risingwave_frontend/handler/

query.rs

// Copyright 2022 RisingWave Labs
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

use std::collections::HashSet;
use std::sync::Arc;
use std::time::Instant;

use itertools::Itertools;
use pgwire::pg_field_descriptor::PgFieldDescriptor;
use pgwire::pg_response::{PgResponse, StatementType};
use pgwire::types::Format;
use risingwave_batch::worker_manager::worker_node_manager::WorkerNodeSelector;
use risingwave_common::bail_not_implemented;
use risingwave_common::catalog::{FunctionId, Schema, SecretId};
use risingwave_common::id::ObjectId;
use risingwave_common::session_config::QueryMode;
use risingwave_common::types::{DataType, Datum};
use risingwave_sqlparser::ast::{SetExpr, Statement};

use super::extended_handle::{PortalResult, PrepareStatement, PreparedResult};
use super::{PgResponseStream, RwPgResponse, create_mv, declare_cursor};
use crate::binder::{Binder, BoundCreateView, BoundStatement};
#[cfg(feature = "datafusion")]
use crate::datafusion::DfBatchQueryPlanResult;
use crate::error::{ErrorCode, Result, RwError};
use crate::handler::HandlerArgs;
use crate::handler::flush::do_flush;
use crate::handler::util::{DataChunkToRowSetAdapter, to_pg_field};
use crate::optimizer::plan_node::{BatchPlanRef, Explain};
use crate::optimizer::{
    BatchPlanRoot, ExecutionModeDecider, OptimizerContext, OptimizerContextRef,
    RelationCollectorVisitor, SysTableVisitor,
};
use crate::planner::Planner;
use crate::scheduler::plan_fragmenter::Query;
use crate::scheduler::{
    BatchPlanFragmenter, DistributedQueryStream, ExecutionContext, ExecutionContextRef,
    LocalQueryExecution, LocalQueryStream,
};
use crate::session::SessionImpl;

/// Choice between running RisingWave's own batch executor (Rw) or a `DataFusion` (DF) logical plan.
pub enum BatchPlanChoice {
    Rw(RwBatchQueryPlanResult),
    #[cfg(feature = "datafusion")]
    Df(DfBatchQueryPlanResult),
}

impl BatchPlanChoice {
    pub fn unwrap_rw(self) -> Result<RwBatchQueryPlanResult> {
        match self {
            BatchPlanChoice::Rw(result) => Ok(result),
            #[cfg(feature = "datafusion")]
            BatchPlanChoice::Df { .. } => {
                risingwave_common::bail!(
                    "Expected RisingWave plan in BatchPlanChoice, but got DataFusion plan"
                )
            }
        }
    }
}

pub async fn handle_query(
    handler_args: HandlerArgs,
    stmt: Statement,
    formats: Vec<Format>,
) -> Result<RwPgResponse> {
    let session = handler_args.session.clone();
    let context = OptimizerContext::from_handler_args(handler_args);

    #[cfg(feature = "datafusion")]
    {
        // We construct a future manually here to make sure this async function is `Send`.
        // `BatchPlanChoice` is non-Send, and rust cannot prove it has dropped before await point.
        // See more details in https://github.com/rust-lang/rust/issues/128095
        use futures::FutureExt;

        use crate::datafusion::execute_datafusion_plan;

        let future = match gen_batch_plan_by_statement(&session, context.into(), stmt)? {
            BatchPlanChoice::Rw(plan_result) => {
                let plan_fragmenter_result = risingwave_expr::expr_context::TIME_ZONE::sync_scope(
                    session.config().timezone(),
                    || gen_batch_plan_fragmenter(&session, plan_result),
                )?;
                execute_risingwave_plan(session, plan_fragmenter_result, formats).left_future()
            }
            BatchPlanChoice::Df(plan_result) => {
                execute_datafusion_plan(session, plan_result, formats).right_future()
            }
        };
        future.await
    }

    #[cfg(not(feature = "datafusion"))]
    {
        let future = match gen_batch_plan_by_statement(&session, context.into(), stmt)? {
            BatchPlanChoice::Rw(plan_result) => {
                let plan_fragmenter_result = risingwave_expr::expr_context::TIME_ZONE::sync_scope(
                    session.config().timezone(),
                    || gen_batch_plan_fragmenter(&session, plan_result),
                )?;
                execute_risingwave_plan(session, plan_fragmenter_result, formats)
            }
        };
        future.await
    }
}

fn handle_parse_inner(binder: Binder, statement: Statement) -> Result<PrepareStatement> {
    let bound_result = gen_bound(binder, statement.clone())?;

    Ok(PrepareStatement::Prepared(PreparedResult {
        statement,
        bound_result,
    }))
}

pub fn handle_parse_for_batch(
    handler_args: HandlerArgs,
    statement: Statement,
    specified_param_types: Vec<Option<DataType>>,
) -> Result<PrepareStatement> {
    let binder = Binder::new_for_batch(&handler_args.session)
        .with_specified_params_types(specified_param_types);
    handle_parse_inner(binder, statement)
}

pub fn handle_parse_for_stream(
    handler_args: HandlerArgs,
    statement: Statement,
    specified_param_types: Vec<Option<DataType>>,
) -> Result<PrepareStatement> {
    let binder = Binder::new_for_stream(&handler_args.session)
        .with_specified_params_types(specified_param_types);
    handle_parse_inner(binder, statement)
}

/// Execute a "Portal", which is a prepared statement with bound parameters.
pub async fn handle_execute(
    handler_args: HandlerArgs,
    portal: PortalResult,
) -> Result<RwPgResponse> {
    let PortalResult {
        bound_result,
        result_formats,
        statement,
    } = portal;
    match statement {
        Statement::Query(_)
        | Statement::Insert { .. }
        | Statement::Delete { .. }
        | Statement::Update { .. } => {
            // Execute a batch query
            let session = handler_args.session.clone();
            let plan_fragmenter_result = {
                let context = OptimizerContext::from_handler_args(handler_args);
                let plan_result =
                    gen_batch_query_plan(&session, context.into(), bound_result)?.unwrap_rw()?;
                // Time zone is used by Hummock time travel query.
                risingwave_expr::expr_context::TIME_ZONE::sync_scope(
                    session.config().timezone(),
                    || gen_batch_plan_fragmenter(&session, plan_result),
                )?
            };
            execute_risingwave_plan(session, plan_fragmenter_result, result_formats).await
        }
        Statement::CreateView { materialized, .. } if materialized => {
            // Execute a CREATE MATERIALIZED VIEW
            let BoundResult {
                bound,
                dependent_relations,
                dependent_udfs,
                dependent_secrets,
                ..
            } = bound_result;
            let create_mv = if let BoundStatement::CreateView(create_mv) = bound {
                create_mv
            } else {
                unreachable!("expect a BoundStatement::CreateView")
            };
            let BoundCreateView {
                or_replace,
                materialized: _,
                if_not_exists,
                name,
                columns,
                query,
                emit_mode,
                with_options,
            } = *create_mv;
            if or_replace {
                bail_not_implemented!("CREATE OR REPLACE VIEW");
            }

            // Hack: replace the `with_options` with the bounded ones.
            let handler_args = HandlerArgs {
                session: handler_args.session.clone(),
                sql: handler_args.sql.clone(),
                normalized_sql: handler_args.normalized_sql.clone(),
                with_options: crate::WithOptions::try_from(with_options.as_slice())?,
            };

            create_mv::handle_create_mv_bound(
                handler_args,
                if_not_exists,
                name,
                *query,
                dependent_relations,
                dependent_udfs,
                dependent_secrets,
                columns,
                emit_mode,
            )
            .await
        }
        Statement::DeclareCursor { stmt } => match stmt.declare_cursor {
            risingwave_sqlparser::ast::DeclareCursor::Query(_) => {
                let session = handler_args.session.clone();
                let plan_fragmenter_result = {
                    let context = OptimizerContext::from_handler_args(handler_args.clone());
                    let plan_result = gen_batch_query_plan(&session, context.into(), bound_result)?
                        .unwrap_rw()?;
                    gen_batch_plan_fragmenter(&session, plan_result)?
                };
                declare_cursor::handle_bound_declare_query_cursor(
                    handler_args,
                    stmt.cursor_name,
                    plan_fragmenter_result,
                )
                .await
            }
            risingwave_sqlparser::ast::DeclareCursor::Subscription(sub_name, rw_timestamp) => {
                declare_cursor::handle_declare_subscription_cursor(
                    handler_args,
                    sub_name,
                    stmt.cursor_name,
                    rw_timestamp,
                )
                .await
            }
        },
        _ => unreachable!(),
    }
}

pub fn gen_batch_plan_by_statement(
    session: &SessionImpl,
    context: OptimizerContextRef,
    stmt: Statement,
) -> Result<BatchPlanChoice> {
    let binder = Binder::new_for_batch(session);
    let bound_result = gen_bound(binder, stmt)?;
    gen_batch_query_plan(session, context, bound_result)
}

#[derive(Clone)]
pub struct BoundResult {
    pub(crate) stmt_type: StatementType,
    pub(crate) must_dist: bool,
    pub(crate) bound: BoundStatement,
    pub(crate) param_types: Vec<DataType>,
    pub(crate) parsed_params: Option<Vec<Datum>>,
    pub(crate) dependent_relations: HashSet<ObjectId>,
    /// TODO(rc): merge with `dependent_relations`
    pub(crate) dependent_udfs: HashSet<FunctionId>,
    pub(crate) dependent_secrets: HashSet<SecretId>,
}

fn gen_bound(mut binder: Binder, stmt: Statement) -> Result<BoundResult> {
    let stmt_type = StatementType::infer_from_statement(&stmt)
        .map_err(|err| RwError::from(ErrorCode::InvalidInputSyntax(err)))?;
    let must_dist = must_run_in_distributed_mode(&stmt)?;

    let bound = binder.bind(stmt)?;

    Ok(BoundResult {
        stmt_type,
        must_dist,
        bound,
        param_types: binder.export_param_types()?,
        parsed_params: None,
        dependent_relations: binder.included_relations().clone(),
        dependent_udfs: binder.included_udfs().clone(),
        dependent_secrets: binder.included_secrets().clone(),
    })
}

pub struct RwBatchQueryPlanResult {
    pub(crate) plan: BatchPlanRef,
    pub(crate) query_mode: QueryMode,
    pub(crate) schema: Schema,
    pub(crate) stmt_type: StatementType,
    // Note that these relations are only resolved in the binding phase, and it may only be a
    // subset of the final one. i.e. the final one may contain more implicit dependencies on
    // indices.
    pub(crate) dependent_relations: Vec<ObjectId>,
    pub(crate) dependent_secrets: Vec<SecretId>,
}

fn gen_batch_query_plan(
    session: &SessionImpl,
    context: OptimizerContextRef,
    bind_result: BoundResult,
) -> Result<BatchPlanChoice> {
    let BoundResult {
        stmt_type,
        must_dist,
        bound,
        dependent_relations,
        dependent_secrets,
        ..
    } = bind_result;

    let mut planner = if matches!(bound, BoundStatement::Query(_)) {
        Planner::new_for_batch_dql(context)
    } else {
        Planner::new_for_batch(context)
    };

    let logical = planner.plan(bound)?;
    let schema = logical.schema();
    let optimized_logical = logical.gen_optimized_logical_plan_for_batch()?;

    #[cfg(feature = "datafusion")]
    {
        if session.config().enable_datafusion_engine() {
            use thiserror_ext::AsReport;

            use crate::datafusion::{GenDataFusionPlanError, try_gen_datafusion_plan};

            match try_gen_datafusion_plan(&optimized_logical) {
                Ok(plan) => {
                    return Ok(BatchPlanChoice::Df(DfBatchQueryPlanResult {
                        plan,
                        schema,
                        stmt_type,
                    }));
                }
                Err(GenDataFusionPlanError::MissingIcebergScan) => {}
                Err(err) => {
                    tracing::warn!(
                        "Failed to generate DataFusion plan, fallback to RisingWave plan: {}",
                        err.as_report()
                    );
                }
            }
        }
    }

    let batch_plan = optimized_logical.gen_batch_plan()?;

    let dependent_relations =
        RelationCollectorVisitor::collect_with(dependent_relations, batch_plan.plan.clone());

    let must_local = must_run_in_local_mode(&batch_plan);

    let query_mode = match (must_dist, must_local) {
        (true, true) => {
            return Err(ErrorCode::InternalError(
                "the query is forced to both local and distributed mode by optimizer".to_owned(),
            )
            .into());
        }
        (true, false) => QueryMode::Distributed,
        (false, true) => QueryMode::Local,
        (false, false) => match session.config().query_mode() {
            QueryMode::Auto => determine_query_mode(&batch_plan),
            QueryMode::Local => QueryMode::Local,
            QueryMode::Distributed => QueryMode::Distributed,
        },
    };

    let physical = match query_mode {
        QueryMode::Auto => unreachable!(),
        QueryMode::Local => batch_plan.gen_batch_local_plan()?,
        QueryMode::Distributed => batch_plan.gen_batch_distributed_plan()?,
    };

    let result = RwBatchQueryPlanResult {
        plan: physical,
        query_mode,
        schema,
        stmt_type,
        dependent_relations: dependent_relations.into_iter().collect_vec(),
        dependent_secrets: dependent_secrets.into_iter().collect_vec(),
    };
    Ok(BatchPlanChoice::Rw(result))
}

fn must_run_in_distributed_mode(stmt: &Statement) -> Result<bool> {
    fn is_insert_using_select(stmt: &Statement) -> bool {
        fn has_select_query(set_expr: &SetExpr) -> bool {
            match set_expr {
                SetExpr::Select(_) => true,
                SetExpr::Query(query) => has_select_query(&query.body),
                SetExpr::SetOperation { left, right, .. } => {
                    has_select_query(left) || has_select_query(right)
                }
                SetExpr::Values(_) => false,
            }
        }

        matches!(
            stmt,
            Statement::Insert {source, ..} if has_select_query(&source.body)
        )
    }

    let stmt_type = StatementType::infer_from_statement(stmt)
        .map_err(|err| RwError::from(ErrorCode::InvalidInputSyntax(err)))?;

    Ok(matches!(
        stmt_type,
        StatementType::UPDATE
            | StatementType::DELETE
            | StatementType::UPDATE_RETURNING
            | StatementType::DELETE_RETURNING
    ) | is_insert_using_select(stmt))
}

fn must_run_in_local_mode(batch_plan: &BatchPlanRoot) -> bool {
    SysTableVisitor::has_sys_table(batch_plan)
}

fn determine_query_mode(batch_plan: &BatchPlanRoot) -> QueryMode {
    if ExecutionModeDecider::run_in_local_mode(batch_plan) {
        QueryMode::Local
    } else {
        QueryMode::Distributed
    }
}

pub struct BatchPlanFragmenterResult {
    pub(crate) plan_fragmenter: BatchPlanFragmenter,
    pub(crate) query_mode: QueryMode,
    pub(crate) schema: Schema,
    pub(crate) stmt_type: StatementType,
}

pub fn gen_batch_plan_fragmenter(
    session: &SessionImpl,
    plan_result: RwBatchQueryPlanResult,
) -> Result<BatchPlanFragmenterResult> {
    let RwBatchQueryPlanResult {
        plan,
        query_mode,
        schema,
        stmt_type,
        ..
    } = plan_result;

    tracing::trace!(
        "Generated query plan: {:?}, query_mode:{:?}",
        plan.explain_to_string(),
        query_mode
    );
    let worker_node_manager_reader = WorkerNodeSelector::new(
        session.env().worker_node_manager_ref(),
        session.is_barrier_read(),
    );
    let plan_fragmenter = BatchPlanFragmenter::new(
        worker_node_manager_reader,
        session.env().catalog_reader().clone(),
        session.config().batch_parallelism().0,
        plan,
    )?;

    Ok(BatchPlanFragmenterResult {
        plan_fragmenter,
        query_mode,
        schema,
        stmt_type,
    })
}

pub async fn create_stream(
    session: Arc<SessionImpl>,
    plan_fragmenter_result: BatchPlanFragmenterResult,
    formats: Vec<Format>,
) -> Result<(PgResponseStream, Vec<PgFieldDescriptor>)> {
    let BatchPlanFragmenterResult {
        plan_fragmenter,
        query_mode,
        schema,
        stmt_type,
        ..
    } = plan_fragmenter_result;

    let mut can_timeout_cancel = true;
    // Acquire the write guard for DML statements.
    match stmt_type {
        StatementType::INSERT
        | StatementType::INSERT_RETURNING
        | StatementType::DELETE
        | StatementType::DELETE_RETURNING
        | StatementType::UPDATE
        | StatementType::UPDATE_RETURNING => {
            session.txn_write_guard()?;
            can_timeout_cancel = false;
        }
        _ => {}
    }

    let query = plan_fragmenter.generate_complete_query().await?;
    tracing::trace!("Generated query after plan fragmenter: {:?}", &query);

    let pg_descs = schema
        .fields()
        .iter()
        .map(to_pg_field)
        .collect::<Vec<PgFieldDescriptor>>();
    let column_types = schema.fields().iter().map(|f| f.data_type()).collect_vec();

    let row_stream = match query_mode {
        QueryMode::Auto => unreachable!(),
        QueryMode::Local => PgResponseStream::LocalQuery(DataChunkToRowSetAdapter::new(
            local_execute(session.clone(), query, can_timeout_cancel).await?,
            column_types,
            formats,
            session.clone(),
        )),
        // Local mode do not support cancel tasks.
        QueryMode::Distributed => {
            PgResponseStream::DistributedQuery(DataChunkToRowSetAdapter::new(
                distribute_execute(session.clone(), query, can_timeout_cancel).await?,
                column_types,
                formats,
                session.clone(),
            ))
        }
    };

    Ok((row_stream, pg_descs))
}

async fn execute_risingwave_plan(
    session: Arc<SessionImpl>,
    plan_fragmenter_result: BatchPlanFragmenterResult,
    formats: Vec<Format>,
) -> Result<RwPgResponse> {
    // Used in counting row count.
    let first_field_format = formats.first().copied().unwrap_or(Format::Text);
    let query_mode = plan_fragmenter_result.query_mode;
    let stmt_type = plan_fragmenter_result.stmt_type;

    let query_start_time = Instant::now();
    let (row_stream, pg_descs) =
        create_stream(session.clone(), plan_fragmenter_result, formats).await?;

    // We need to do some post work after the query is finished and before the `Complete` response
    // it sent. This is achieved by the `callback` in `PgResponse`.
    let callback = async move {
        // Implicitly flush the writes.
        if session.config().implicit_flush() && stmt_type.is_dml() {
            do_flush(&session).await?;
        }

        // update some metrics
        match query_mode {
            QueryMode::Auto => unreachable!(),
            QueryMode::Local => {
                session
                    .env()
                    .frontend_metrics
                    .latency_local_execution
                    .observe(query_start_time.elapsed().as_secs_f64());

                session
                    .env()
                    .frontend_metrics
                    .query_counter_local_execution
                    .inc();
            }
            QueryMode::Distributed => {
                session
                    .env()
                    .query_manager()
                    .query_metrics
                    .query_latency
                    .observe(query_start_time.elapsed().as_secs_f64());

                session
                    .env()
                    .query_manager()
                    .query_metrics
                    .completed_query_counter
                    .inc();
            }
        }

        Ok(())
    };

    Ok(PgResponse::builder(stmt_type)
        .row_cnt_format_opt(Some(first_field_format))
        .values(row_stream, pg_descs)
        .callback(callback)
        .into())
}

pub async fn distribute_execute(
    session: Arc<SessionImpl>,
    query: Query,
    can_timeout_cancel: bool,
) -> Result<DistributedQueryStream> {
    let timeout = if cfg!(madsim) {
        None
    } else if can_timeout_cancel {
        Some(session.statement_timeout())
    } else {
        None
    };
    let execution_context: ExecutionContextRef =
        ExecutionContext::new(session.clone(), timeout).into();
    let query_manager = session.env().query_manager().clone();

    query_manager
        .schedule(execution_context, query)
        .await
        .map_err(|err| err.into())
}

pub async fn local_execute(
    session: Arc<SessionImpl>,
    mut query: Query,
    can_timeout_cancel: bool,
) -> Result<LocalQueryStream> {
    let timeout = if cfg!(madsim) {
        None
    } else if can_timeout_cancel {
        Some(session.statement_timeout())
    } else {
        None
    };
    let front_env = session.env();

    let snapshot = session.pinned_snapshot();

    snapshot.fill_batch_query_epoch(&mut query)?;

    let execution = LocalQueryExecution::new(
        query,
        front_env.clone(),
        snapshot.support_barrier_read(),
        session,
        timeout,
    );

    Ok(execution.stream_rows())
}