
    }

    #[test]
    fn defaults_produce_no_sql() {
        let r = DefaultRenderer;
        let plan = StoragePlan::none();
        assert!(r.render_partition_function(entity(), &plan).is_empty());
        assert!(r.render_partition_scheme(entity(), &plan).is_empty());
        assert!(r.render_table_storage_clause(entity(), &plan).is_none());
        assert!(r.render_post_create_storage(entity(), &plan).is_empty());
        assert!(r.render_partition_extend_procedure(entity(), &plan).is_none());
        assert!(r.render_partition_retain_procedure(entity(), &plan, None).is_none());
        assert!(r.partition_extend_call(entity(), &plan).is_none());
        assert!(
            r.render_slice_replacement(
                entity(),
                &plan,
                SliceRef { column: "day", value: "$p_slice" }
            )
            .is_none()
        );
    }

    /// The default reads the shape: a conditional is a procedure body, a
    /// statement is not, and a clause the write carries never is.
    #[test]
    fn the_default_reads_the_shape_of_the_answer() {
        let r = DefaultRenderer;
        assert!(r.slice_replacement_is_procedure_body(&SliceReplacement::Emptied(vec![
            "IF to_regclass('core.events_20260917') IS NOT NULL THEN\n    TRUNCATE TABLE \
             core.events_20260917;\nEND IF;"
                .to_owned()
        ])));
        assert!(!r.slice_replacement_is_procedure_body(&SliceReplacement::Emptied(vec![
            "ALTER TABLE core.events DROP PARTITION 20260917;".to_owned()
        ])));
        assert!(!r.slice_replacement_is_procedure_body(&SliceReplacement::WrittenByInsert(
            "REPLACE WHERE day = '2026-09-17'".to_owned()
        )));
    }

    /// **An engine whose conditional runs on its own says so, and the shape is
    /// not consulted.** The text below is the one the default reads wrongly:
    /// it opens with `IF` and closes with `END IF`, and on an engine that
    /// executes a compound statement from a plain client it needs no procedure
    /// around it.
    #[test]
    fn an_engine_that_states_it_runs_alone_is_believed() {
        struct RunsAlone;
        impl StorageRenderer for RunsAlone {
            fn slice_replacement_is_procedure_body(&self, _replacement: &SliceReplacement) -> bool {
                false
            }
        }
        let conditional = SliceReplacement::Emptied(vec![
            "IF EXISTS (SELECT 1 FROM core.events WHERE day = '2026-09-17') THEN\n    DELETE FROM \
             core.events WHERE day = '2026-09-17';\nEND IF;"
                .to_owned(),
        ]);
        assert!(DefaultRenderer.slice_replacement_is_procedure_body(&conditional));
        assert!(!RunsAlone.slice_replacement_is_procedure_body(&conditional));
    }
}