Skip to content

Apache DataFusion SQL Query Engine

11.8 KB
Raw Blame
// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: See the NOTICE file distributed with this work for additional information regarding copyright ownership
// SPDX-FileContributor: Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements

//! external_aggr binary entrypoint

use clap::{Args, Parser, Subcommand};
use datafusion::execution::memory_pool::GreedyMemoryPool;
use datafusion::execution::memory_pool::MemoryPool;
use std::collections::HashMap;
use std::path::PathBuf;
use std::sync::Arc;
use std::sync::LazyLock;

use arrow::record_batch::RecordBatch;
use arrow::util::pretty;
use datafusion::datasource::file_format::parquet::ParquetFormat;
use datafusion::datasource::listing::{
    ListingOptions, ListingTable, ListingTableConfig, ListingTableUrl,
};
use datafusion::datasource::{MemTable, TableProvider};
use datafusion::error::Result;
use datafusion::execution::SessionStateBuilder;
use datafusion::execution::memory_pool::{FairSpillPool, PeakRecordingPool};
use datafusion::execution::runtime_env::RuntimeEnvBuilder;
use datafusion::physical_plan::display::DisplayableExecutionPlan;
use datafusion::physical_plan::{collect, displayable};
use datafusion::prelude::*;
use datafusion_benchmarks::util::{BenchmarkRun, CommonOpt, QueryResult};
use datafusion_common::instant::Instant;
use datafusion_common::utils::get_available_parallelism;
use datafusion_common::{DEFAULT_PARQUET_EXTENSION, exec_err};
use datafusion_common::{human_readable_size, units};

#[derive(Debug, Parser)]
#[command(
    name = "datafusion-external-aggregation",
    about = "DataFusion external aggregation benchmark"
)]
struct Cli {
    #[command(subcommand)]
    command: ExternalAggrOpt,
}

#[derive(Debug, Subcommand)]
enum ExternalAggrOpt {
    Benchmark(ExternalAggrConfig),
}

#[derive(Debug, Args)]
struct ExternalAggrConfig {
    /// Query number. If not specified, runs all queries
    #[arg(short, long)]
    query: Option<usize>,

    /// Common options
    #[command(flatten)]
    common: CommonOpt,

    /// Path to data files (lineitem). Only parquet format is supported
    #[arg(required = true, short = 'p', long = "path")]
    path: PathBuf,

    /// Load the data into a MemTable before executing the query
    #[arg(short = 'm', long = "mem-table")]
    mem_table: bool,

    /// Path to JSON benchmark result to be compare using `compare.py`
    #[arg(short = 'o', long = "output")]
    output_path: Option<PathBuf>,
}

/// Query Memory Limits
/// Map query id to predefined memory limits
///
/// Q1 requires 36MiB for aggregation
/// Memory limits to run: 64MiB, 32MiB, 16MiB
/// Q2 requires 250MiB for aggregation
/// Memory limits to run: 512MiB, 256MiB, 128MiB, 64MiB, 32MiB
static QUERY_MEMORY_LIMITS: LazyLock<HashMap<usize, Vec<u64>>> = LazyLock::new(|| {
    use units::*;
    let mut map = HashMap::new();
    map.insert(1, vec![64 * MB, 32 * MB, 16 * MB]);
    map.insert(2, vec![512 * MB, 256 * MB, 128 * MB, 64 * MB, 32 * MB]);
    map
});

impl ExternalAggrConfig {
    const AGGR_TABLES: [&'static str; 1] = ["lineitem"];
    const AGGR_QUERIES: [&'static str; 2] = [
        // Q1: Output size is ~25% of lineitem table
        r#"
        SELECT count(*)
        FROM (
            SELECT DISTINCT l_orderkey
            FROM lineitem
        )
        "#,
        // Q2: Output size is ~99% of lineitem table
        r#"
        SELECT count(*)
        FROM (
            SELECT DISTINCT l_orderkey, l_suppkey
            FROM lineitem
        )
        "#,
    ];

    /// If `--query` and `--memory-limit` is not specified, run all queries
    /// with pre-configured memory limits
    /// If only `--query` is specified, run the query with all memory limits
    /// for this query
    /// If both `--query` and `--memory-limit` are specified, run the query
    /// with the specified memory limit
    pub async fn run(&self) -> Result<()> {
        let mut benchmark_run = BenchmarkRun::new();

        let memory_limit = self.common.memory_limit.map(|limit| limit as u64);
        let mem_pool_type = self.common.mem_pool_type.as_str();

        let query_range = match self.query {
            Some(query_id) => query_id..=query_id,
            None => 1..=Self::AGGR_QUERIES.len(),
        };

        // Each element is (query_id, memory_limit)
        // e.g. [(1, 64_000), (1, 32_000)...] means first run Q1 with 64KiB
        // memory limit, next run Q1 with 32KiB memory limit, etc.
        let mut query_executions = vec![];
        // Setup `query_executions`
        for query_id in query_range {
            if query_id > Self::AGGR_QUERIES.len() {
                return exec_err!(
                    "Invalid '--query'(query number) {} for external aggregation benchmark.",
                    query_id
                );
            }

            match memory_limit {
                Some(limit) => {
                    query_executions.push((query_id, limit));
                }
                None => {
                    let memory_limits = QUERY_MEMORY_LIMITS.get(&query_id).unwrap();
                    for limit in memory_limits {
                        query_executions.push((query_id, *limit));
                    }
                }
            }
        }

        for (query_id, mem_limit) in query_executions {
            benchmark_run.start_new_case(&format!(
                "{query_id}({})",
                human_readable_size(mem_limit as usize)
            ));

            let query_results = self
                .benchmark_query(query_id, mem_limit, mem_pool_type, &mut benchmark_run)
                .await?;
            for iter in query_results {
                benchmark_run.write_iter(iter.elapsed, iter.row_count);
            }
        }

        benchmark_run.maybe_write_json(self.output_path.as_ref())?;

        Ok(())
    }

    /// Benchmark query `query_id` in `AGGR_QUERIES`
    ///
    /// `benchmark_run` is handed this query's runtime, which is built here
    /// because each query runs under its own memory limit.
    async fn benchmark_query(
        &self,
        query_id: usize,
        mem_limit: u64,
        mem_pool_type: &str,
        benchmark_run: &mut BenchmarkRun,
    ) -> Result<Vec<QueryResult>> {
        let query_name =
            format!("Q{query_id}({})", human_readable_size(mem_limit as usize));
        let config = self.common.config()?;
        let memory_pool: Arc<dyn MemoryPool> = match mem_pool_type {
            "fair" => Arc::new(FairSpillPool::new(mem_limit as usize)),
            "greedy" => Arc::new(GreedyMemoryPool::new(mem_limit as usize)),
            _ => {
                return exec_err!("Invalid memory pool type: {}", mem_pool_type);
            }
        };
        // This benchmark builds its pool directly rather than going through
        // `CommonOpt::runtime_env_builder`, so it has to install the recorder
        // itself to report a peak.
        let memory_pool: Arc<dyn MemoryPool> =
            Arc::new(PeakRecordingPool::new(memory_pool));
        benchmark_run.set_memory_pool(&memory_pool);
        let runtime_env = RuntimeEnvBuilder::new()
            .with_memory_pool(memory_pool)
            .build_arc()?;
        let state = SessionStateBuilder::new()
            .with_config(config)
            .with_runtime_env(runtime_env)
            .with_default_features()
            .build();
        let ctx = SessionContext::from(state);

        // register tables
        self.register_tables(&ctx).await?;

        let mut millis = vec![];
        // run benchmark
        let mut query_results = vec![];
        for i in 0..self.iterations() {
            let start = Instant::now();

            let query_idx = query_id - 1; // 1-indexed -> 0-indexed
            let sql = Self::AGGR_QUERIES[query_idx];

            let result = self.execute_query(&ctx, sql).await?;

            let elapsed = start.elapsed(); //.as_secs_f64() * 1000.0;
            let ms = elapsed.as_secs_f64() * 1000.0;
            millis.push(ms);

            let row_count = result.iter().map(|b| b.num_rows()).sum();
            println!(
                "{query_name} iteration {i} took {ms:.1} ms and returned {row_count} rows"
            );
            query_results.push(QueryResult { elapsed, row_count });
        }

        let avg = millis.iter().sum::<f64>() / millis.len() as f64;
        println!("{query_name} avg time: {avg:.2} ms");

        Ok(query_results)
    }

    async fn register_tables(&self, ctx: &SessionContext) -> Result<()> {
        for table in Self::AGGR_TABLES {
            let table_provider = { self.get_table(ctx, table).await? };

            if self.mem_table {
                println!("Loading table '{table}' into memory");
                let start = Instant::now();
                let memtable =
                    MemTable::load(table_provider, Some(self.partitions()), &ctx.state())
                        .await?;
                println!(
                    "Loaded table '{}' into memory in {} ms",
                    table,
                    start.elapsed().as_millis()
                );
                ctx.register_table(table, Arc::new(memtable))?;
            } else {
                ctx.register_table(table, table_provider)?;
            }
        }
        Ok(())
    }

    async fn execute_query(
        &self,
        ctx: &SessionContext,
        sql: &str,
    ) -> Result<Vec<RecordBatch>> {
        let debug = self.common.debug;
        let plan = ctx.sql(sql).await?;
        let (state, plan) = plan.into_parts();

        if debug {
            println!("=== Logical plan ===\n{plan}\n");
        }

        let plan = state.optimize(&plan)?;
        if debug {
            println!("=== Optimized logical plan ===\n{plan}\n");
        }
        let physical_plan = state.create_physical_plan(&plan).await?;
        if debug {
            println!(
                "=== Physical plan ===\n{}\n",
                displayable(physical_plan.as_ref()).indent(true)
            );
        }
        let result = collect(physical_plan.clone(), state.task_ctx()).await?;
        if debug {
            println!(
                "=== Physical plan with metrics ===\n{}\n",
                DisplayableExecutionPlan::with_metrics(physical_plan.as_ref())
                    .indent(true)
            );
            if !result.is_empty() {
                // do not call print_batches if there are no batches as the result is confusing
                // and makes it look like there is a batch with no columns
                pretty::print_batches(&result)?;
            }
        }
        Ok(result)
    }

    async fn get_table(
        &self,
        ctx: &SessionContext,
        table: &str,
    ) -> Result<Arc<dyn TableProvider>> {
        let path = self.path.to_str().unwrap();

        // Obtain a snapshot of the SessionState
        let state = ctx.state();
        let path = format!("{path}/{table}");
        let format = Arc::new(
            ParquetFormat::default()
                .with_options(ctx.state().table_options().parquet.clone()),
        );
        let extension = DEFAULT_PARQUET_EXTENSION;

        let options = ListingOptions::new(format).with_file_extension(extension);

        let table_path = ListingTableUrl::parse(path)?;
        let config = ListingTableConfig::new(table_path).with_listing_options(options);
        let config = config.infer_schema(&state).await?;

        Ok(Arc::new(ListingTable::try_new(config)?.with_cache(
            ctx.runtime_env().cache_manager.get_file_statistic_cache(),
        )))
    }

    fn iterations(&self) -> usize {
        self.common.iterations
    }

    fn partitions(&self) -> usize {
        self.common
            .partitions
            .unwrap_or_else(get_available_parallelism)
    }
}

#[tokio::main]
pub async fn main() -> Result<()> {
    env_logger::init();

    let cli = Cli::parse();
    match cli.command {
        ExternalAggrOpt::Benchmark(opt) => opt.run().await?,
    }

    Ok(())
}