duckdb.rs

/*
Copyright 2024-2025 The Spice.ai OSS Authors

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

     https://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 super::{AccelerationSource, BootstrapStatus, DataAccelerator};
use crate::{
    App, Runtime,
    component::{
        dataset::{
            Dataset,
            acceleration::{Acceleration, Engine, Mode, RefreshMode},
        },
        view::View,
    },
    dataaccelerator::{
        FilePathError,
        snapshots::{download_snapshot_if_needed, snapshot_before_recreate},
        storage::{ResolvedAccelerationStorage, resolve_acceleration_storage_async},
    },
    datafusion::{
        dialect::new_duckdb_dialect,
        sort_columns::{SortColumn, parse_sort_columns},
        udf::deny_spice_functions_for_duckdb,
    },
    make_spice_data_directory,
    parameters::ParameterSpec,
    register_data_accelerator, spice_data_base_path,
};
use async_trait::async_trait;
use data_components::poly::PolyTableProvider;
use datafusion::error::DataFusionError;
use datafusion::execution::runtime_env::RuntimeEnv;
use datafusion::{
    catalog::TableProviderFactory,
    datasource::TableProvider,
    execution::context::SessionContext,
    logical_expr::CreateExternalTable,
    sql::sqlparser::ast::{
        Delete, FromTable, Ident, ObjectName, ObjectNamePart, Statement as SQLStatement,
        TableFactor,
    },
};
use datafusion_table_providers::{
    duckdb::{
        DuckDB, DuckDBSettingsRegistry, DuckDBTableProviderFactory,
        write::{DuckDBTableWriter, WriteCompletionHandler},
    },
    sql::db_connection_pool::{
        self as db_connection_pool,
        duckdbpool::{DuckDbConnectionPool, DuckDbConnectionPoolBuilder},
    },
};
use duckdb::AccessMode;
use itertools::Itertools;
use runtime_acceleration::snapshot::AccelerationEngine;
use runtime_table_partition::expression::PartitionedBy;
use settings::OrderByNonIntegerLiteral;
use snafu::prelude::*;
use std::collections::HashMap;
use std::{
    any::Any,
    cmp::max,
    collections::HashSet,
    ffi::OsStr,
    path::PathBuf,
    sync::{Arc, Once},
};

pub(crate) mod settings;

/// Creates a [`DuckDBTableProviderFactory`] with standard Spice settings (dialect, timezone,
/// index scan tuning, function deny-list). All `DuckDB` accelerator consumers should use this
/// to avoid divergent configurations.
pub(crate) fn create_factory() -> DuckDBTableProviderFactory {
    DuckDBTableProviderFactory::new(AccessMode::ReadWrite)
        .with_dialect(new_duckdb_dialect())
        .with_settings_registry(
            DuckDBSettingsRegistry::new()
                .with_setting(Box::new(OrderByNonIntegerLiteral))
                .with_setting(Box::new(settings::IndexScanPercentage))
                .with_setting(Box::new(settings::IndexScanMaxCount))
                .with_setting(Box::new(settings::TimeZone)),
        )
        .with_function_support(deny_spice_functions_for_duckdb().as_ref().clone())
}

pub(crate) const DEFAULT_CONNECTION_POOL_SIZE: u32 = 10;
pub(crate) const DEFAULT_EBS_CONNECTION_POOL_SIZE: u32 = 4;
pub(crate) const SPICE_ACCELERATOR_METADATA_KEY: &str = "spice.accelerator";
pub(crate) const SPICE_OPT_DUCKDB_AGG_PUSHDOWN_KEY: &str =
    "spice.optimizer.duckdb_aggregate_pushdown";

use super::upsert_dedup;

#[derive(Debug, Snafu)]
pub enum Error {
    #[snafu(display("Unable to create table: {source}"))]
    UnableToCreateTable {
        source: datafusion::error::DataFusionError,
    },

    #[snafu(display("Acceleration creation failed: {source}"))]
    AccelerationCreationFailed {
        source: Box<dyn std::error::Error + Send + Sync>,
    },

    #[snafu(display("Acceleration initialization failed: {source}"))]
    AccelerationInitializationFailed {
        source: Box<dyn std::error::Error + Send + Sync>,
    },

    #[snafu(display(r#"The "duckdb_file" acceleration parameter has an invalid extension. Expected one of "{valid_extensions}" but got "{extension}"."#))]
    InvalidFileExtension {
        valid_extensions: String,
        extension: String,
    },

    #[snafu(display(r#"The "duckdb_file" acceleration parameter is a directory."#))]
    InvalidFileIsDirectory,

    #[snafu(display("Acceleration not enabled for dataset: {dataset}"))]
    AccelerationNotEnabled { dataset: Arc<str> },

    #[snafu(display("Invalid DuckDB acceleration configuration: {detail}"))]
    InvalidConfiguration { detail: Arc<str> },
}

type Result<T, E = Error> = std::result::Result<T, E>;

pub struct DuckDBAccelerator {
    duckdb_factory: DuckDBTableProviderFactory,
}

impl DuckDBAccelerator {
    #[must_use]
    pub fn new() -> Self {
        Self {
            duckdb_factory: create_factory(),
        }
    }

    /// Returns the `DuckDB` file path that would be used for a file-based `DuckDB` accelerator from this dataset
    pub fn duckdb_file_path(&self, source: &dyn AccelerationSource) -> Result<String> {
        duckdb_file_path(&self.duckdb_factory, source, "accelerated_duckdb")
    }

    /// Returns an existing `DuckDB` connection pool for the given dataset, or creates a new one if it doesn't exist.
    pub async fn get_shared_pool(
        &self,
        source: &dyn AccelerationSource,
    ) -> Result<DuckDbConnectionPool> {
        let duckdb_file = self.duckdb_file_path(source);

        let acceleration = source.acceleration().context(AccelerationNotEnabledSnafu {
            dataset: source.name().to_string(),
        })?;

        let pool = match (duckdb_file, acceleration.mode) {
            (Ok(duckdb_file), Mode::File | Mode::FileCreate | Mode::FileUpdate) => {
                let num_accelerating_datasets = self.get_num_accelerating_datasets(
                    Some(duckdb_file.as_str()),
                    &source.app(),
                    source.runtime(),
                );
                let storage =
                    resolve_acceleration_storage_async(acceleration.storage_profile, &duckdb_file)
                        .await;
                tracing::debug!(
                    dataset = %source.name(),
                    storage = %storage,
                    "Resolved DuckDB acceleration storage profile"
                );
                let max_size =
                    Self::get_pool_max_size(num_accelerating_datasets, acceleration, storage);
                let min_idle = Self::get_pool_min_idle(storage, max_size);
                let mut pool_builder = DuckDbConnectionPoolBuilder::file(&duckdb_file)
                    .with_max_size(Some(max_size))
                    .with_min_idle(Some(min_idle))
                    .with_connection_setup_query("PRAGMA enable_checkpoint_on_shutdown");
                for pragma in Self::storage_setup_queries(storage) {
                    pool_builder = pool_builder.with_connection_setup_query(*pragma);
                }
                self.duckdb_factory
                    .get_or_init_instance_with_builder(pool_builder)
                    .await
                    .boxed()
                    .context(AccelerationCreationFailedSnafu)?
            }
            (_, Mode::Memory) => {
                let num_accelerating_datasets =
                    self.get_num_accelerating_datasets(None, &source.app(), source.runtime());
                let max_size = Self::get_pool_max_size(
                    num_accelerating_datasets,
                    acceleration,
                    ResolvedAccelerationStorage::Unknown,
                );
                let min_idle =
                    Self::get_pool_min_idle(ResolvedAccelerationStorage::Unknown, max_size);
                let pool_builder = DuckDbConnectionPoolBuilder::memory()
                    .with_max_size(Some(max_size))
                    .with_min_idle(Some(min_idle))
                    .with_connection_setup_query("PRAGMA enable_checkpoint_on_shutdown");
                self.duckdb_factory
                    .get_or_init_instance_with_builder(pool_builder)
                    .await
                    .boxed()
                    .context(AccelerationCreationFailedSnafu)?
            }
            (Err(e), Mode::File | Mode::FileCreate | Mode::FileUpdate) => {
                return Err(Error::InvalidConfiguration {
                    detail: Arc::from(e.to_string()),
                });
            }
        };

        Ok(pool)
    }

    fn get_num_accelerating_datasets(
        &self,
        path: Option<&str>,
        app: &Arc<App>,
        rt: Arc<Runtime>,
    ) -> u32 {
        let mut instance_usage: u32 = 1;

        let datasets = rt.get_valid_datasets(app, crate::LogErrors(false));
        for ds in datasets {
            if let Some(acceleration) = &ds.acceleration {
                if acceleration.engine != Engine::DuckDB {
                    continue;
                }

                // If the path is Some, we're counting the number of file instances
                if let Some(this_file_path) = path {
                    if matches!(
                        acceleration.mode,
                        Mode::File | Mode::FileCreate | Mode::FileUpdate
                    ) && let Ok(file_path) = self.file_path(ds.as_ref())
                        && this_file_path == file_path
                    {
                        instance_usage += 1;
                    }
                } else {
                    // If the path is None, we're just counting the number of memory instances
                    if acceleration.mode == Mode::Memory {
                        instance_usage += 1;
                    }
                }
            }
        }

        instance_usage
    }

    pub(crate) fn default_connection_pool_size(storage: ResolvedAccelerationStorage) -> u32 {
        match storage {
            ResolvedAccelerationStorage::Ebs => DEFAULT_EBS_CONNECTION_POOL_SIZE,
            ResolvedAccelerationStorage::LocalSsd
            | ResolvedAccelerationStorage::Tmpfs
            | ResolvedAccelerationStorage::Unknown => DEFAULT_CONNECTION_POOL_SIZE,
        }
    }

    pub(crate) fn get_pool_min_idle(storage: ResolvedAccelerationStorage, max_size: u32) -> u32 {
        Self::default_connection_pool_size(storage).min(max_size)
    }

    /// Storage-profile-specific `DuckDB` pragmas applied to every connection in
    /// the pool. These tune `DuckDB`'s I/O behavior to match the underlying
    /// medium's latency and durability profile.
    pub(crate) fn storage_setup_queries(
        storage: ResolvedAccelerationStorage,
    ) -> &'static [&'static str] {
        match storage {
            // Network-attached block storage (e.g. EBS, Azure Managed Disks)
            // pays per-IO latency on every flush. Raise the checkpoint
            // threshold so WAL flushes are larger and less frequent, which
            // reduces write amplification on the slow link.
            ResolvedAccelerationStorage::Ebs => &["PRAGMA checkpoint_threshold='256MiB'"],
            // tmpfs/ramfs is volatile and effectively free to write, but
            // checkpointing still copies pages around. Push the threshold up
            // so steady-state workloads don't pay checkpoint cost on tiny
            // amounts of dirty data.
            ResolvedAccelerationStorage::Tmpfs => &["PRAGMA checkpoint_threshold='1GiB'"],
            // Local SSD/NVMe handles small frequent flushes well; keep
            // DuckDB defaults.
            ResolvedAccelerationStorage::LocalSsd | ResolvedAccelerationStorage::Unknown => &[],
        }
    }

    fn get_pool_max_size(
        num_accelerating_datasets: u32,
        acceleration: &Acceleration,
        storage: ResolvedAccelerationStorage,
    ) -> u32 {
        let pool_size_param = acceleration
            .params
            .get("connection_pool_size")
            .and_then(|size_str| size_str.parse::<u32>().ok());

        pool_size_param.unwrap_or_else(|| {
            max(
                Self::default_connection_pool_size(storage),
                num_accelerating_datasets,
            )
        })
    }
}

/// Returns the `DuckDB` file path that would be used for a file-based `DuckDB` acceleration for this acceleration source
///
/// # Parameters
///
/// * `duckdb_factory` - The `DuckDB` table provider factory used to generate the file path
/// * `source` - The acceleration source (dataset or view) containing acceleration configuration
/// * `default_db_name` - Default database file name to use if the `duckdb_file` parameter is not specified
pub fn duckdb_file_path(
    duckdb_factory: &DuckDBTableProviderFactory,
    source: &dyn AccelerationSource,
    default_db_name: &str,
) -> Result<String> {
    if !source.is_file_accelerated() {
        Err(Error::InvalidConfiguration {
            detail: Arc::from("Dataset is not file accelerated"),
        })
    } else if let Some(acceleration) = source.acceleration().as_ref() {
        let mut params = acceleration.params.clone();
        let mut using_duckdb_data_dir = true;
        let data_directory = params.remove("duckdb_data_dir").unwrap_or_else(|| {
            using_duckdb_data_dir = false;
            spice_data_base_path()
        });
        params.insert("data_directory".to_string(), data_directory);

        if let Some(duckdb_file) = params.remove("duckdb_file") {
            if using_duckdb_data_dir {
                static WARN_ONCE: Once = Once::new();
                WARN_ONCE.call_once(|| {
                    tracing::warn!(
                        "'duckdb_data_dir' and 'duckdb_file' were both specified but 'duckdb_file' ({duckdb_file}) will be used."
                    );
                });
            }
            params.insert("duckdb_open".to_string(), duckdb_file);
        }

        duckdb_factory
            .duckdb_file_path(default_db_name, &mut params)
            .map_err(|err| Error::InvalidConfiguration {
                detail: Arc::from(err.to_string()),
            })
    } else {
        unreachable!("Expected dataset to have acceleration parameters, but none were found")
    }
}

impl Default for DuckDBAccelerator {
    fn default() -> Self {
        Self::new()
    }
}

const PARAMETERS: &[ParameterSpec] = &[
    ParameterSpec::runtime("file_watcher"),
    ParameterSpec::component("file"),
    ParameterSpec::component("data_dir"),
    ParameterSpec::component("memory_limit"),
    ParameterSpec::component("preserve_insertion_order"),
    ParameterSpec::component("index_scan_percentage"),
    ParameterSpec::component("index_scan_max_count"),
    ParameterSpec::runtime("partition_mode"),
    ParameterSpec::component("partitioned_write_flush_threshold_rows"),
    ParameterSpec::runtime("connection_pool_size").description(
        "The maximum number of client connections created in the duckdb connection pool.",
    ),
    ParameterSpec::runtime("on_refresh_recompute_statistics"),
    ParameterSpec::runtime("on_refresh_sort_columns"),
    ParameterSpec::runtime("partitioned_write_buffer"),
    ParameterSpec::runtime("optimizer_duckdb_aggregate_pushdown"),
];

#[async_trait]
impl DataAccelerator for DuckDBAccelerator {
    fn as_any(&self) -> &dyn Any {
        self
    }

    fn name(&self) -> &'static str {
        "duckdb"
    }

    fn valid_file_extensions(&self) -> Vec<&'static str> {
        vec!["db", "ddb", "duckdb"]
    }

    fn file_path(&self, source: &dyn AccelerationSource) -> Result<String, FilePathError> {
        self.duckdb_file_path(source)
            .map_err(|e| FilePathError::External {
                engine: Engine::DuckDB,
                source: e.into(),
            })
    }

    fn is_initialized(&self, source: &dyn AccelerationSource) -> bool {
        if !source.is_file_accelerated() {
            return true; // memory mode DuckDB is always initialized
        }

        // otherwise, we're initialized if the file exists
        self.has_existing_file(source)
    }

    async fn init(
        &self,
        source: &dyn AccelerationSource,
    ) -> Result<BootstrapStatus, Box<dyn std::error::Error + Send + Sync>> {
        if !source.is_file_accelerated() {
            return Ok(BootstrapStatus::none());
        }

        let path = self.file_path(source)?;

        if let Some(acceleration) = source.acceleration() {
            if !acceleration.params.contains_key("duckdb_file") {
                make_spice_data_directory().map_err(|err| {
                    Error::AccelerationInitializationFailed { source: err.into() }
                })?;
            } else if !self.is_valid_file(source) {
                if std::path::Path::new(&path).is_dir() {
                    return Err(Error::InvalidFileIsDirectory.into());
                }

                let extension = std::path::Path::new(&path)
                    .extension()
                    .and_then(OsStr::to_str)
                    .unwrap_or("");

                return Err(Error::InvalidFileExtension {
                    valid_extensions: self.valid_file_extensions().join(","),
                    extension: extension.to_string(),
                }
                .into());
            }

            // If mode is FileCreate, snapshot the existing file (if enabled) then delete it to start fresh
            if acceleration.mode == Mode::FileCreate {
                let file_path = std::path::Path::new(&path);
                if file_path.exists() {
                    snapshot_before_recreate(
                        acceleration,
                        &source.name().to_string(),
                        runtime_acceleration::snapshot::AccelerationLayout::file(PathBuf::from(
                            &path,
                        )),
                        AccelerationEngine::DuckDB,
                        Arc::new(arrow_schema::Schema::empty()),
                        None,
                    )
                    .await;

                    tracing::warn!(
                        "DuckDB acceleration mode is 'file_create', removing existing file: {}",
                        path
                    );
                    std::fs::remove_file(file_path).map_err(|err| {
                        Error::AccelerationInitializationFailed { source: err.into() }
                    })?;
                }
            }

            let bootstrap_status = download_snapshot_if_needed(
                acceleration,
                source,
                runtime_acceleration::snapshot::AccelerationLayout::file(PathBuf::from(path)),
                AccelerationEngine::DuckDB,
                None,
            )
            .await;

            self.get_shared_pool(source).await?;

            return Ok(bootstrap_status);
        }

        Ok(BootstrapStatus::none())
    }

    /// Creates a new table in the accelerator engine, returning a `TableProvider` that supports reading and writing.
    async fn create_external_table(
        &self,
        mut cmd: CreateExternalTable,
        source: Option<&dyn AccelerationSource>,
        _partition_by: Vec<PartitionedBy>,
        _runtime_env: Option<Arc<RuntimeEnv>>,
    ) -> Result<Arc<dyn TableProvider>, Box<dyn std::error::Error + Send + Sync>> {
        if let Some(duckdb_file) = cmd.options.remove("file") {
            cmd.options.insert("open".to_string(), duckdb_file);
        }

        if let Some(recompute_statistics_on_write) =
            cmd.options.remove("on_refresh_recompute_statistics")
        {
            // Translate Spice parameter to DuckDB write setting
            cmd.options.insert(
                "recompute_statistics_on_write".to_string(),
                recompute_statistics_on_write,
            );
        }

        let is_changes_refresh = source
            .and_then(|src| src.acceleration())
            .and_then(|acceleration| acceleration.refresh_mode)
            .is_some_and(|refresh_mode| refresh_mode == RefreshMode::Changes);
        apply_changes_refresh_write_defaults(&mut cmd, is_changes_refresh);

        // Modify the `cmd` by adding options to attach other databases
        if let Some(source) = source {
            if let Some(temp_directory) = source
                .app()
                .runtime
                .query
                .clone()
                .unwrap_or_default()
                .temp_directory
            {
                cmd.options
                    .insert("temp_directory".to_string(), temp_directory);
            }

            if source.is_file_accelerated() {
                // If the user didn't specify a DuckDB file and this is a file-mode DuckDB,
                // then use the shared DuckDB file `accelerated_duckdb.db`
                if !cmd.options.contains_key("open") {
                    let duckdb_file = self.duckdb_file_path(source)?;
                    cmd.options.insert("open".to_string(), duckdb_file);
                }

                let datasets: Vec<Arc<Dataset>> = Arc::clone(&source.runtime())
                    .get_initialized_datasets(&source.app(), crate::LogErrors(false))
                    .await;

                let views: Vec<Arc<View>> = Arc::clone(&source.runtime())
                    .get_initialized_views(&source.app(), crate::LogErrors(false))
                    .await;

                let self_path = self.file_path(source)?;
                let attach_databases = datasets
                    .into_iter()
                    .map(|ds| ds as Arc<dyn AccelerationSource>)
                    .chain(
                        views
                            .into_iter()
                            .map(|view| view as Arc<dyn AccelerationSource>),
                    )
                    .filter_map(|other_source| {
                        if other_source.acceleration().is_some_and(|a| {
                            a.engine == Engine::DuckDB
                                && matches!(
                                    a.mode,
                                    Mode::File | Mode::FileCreate | Mode::FileUpdate
                                )
                        }) {
                            if other_source.name() == source.name() {
                                None
                            } else {
                                let other_path = self.file_path(other_source.as_ref());
                                other_path.ok().filter(|p| p != &self_path)
                            }
                        } else {
                            None
                        }
                    })
                    .collect::<HashSet<_>>(); // collect unique paths using HashSet

                if !attach_databases.is_empty() {
                    cmd.options.insert(
                        "attach_databases".to_string(),
                        attach_databases.iter().join(";"),
                    );
                }
            }
        }

        let write_completion_handler = source.and_then(|src| {
            let acceleration = src.acceleration()?;
            let dataset_name = src.name().to_string();
            let schema = Arc::new(cmd.schema.as_arrow().clone());

            let mut config = OnRefreshConfig::empty();

            // Parse retention SQL if configured
            if let Some(retention_sql) = acceleration
                .retention_sql
                .as_deref()
                .map(str::trim)
                .filter(|sql| !sql.is_empty())
            {
                match crate::datafusion::retention_sql::parse_retention_sql(
                    src.name(),
                    retention_sql,
                    Arc::clone(&schema),
                ) {
                    Ok(parsed_sql) => {
                        config = config.with_retention(parsed_sql.delete_statement);
                    }
                    Err(e) => {
                        tracing::warn!(
                            "Failed to parse retention_sql for dataset {dataset_name}: {e}. Retention SQL will not be applied.");
                    }
                }
            }

            // Parse on_refresh_sort_columns if configured
            if let Some(sort_columns_str) = acceleration
                .params
                .get("on_refresh_sort_columns")
                .map(String::as_str)
                .map(str::trim)
                .filter(|s| !s.is_empty())
            {
                match parse_sort_columns(sort_columns_str, &schema) {
                    Ok(sort_columns) => {
                        tracing::debug!(
                            dataset = %dataset_name,
                            sort_columns = ?sort_columns,
                            "Parsed on_refresh_sort_columns"
                        );
                        config = config.with_sort(sort_columns);
                    }
                    Err(e) => {
                        tracing::warn!(
                            "Failed to parse on_refresh_sort_columns for dataset {dataset_name}: {e}. Sorting will not be applied."
                        );
                    }
                }
            }

            Some(make_on_refresh_write_handler(dataset_name, config))
        });

        Ok(create_table_provider(&self.duckdb_factory, &cmd, write_completion_handler).await?)
    }

    fn prefix(&self) -> &'static str {
        "duckdb"
    }

    fn parameters(&self) -> &'static [ParameterSpec] {
        PARAMETERS
    }

    fn supports_snapshot_reload(&self) -> bool {
        true
    }

    /// Reloads the `DuckDB`-backed table provider from the snapshot file
    /// that was just written to the primary path.
    ///
    /// Drops the previous provider, evicts the cached connection pool from
    /// the upstream `DuckDBTableProviderFactory` registry, and then re-runs
    /// the registry factory to build a fresh provider over the on-disk file.
    /// The pool eviction is required because the registry caches pool
    /// instances by file path; without it, the freshly built provider would
    /// reuse the prior pool's open connections — which keep observing the
    /// previous file inode — and queries would continue to return stale data
    /// even after the file has been atomically replaced on disk.
    async fn reload_from_snapshot(
        &self,
        source: &dyn AccelerationSource,
        previous_provider: Arc<dyn TableProvider>,
        provider_factory: super::ReloadProviderFactory,
    ) -> Result<Arc<dyn TableProvider>, Box<dyn std::error::Error + Send + Sync>> {
        // Drop the caller's clone first so the only remaining strong refs to
        // the prior pool are the registry entry (which we are about to evict)
        // and any in-flight queries (which will drain naturally).
        drop(previous_provider);

        // Evict the cached pool. For file mode this matches the path the
        // factory keyed on at construction time; for memory mode this falls
        // back to the in-memory key. Snapshot reload is only meaningful for
        // file mode in practice (memory accelerators cannot be snapshotted),
        // but the memory branch is kept for completeness.
        let acceleration =
            source
                .acceleration()
                .ok_or_else(|| -> Box<dyn std::error::Error + Send + Sync> {
                    "acceleration not configured for snapshot reload".into()
                })?;
        match acceleration.mode {
            Mode::File | Mode::FileCreate | Mode::FileUpdate => {
                let path = self.duckdb_file_path(source).boxed()?;
                self.duckdb_factory.invalidate_file_instance(path).await;
            }
            Mode::Memory => {
                self.duckdb_factory
                    .invalidate_instance(&db_connection_pool::DbInstanceKey::memory())
                    .await;
            }
        }

        provider_factory().await
    }

    async fn drop_table(
        &self,
        table_name: &str,
        source: &dyn AccelerationSource,
    ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
        let pool = Arc::new(self.get_shared_pool(source).await?);
        let table_name = table_name.to_owned();

        tokio::task::spawn_blocking(
            move || -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
                let mut conn = pool.connect_sync()?;
                let duckdb_conn = DuckDB::duckdb_conn(&mut conn).boxed()?;
                let escaped = table_name.replace('"', "\"\"");
                let drop_sql = format!("DROP TABLE IF EXISTS \"{escaped}\"");
                duckdb_conn
                    .get_underlying_conn_mut()
                    .execute(&drop_sql, [])
                    .boxed()?;
                // Also drop any internal DuckDB tables associated with this table
                let internal_name = format!("__data_{table_name}").replace('"', "\"\"");
                let internal_drop = format!("DROP TABLE IF EXISTS \"{internal_name}\"");
                let _ = duckdb_conn
                    .get_underlying_conn_mut()
                    .execute(&internal_drop, []);
                tracing::info!(
                    "Dropped DuckDB table '{table_name}' for schema recreation (file_update mode)"
                );
                Ok(())
            },
        )
        .await
        .map_err(|e| -> Box<dyn std::error::Error + Send + Sync> { Box::new(e) })?
    }
}

fn apply_changes_refresh_write_defaults(cmd: &mut CreateExternalTable, is_changes_refresh: bool) {
    if is_changes_refresh && !cmd.options.contains_key("recompute_statistics_on_write") {
        cmd.options.insert(
            "recompute_statistics_on_write".to_string(),
            "false".to_string(),
        );
    }
}

pub(crate) async fn create_table_provider(
    duckdb_factory: &DuckDBTableProviderFactory,
    cmd: &CreateExternalTable,
    on_data_written: Option<WriteCompletionHandler>,
) -> Result<Arc<dyn TableProvider>, Box<dyn std::error::Error + Send + Sync>> {
    let ctx = SessionContext::new();

    let table_provider = duckdb_factory
        .create(&ctx.state(), cmd)
        .await
        .context(UnableToCreateTableSnafu)
        .boxed()?;

    let Some(duckdb_writer) = table_provider.as_any().downcast_ref::<DuckDBTableWriter>() else {
        unreachable!("DuckDBTableWriter should be returned from DuckDBTableProviderFactory")
    };

    let read_provider = Arc::clone(&duckdb_writer.read_provider);
    let duckdb_writer: Arc<DuckDBTableWriter> = match on_data_written {
        Some(handler) => Arc::new(duckdb_writer.clone().with_on_data_written_handler(handler)),
        None => Arc::new(duckdb_writer.clone()),
    };

    // Wrap with upsert deduplication if needed
    let write_provider = upsert_dedup::wrap_with_upsert_dedup_if_needed(
        duckdb_writer,
        &cmd.options,
        cmd.constraints.clone(),
    );

    let mut schema_metadata = HashMap::new();
    schema_metadata.insert(
        SPICE_ACCELERATOR_METADATA_KEY.to_string(),
        "duckdb".to_string(),
    );

    let agg_pushdown_optimization = cmd
        .options
        .get("optimizer_duckdb_aggregate_pushdown")
        .map_or("disabled", |v| v.as_str())
        .to_lowercase();

    schema_metadata.insert(
        SPICE_OPT_DUCKDB_AGG_PUSHDOWN_KEY.to_string(),
        agg_pushdown_optimization,
    );

    let table_provider = Arc::new(PolyTableProvider::new_with_schema_metadata(
        write_provider,
        read_provider,
        schema_metadata,
    ));

    Ok(table_provider)
}

/// Reconstruct the DELETE statement with the internal `DuckDB` table name.
fn reconstruct_retention_sql_with_table_name(
    delete: &Delete,
    internal_table_name: &str,
) -> Result<String, String> {
    // Clone the delete statement and modify the table name
    let mut modified_delete = delete.clone();

    // Replace the table name with the internal table name
    // DuckDB internal table names should be used as-is without schema qualification
    let FromTable::WithFromKeyword(from_tables) = &mut modified_delete.from else {
        return Err("DELETE statement must use FROM keyword".to_string());
    };

    // Replace the first table's name, keeping all other properties
    let Some(table_relation) = from_tables.first_mut() else {
        return Err("No table specified in DELETE statement".to_string());
    };

    if let TableFactor::Table { name, .. } = &mut table_relation.relation {
        *name = ObjectName(vec![ObjectNamePart::Identifier(Ident::new(
            internal_table_name,
        ))]);
    } else {
        return Err("DELETE statement must reference a simple table".to_string());
    }

    // Simply convert the AST to string using Display trait
    let statement = SQLStatement::Delete(modified_delete);
    Ok(statement.to_string())
}

/// Configuration for on-refresh operations that run after data is written but before commit.
#[derive(Clone, Default)]
struct OnRefreshConfig {
    /// Parsed DELETE statement for retention SQL
    retention_delete: Option<Delete>,
    /// Sort columns configuration
    sort_columns: Vec<SortColumn>,
}

impl OnRefreshConfig {
    /// Creates an empty config.
    fn empty() -> Self {
        Self::default()
    }

    /// Sets retention SQL configuration.
    #[must_use]
    fn with_retention(mut self, delete: Delete) -> Self {
        self.retention_delete = Some(delete);
        self
    }

    /// Sets sort columns configuration.
    #[must_use]
    fn with_sort(mut self, columns: Vec<SortColumn>) -> Self {
        self.sort_columns = columns;
        self
    }
}

fn make_on_refresh_write_handler(
    dataset_name: String,
    config: OnRefreshConfig,
) -> WriteCompletionHandler {
    Arc::new(move |tx, table_manager, _schema, inserted_rows| {
        let internal_table_name = table_manager.table_name().to_string();

        // Apply retention SQL if configured
        if let Some(ref parsed_delete) = config.retention_delete {
            tracing::debug!(
                dataset = %dataset_name,
                table = %internal_table_name,
                inserted_rows,
                "Applying retention SQL before commit"
            );

            let reconstructed_sql = match reconstruct_retention_sql_with_table_name(
                parsed_delete,
                &internal_table_name,
            ) {
                Ok(sql) => sql,
                Err(e) => {
                    return Err(DataFusionError::Execution(format!(
                        "Failed to reconstruct retention SQL for dataset {dataset_name}: {e}"
                    )));
                }
            };

            tracing::debug!(
                dataset = %dataset_name,
                table = %internal_table_name,
                sql = %reconstructed_sql,
                "Reconstructed retention SQL with internal table name"
            );

            match tx.execute(reconstructed_sql.as_str(), []) {
                Ok(affected_rows) => {
                    tracing::debug!(
                        dataset = %dataset_name,
                        table = %internal_table_name,
                        affected_rows,
                        "Retention SQL applied before commit"
                    );
                }
                Err(err) => {
                    return Err(DataFusionError::Execution(format!(
                        "Failed to apply retention SQL for dataset {dataset_name} (table {internal_table_name}): {err}"
                    )));
                }
            }
        }

        // Current sorting implementation is experimental. CREATE OR REPLACE TABLE recreates
        // the table from scratch, which means:
        // - Primary keys and unique constraints (table-level definitions) are NOT preserved
        // - Indexes (separate DB objects) are dropped when the original table is replaced
        // - Foreign key constraints are NOT preserved
        if !config.sort_columns.is_empty() {
            let order_by_clause: String = config
                .sort_columns
                .iter()
                .map(|sc| format!("\"{}\" {}", sc.column, sc.direction))
                .collect::<Vec<_>>()
                .join(", ");

            tracing::debug!(
                dataset = %dataset_name,
                table = %internal_table_name,
                inserted_rows,
                order_by = %order_by_clause,
                "Applying on-refresh sort before commit"
            );

            let sort_sql = format!(
                "CREATE OR REPLACE TABLE \"{internal_table_name}\" AS SELECT * FROM \"{internal_table_name}\" ORDER BY {order_by_clause}"
            );

            tracing::debug!(
                dataset = %dataset_name,
                table = %internal_table_name,
                sql = %sort_sql,
                "Executing on-refresh sort SQL"
            );

            if let Err(err) = tx.execute(sort_sql.as_str(), []) {
                return Err(DataFusionError::Execution(format!(
                    "Failed to apply on-refresh sort for dataset {dataset_name} (table {internal_table_name}): {err}"
                )));
            }

            tracing::debug!(
                dataset = %dataset_name,
                table = %internal_table_name,
                "On-refresh sort applied before commit"
            );
        }

        table_manager.create_indexes(tx).map_err(|err| {
            DataFusionError::Execution(format!(
                "Failed to create DuckDB indexes for dataset {dataset_name} (table {internal_table_name}) before refresh commit: {err}"
            ))
        })?;

        Ok(())
    })
}

register_data_accelerator!(Engine::DuckDB, DuckDBAccelerator);

#[cfg(test)]
mod tests {
    use std::{collections::HashMap, sync::Arc};

    use crate::component::dataset::builder::DatasetBuilder;
    use arrow::{
        array::{Int64Array, RecordBatch, StringArray, TimestampSecondArray, UInt64Array},
        datatypes::{DataType, Field, Schema},
    };
    use datafusion::{
        common::{Constraints, TableReference, ToDFSchema},
        execution::context::SessionContext,
        logical_expr::{CreateExternalTable, cast, col, dml::InsertOp, lit},
        physical_plan::collect,
        scalar::ScalarValue,
    };
    use datafusion_table_providers::util::test::MockExec;

    use crate::component::dataset::acceleration::Acceleration;
    use crate::component::dataset::acceleration::{Engine, Mode};
    use crate::dataaccelerator::{DataAccelerator, duckdb::DuckDBAccelerator};

    fn external_table_with_options(options: HashMap<String, String>) -> CreateExternalTable {
        let schema = Arc::new(Schema::new(vec![Field::new(
            "value",
            DataType::Int64,
            false,
        )]));
        let df_schema = ToDFSchema::to_dfschema_ref(schema)
            .expect("to convert Arrow schema to DataFusion schema");

        CreateExternalTable {
            schema: df_schema,
            name: TableReference::bare("write_settings_table"),
            location: String::new(),
            file_type: String::new(),
            table_partition_cols: vec![],
            if_not_exists: true,
            or_replace: false,
            definition: None,
            order_exprs: vec![],
            unbounded: false,
            options,
            constraints: Constraints::new_unverified(vec![]),
            column_defaults: HashMap::default(),
            temporary: false,
        }
    }

    #[test]
    fn duckdb_write_settings_changes_refresh_disables_recompute_statistics_by_default() {
        let mut external_table = external_table_with_options(HashMap::new());

        super::apply_changes_refresh_write_defaults(&mut external_table, true);

        assert_eq!(
            external_table.options.get("recompute_statistics_on_write"),
            Some(&"false".to_string())
        );
    }

    #[test]
    fn duckdb_write_settings_changes_refresh_preserves_explicit_recompute_statistics_setting() {
        let mut options = HashMap::new();
        options.insert(
            "recompute_statistics_on_write".to_string(),
            "true".to_string(),
        );
        let mut external_table = external_table_with_options(options);

        super::apply_changes_refresh_write_defaults(&mut external_table, true);

        assert_eq!(
            external_table.options.get("recompute_statistics_on_write"),
            Some(&"true".to_string())
        );
    }

    #[test]
    fn duckdb_write_settings_non_changes_refresh_keeps_recompute_statistics_unset() {
        let mut external_table = external_table_with_options(HashMap::new());

        super::apply_changes_refresh_write_defaults(&mut external_table, false);

        assert!(
            !external_table
                .options
                .contains_key("recompute_statistics_on_write")
        );
    }

    #[tokio::test]
    async fn retention_sql_applies_before_commit() {
        let schema = Arc::new(Schema::new(vec![Field::new(
            "value",
            DataType::Int64,
            false,
        )]));
        let df_schema = ToDFSchema::to_dfschema_ref(Arc::clone(&schema))
            .expect("to convert Arrow schema to DataFusion schema");

        let external_table = CreateExternalTable {
            schema: df_schema,
            name: TableReference::bare("retention_table"),
            location: String::new(),
            file_type: String::new(),
            table_partition_cols: vec![],
            if_not_exists: true,
            or_replace: false,
            definition: None,
            order_exprs: vec![],
            unbounded: false,
            options: HashMap::new(),
            constraints: Constraints::new_unverified(vec![]),
            column_defaults: HashMap::default(),
            temporary: false,
        };

        let duckdb_accelerator = DuckDBAccelerator::new();
        let retention_sql = "DELETE FROM retention_table WHERE value < 5";
        let parsed_delete = crate::datafusion::retention_sql::parse_retention_sql(
            &TableReference::bare("retention_table"),
            retention_sql,
            Arc::clone(&schema),
        )
        .expect("should parse retention SQL")
        .delete_statement;
        let handler = super::make_on_refresh_write_handler(
            "retention_dataset".to_string(),
            super::OnRefreshConfig::empty().with_retention(parsed_delete),
        );

        let table = super::create_table_provider(
            &duckdb_accelerator.duckdb_factory,
            &external_table,
            Some(handler),
        )
        .await
        .expect("table should be created");

        let input = RecordBatch::try_new(
            Arc::clone(&schema),
            vec![Arc::new(Int64Array::from(vec![1, 3, 5, 7]))],
        )
        .expect("to create RecordBatch");

        let exec = Arc::new(MockExec::new(vec![Ok(input)], schema));

        let write_ctx = SessionContext::new();
        let insert_plan = table
            .insert_into(
                &write_ctx.state(),
                Arc::<MockExec>::clone(&exec),
                InsertOp::Append,
            )
            .await
            .expect("to create insert plan");

        collect(insert_plan, write_ctx.task_ctx())
            .await
            .expect("to execute insert");

        let read_ctx = SessionContext::new();
        let scan_plan = table
            .scan(&read_ctx.state(), None, &[], None)
            .await
            .expect("to create scan plan");

        let batches = collect(scan_plan, read_ctx.task_ctx())
            .await
            .expect("to execute scan");

        let mut values = Vec::new();
        for batch in &batches {
            let column = batch
                .column(0)
                .as_any()
                .downcast_ref::<Int64Array>()
                .expect("to downcast column to Int64Array");
            values.extend((0..column.len()).map(|idx| column.value(idx)));
        }

        assert_eq!(values, vec![5, 7]);
    }

    #[tokio::test]
    async fn overwrite_index_failure_keeps_previous_duckdb_view() {
        let schema = Arc::new(Schema::new(vec![Field::new(
            "value",
            DataType::Int64,
            false,
        )]));
        let df_schema = ToDFSchema::to_dfschema_ref(Arc::clone(&schema))
            .expect("to convert Arrow schema to DataFusion schema");

        let mut options = HashMap::new();
        options.insert("indexes".to_string(), "value:unique".to_string());

        let external_table = CreateExternalTable {
            schema: df_schema,
            name: TableReference::bare("indexed_overwrite_table"),
            location: String::new(),
            file_type: String::new(),
            table_partition_cols: vec![],
            if_not_exists: true,
            or_replace: false,
            definition: None,
            order_exprs: vec![],
            unbounded: false,
            options,
            constraints: Constraints::new_unverified(vec![]),
            column_defaults: HashMap::default(),
            temporary: false,
        };

        let duckdb_accelerator = DuckDBAccelerator::new();
        let handler = super::make_on_refresh_write_handler(
            "indexed_overwrite_dataset".to_string(),
            super::OnRefreshConfig::empty(),
        );

        let table = super::create_table_provider(
            &duckdb_accelerator.duckdb_factory,
            &external_table,
            Some(handler),
        )
        .await
        .expect("table should be created");

        let write_ctx = SessionContext::new();
        let initial_input = RecordBatch::try_new(
            Arc::clone(&schema),
            vec![Arc::new(Int64Array::from(vec![1, 2]))],
        )
        .expect("to create initial RecordBatch");
        let initial_exec = Arc::new(MockExec::new(vec![Ok(initial_input)], Arc::clone(&schema)));
        let initial_insert = table
            .insert_into(&write_ctx.state(), initial_exec, InsertOp::Overwrite)
            .await
            .expect("to create initial insert plan");
        collect(initial_insert, write_ctx.task_ctx())
            .await
            .expect("initial overwrite should succeed");

        let duplicate_input = RecordBatch::try_new(
            Arc::clone(&schema),
            vec![Arc::new(Int64Array::from(vec![3, 3]))],
        )
        .expect("to create duplicate RecordBatch");
        let duplicate_exec = Arc::new(MockExec::new(
            vec![Ok(duplicate_input)],
            Arc::clone(&schema),
        ));
        let duplicate_insert = table
            .insert_into(&write_ctx.state(), duplicate_exec, InsertOp::Overwrite)
            .await
            .expect("to create duplicate insert plan");
        let duplicate_result = collect(duplicate_insert, write_ctx.task_ctx()).await;
        assert!(
            duplicate_result.is_err(),
            "duplicate unique-index overwrite should fail"
        );

        let read_ctx = SessionContext::new();
        let scan_plan = table
            .scan(&read_ctx.state(), None, &[], None)
            .await
            .expect("to create scan plan");
        let batches = collect(scan_plan, read_ctx.task_ctx())
            .await
            .expect("to execute scan");

        let mut values = Vec::new();
        for batch in &batches {
            let column = batch
                .column(0)
                .as_any()
                .downcast_ref::<Int64Array>()
                .expect("to downcast column to Int64Array");
            values.extend((0..column.len()).map(|idx| column.value(idx)));
        }
        values.sort_unstable();

        assert_eq!(values, vec![1, 2]);
    }

    #[tokio::test]
    async fn retention_sql_fails_with_internal_tables() {
        use datafusion_table_providers::duckdb::DuckDB;
        use datafusion_table_providers::sql::db_connection_pool::duckdbpool::DuckDbConnectionPool;
        use tempfile::TempDir;

        // This test reproduces the bug where retention SQL fails with internal tables.
        //
        // When DuckDB uses internal tables (for indexes/constraints via preserve_insertion_order),
        // the write completion handler receives the internal table name (like __data_table_123)
        // from table_manager.table_name(), but the retention SQL references the logical table name.
        //
        // DuckDB's error: "Can only delete from base table!" occurs because DELETE statements
        // must target the base/view table name, not the internal table directly.

        let temp_dir = TempDir::new().expect("to create temp directory");
        let db_path = temp_dir.path().join("test_retention.db");

        let schema = Arc::new(Schema::new(vec![
            Field::new("id", DataType::Int64, false),
            Field::new("value", DataType::Int64, false),
        ]));
        let df_schema = ToDFSchema::to_dfschema_ref(Arc::clone(&schema))
            .expect("to convert Arrow schema to DataFusion schema");

        let mut options = HashMap::new();
        // Use file mode to enable full DuckDB features
        options.insert(
            "open".to_string(),
            db_path
                .to_str()
                .expect("to convert path to string")
                .to_string(),
        );
        // Enable preserve_insertion_order which triggers internal table creation
        options.insert("preserve_insertion_order".to_string(), "true".to_string());

        let external_table = CreateExternalTable {
            schema: df_schema,
            name: TableReference::bare("taxi_trips"),
            location: String::new(),
            file_type: String::new(),
            table_partition_cols: vec![],
            if_not_exists: true,
            or_replace: false,
            definition: None,
            order_exprs: vec![],
            unbounded: false,
            options,
            constraints: Constraints::new_unverified(vec![]),
            column_defaults: HashMap::default(),
            temporary: false,
        };

        let duckdb_accelerator = DuckDBAccelerator::new();

        // The retention SQL references the logical table name "taxi_trips"
        let retention_sql = "DELETE FROM taxi_trips WHERE value < 5";
        let parsed_delete = crate::datafusion::retention_sql::parse_retention_sql(
            &TableReference::bare("taxi_trips"),
            retention_sql,
            Arc::clone(&schema),
        )
        .expect("should parse retention SQL")
        .delete_statement;
        let handler = super::make_on_refresh_write_handler(
            "taxi_trips".to_string(),
            super::OnRefreshConfig::empty().with_retention(parsed_delete),
        );

        let table = super::create_table_provider(
            &duckdb_accelerator.duckdb_factory,
            &external_table,
            Some(handler),
        )
        .await
        .expect("table should be created");

        // Insert initial data
        let input = RecordBatch::try_new(
            Arc::clone(&schema),
            vec![
                Arc::new(Int64Array::from(vec![1, 2, 3, 4])),
                Arc::new(Int64Array::from(vec![1, 3, 5, 7])),
            ],
        )
        .expect("to create RecordBatch");

        let exec = Arc::new(MockExec::new(vec![Ok(input.clone())], Arc::clone(&schema)));

        let write_ctx = SessionContext::new();
        let insert_plan = table
            .insert_into(
                &write_ctx.state(),
                Arc::<MockExec>::clone(&exec),
                InsertOp::Append,
            )
            .await
            .expect("to create insert plan");

        // First insert should succeed
        collect(insert_plan, write_ctx.task_ctx())
            .await
            .expect("first insert should succeed");

        // Verify internal tables were created by checking DuckDB directly
        let pool = Arc::new(
            DuckDbConnectionPool::new_file(
                db_path.to_str().expect("to convert path to string"),
                &duckdb::AccessMode::ReadWrite,
            )
            .expect("to create DuckDB connection pool"),
        );

        let mut conn = pool.connect_sync().expect("to get connection from pool");
        let duckdb_conn = DuckDB::duckdb_conn(&mut conn).expect("to get DuckDB connection");

        // Check for internal tables (they follow the pattern __data_*)
        let internal_tables: Vec<String> = duckdb_conn
            .get_underlying_conn_mut()
            .prepare(
                "SELECT table_name FROM information_schema.tables WHERE table_name LIKE '__data_%'",
            )
            .expect("to prepare SQL statement")
            .query_map([], |row| row.get(0))
            .expect("to execute query")
            .collect::<Result<Vec<_>, _>>()
            .expect("to collect query results");

        if internal_tables.is_empty() {
            eprintln!(
                "WARNING: No internal tables found. Test may not be validating the bug correctly."
            );
            eprintln!(
                "This could mean preserve_insertion_order didn't trigger internal table creation."
            );
            return;
        }

        eprintln!("Found internal tables: {internal_tables:?}");

        // Now try to insert more data - this should trigger the retention SQL
        // and fail because it tries to DELETE from the internal table name
        let exec2 = Arc::new(MockExec::new(vec![Ok(input)], schema));

        let insert_plan2 = table
            .insert_into(&write_ctx.state(), exec2, InsertOp::Append)
            .await
            .expect("to create second insert plan");

        let result = collect(insert_plan2, write_ctx.task_ctx()).await;

        // This should fail with "Can only delete from base table!"
        assert!(
            result.is_err(),
            "Expected an error due to retention SQL targeting internal table, but insert succeeded"
        );

        let error_msg = result.expect_err("Expected error").to_string();
        assert!(
            error_msg.contains("Can only delete from base table")
                || error_msg.contains("Binder Error")
                || error_msg.contains("to apply retention SQL"),
            "Expected error about deleting from base table, got: {error_msg}"
        );

        eprintln!("✓ Test correctly reproduced the error: {error_msg}");
    }

    #[tokio::test]
    #[expect(clippy::unreadable_literal)]
    async fn test_round_trip_duckdb() {
        let schema = Arc::new(Schema::new(vec![
            arrow::datatypes::Field::new("time_in_string", DataType::Utf8, false),
            arrow::datatypes::Field::new(
                "time",
                DataType::Timestamp(arrow::datatypes::TimeUnit::Second, None),
                false,
            ),
            arrow::datatypes::Field::new("time_int", DataType::Int64, false),
            arrow::datatypes::Field::new(
                "time_with_zone",
                DataType::Timestamp(
                    arrow::datatypes::TimeUnit::Second,
                    Some("Etc/UTC".to_string().into()),
                ),
                false,
            ),
        ]));
        let df_schema = ToDFSchema::to_dfschema_ref(Arc::clone(&schema))
            .expect("to convert Arrow schema to DataFusion schema");
        let external_table = CreateExternalTable {
            schema: df_schema,
            name: TableReference::bare("test_table"),
            location: String::new(),
            file_type: String::new(),
            table_partition_cols: vec![],
            if_not_exists: true,
            or_replace: false,
            definition: None,
            order_exprs: vec![],
            unbounded: false,
            options: HashMap::new(),
            constraints: Constraints::new_unverified(vec![]),
            column_defaults: HashMap::default(),
            temporary: false,
        };
        let duckdb_accelerator = DuckDBAccelerator::new();
        let ctx = SessionContext::new();
        let table = duckdb_accelerator
            .create_external_table(external_table, None, vec![], None)
            .await
            .expect("table should be created");

        let arr1 = StringArray::from(vec![
            "1970-01-01",
            "2012-12-01T11:11:11Z",
            "2012-12-01T11:11:12Z",
        ]);
        let arr2 = TimestampSecondArray::from(vec![0, 1354360271, 1354360272]);
        let arr3 = Int64Array::from(vec![0, 1354360271, 1354360272]);
        let arr4 = arrow::compute::cast(
            &arr2,
            &DataType::Timestamp(
                arrow::datatypes::TimeUnit::Second,
                Some("Etc/UTC".to_string().into()),
            ),
        )
        .expect("casting works");
        let data = RecordBatch::try_new(
            Arc::clone(&schema),
            vec![
                Arc::new(arr1),
                Arc::new(arr2),
                Arc::new(arr3),
                Arc::new(arr4),
            ],
        )
        .expect("data should be created");

        let exec = Arc::new(MockExec::new(vec![Ok(data)], schema));

        let insertion = table
            .insert_into(
                &ctx.state(),
                Arc::<MockExec>::clone(&exec),
                InsertOp::Append,
            )
            .await
            .expect("insertion should be successful");

        collect(insertion, ctx.task_ctx())
            .await
            .expect("insert successful");

        let filter = cast(
            col("time_in_string"),
            DataType::Timestamp(arrow::datatypes::TimeUnit::Millisecond, None),
        )
        .lt(lit(ScalarValue::TimestampMillisecond(
            Some(1354360272000),
            None,
        )));
        let plan = table
            .delete_from(&ctx.state(), vec![filter])
            .await
            .expect("deletion should be successful");

        let result = collect(plan, ctx.task_ctx())
            .await
            .expect("deletion successful");
        let actual = result
            .first()
            .expect("result should have at least one batch")
            .column(0)
            .as_any()
            .downcast_ref::<UInt64Array>()
            .expect("result should be UInt64Array");
        let expected = UInt64Array::from(vec![2]);
        assert_eq!(actual, &expected);

        let filter = col("time_int").lt(lit(1354360273));
        let plan = table
            .delete_from(&ctx.state(), vec![filter])
            .await
            .expect("deletion should be successful");

        let result = collect(plan, ctx.task_ctx())
            .await
            .expect("deletion successful");
        let actual = result
            .first()
            .expect("result should have at least one batch")
            .column(0)
            .as_any()
            .downcast_ref::<UInt64Array>()
            .expect("result should be UInt64Array");
        let expected = UInt64Array::from(vec![1]);
        assert_eq!(actual, &expected);

        let insertion = table
            .insert_into(
                &ctx.state(),
                Arc::<MockExec>::clone(&exec),
                InsertOp::Append,
            )
            .await
            .expect("insertion should be successful");

        collect(insertion, ctx.task_ctx())
            .await
            .expect("insert successful");

        let filter = col("time").lt(lit(ScalarValue::TimestampMillisecond(
            Some(1354360272000),
            None,
        )));
        let plan = table
            .delete_from(&ctx.state(), vec![filter])
            .await
            .expect("deletion should be successful");

        let result = collect(plan, ctx.task_ctx())
            .await
            .expect("deletion successful");
        let actual = result
            .first()
            .expect("result should have at least one batch")
            .column(0)
            .as_any()
            .downcast_ref::<UInt64Array>()
            .expect("result should be UInt64Array");
        let expected = UInt64Array::from(vec![2]);
        assert_eq!(actual, &expected);

        let insertion = table
            .insert_into(&ctx.state(), exec, InsertOp::Append)
            .await
            .expect("insertion should be successful");

        collect(insertion, ctx.task_ctx())
            .await
            .expect("insert successful");

        let filter = col("time_with_zone").lt(lit(ScalarValue::TimestampMillisecond(
            Some(1354360272000),
            None,
        )));
        let plan = table
            .delete_from(&ctx.state(), vec![filter])
            .await
            .expect("deletion should be successful");

        let result = collect(plan, ctx.task_ctx())
            .await
            .expect("deletion successful");
        let actual = result
            .first()
            .expect("result should have at least one batch")
            .column(0)
            .as_any()
            .downcast_ref::<UInt64Array>()
            .expect("result should be UInt64Array");
        let expected = UInt64Array::from(vec![2]);
        assert_eq!(actual, &expected);
    }

    #[tokio::test]
    async fn test_duckdb_file_initialization() {
        let app = app::AppBuilder::new("test").build();
        let rt = crate::Runtime::builder().build().await;

        let mut dataset = DatasetBuilder::try_new(
            "duckdb_file_accelerator_init".to_string(),
            "duckdb_file_accelerator_init",
        )
        .expect("to create builder")
        .with_app(Arc::new(app))
        .with_runtime(Arc::new(rt))
        .build()
        .expect("to build dataset");

        dataset.acceleration = Some(Acceleration {
            engine: Engine::DuckDB,
            mode: Mode::File,
            ..Default::default()
        });

        let accelerator = DuckDBAccelerator::new();
        assert!(!accelerator.is_initialized(&dataset));

        accelerator
            .init(&dataset)
            .await
            .expect("initialization should be successful");

        assert!(accelerator.is_initialized(&dataset));

        let path = accelerator.file_path(&dataset).expect("path should exist");
        assert!(std::path::Path::new(&path).exists());

        // cleanup
        std::fs::remove_file(&path).expect("file should be removed");
    }

    #[tokio::test]
    async fn test_retention_sql_with_duckdb_accelerator() {
        use tempfile::TempDir;

        // Create a temporary directory for the DuckDB file
        let temp_dir = TempDir::new().expect("to create temp dir");
        let db_path = temp_dir.path().join("test_retention.db");

        // Create schema
        let schema = Arc::new(Schema::new(vec![
            Field::new("id", DataType::Int64, false),
            Field::new("value", DataType::Int64, false),
        ]));
        let df_schema = ToDFSchema::to_dfschema_ref(Arc::clone(&schema))
            .expect("to convert Arrow schema to DataFusion schema");

        // Prepare the external table command with file path
        let mut external_table = CreateExternalTable {
            schema: df_schema,
            name: TableReference::bare("retention_test_dataset"),
            location: String::new(),
            file_type: String::new(),
            table_partition_cols: vec![],
            if_not_exists: true,
            or_replace: false,
            definition: None,
            order_exprs: vec![],
            unbounded: false,
            options: HashMap::new(),
            constraints: Constraints::new_unverified(vec![]),
            column_defaults: HashMap::default(),
            temporary: false,
        };
        external_table.options.insert(
            "open".to_string(),
            db_path
                .to_str()
                .expect("to convert path to string")
                .to_string(),
        );

        // Parse retention SQL and create handler
        let retention_sql = "DELETE FROM retention_test_dataset WHERE value < 5";
        let parsed_delete = crate::datafusion::retention_sql::parse_retention_sql(
            &TableReference::bare("retention_test_dataset"),
            retention_sql,
            Arc::clone(&schema),
        )
        .expect("should parse retention SQL")
        .delete_statement;
        let handler = super::make_on_refresh_write_handler(
            "retention_test_dataset".to_string(),
            super::OnRefreshConfig::empty().with_retention(parsed_delete),
        );

        // Create the accelerator and table
        let accelerator = DuckDBAccelerator::new();
        let table = super::create_table_provider(
            &accelerator.duckdb_factory,
            &external_table,
            Some(handler),
        )
        .await
        .expect("table should be created");

        // Insert initial data with values both above and below the retention threshold
        let input = RecordBatch::try_new(
            Arc::clone(&schema),
            vec![
                Arc::new(Int64Array::from(vec![1, 2, 3, 4, 5, 6])),
                Arc::new(Int64Array::from(vec![2, 3, 4, 6, 7, 8])), // values: 2, 3, 4 should be deleted (< 5)
            ],
        )
        .expect("to create RecordBatch");

        let exec = Arc::new(MockExec::new(vec![Ok(input.clone())], Arc::clone(&schema)));

        let write_ctx = SessionContext::new();
        let insert_plan = table
            .insert_into(
                &write_ctx.state(),
                Arc::<MockExec>::clone(&exec),
                InsertOp::Append,
            )
            .await
            .expect("to create insert plan");

        // Execute the insert - this should trigger the retention SQL
        collect(insert_plan, write_ctx.task_ctx())
            .await
            .expect("insert should succeed");

        // Query the table to verify retention SQL was applied
        let read_ctx = SessionContext::new();
        let scan = table
            .scan(&read_ctx.state(), None, &[], None)
            .await
            .expect("scan should succeed");

        let results = collect(scan, read_ctx.task_ctx())
            .await
            .expect("collect should succeed");

        // Verify that only rows with value >= 5 remain
        let mut total_rows = 0;
        let mut values = Vec::new();
        for batch in &results {
            total_rows += batch.num_rows();
            let value_array = batch
                .column(1)
                .as_any()
                .downcast_ref::<Int64Array>()
                .expect("value column should be Int64Array");

            // Collect all values
            for i in 0..value_array.len() {
                values.push(value_array.value(i));
            }
        }

        // All values should be >= 5
        for value in &values {
            assert!(
                *value >= 5,
                "Found value {value} which should have been deleted by retention SQL"
            );
        }

        // We should have 3 rows remaining (values 6, 7, 8)
        assert_eq!(
            total_rows, 3,
            "Expected 3 rows after retention (values >= 5), found {total_rows}. Values: {values:?}"
        );

        // cleanup
        drop(table);
        drop(temp_dir);
    }

    #[tokio::test]
    async fn test_reconstruct_retention_sql() {
        let sql = "DELETE FROM taxi_trips WHERE status = 'expired'";
        let schema = Arc::new(Schema::new(vec![Field::new(
            "status",
            DataType::Utf8,
            true,
        )]));
        let parsed = crate::datafusion::retention_sql::parse_retention_sql(
            &TableReference::bare("taxi_trips"),
            sql,
            schema,
        )
        .expect("should parse");

        let internal_name = "__data_taxi_trips_1234567890";
        let result = super::reconstruct_retention_sql_with_table_name(
            &parsed.delete_statement,
            internal_name,
        );

        assert!(result.is_ok(), "Should reconstruct SQL successfully");
        let reconstructed = result.expect("reconstructed");

        // Verify the internal table name is used
        assert!(
            reconstructed.contains(internal_name),
            "Should contain internal table name"
        );

        // Verify the WHERE clause is preserved
        assert!(
            reconstructed.contains("status = 'expired'")
                || reconstructed.contains("status = \"expired\""),
            "Should preserve WHERE clause"
        );

        // Verify it's still a DELETE statement
        assert!(
            reconstructed.to_lowercase().starts_with("delete from"),
            "Should start with DELETE FROM"
        );
    }

    #[tokio::test]
    async fn test_reconstruct_retention_sql_complex_where() {
        let sql = "DELETE FROM orders WHERE created_at < NOW() - INTERVAL '30 days' AND status IN ('cancelled', 'expired')";
        let schema = Arc::new(Schema::new(vec![
            Field::new("created_at", DataType::Utf8, true),
            Field::new("status", DataType::Utf8, true),
        ]));
        let parsed = crate::datafusion::retention_sql::parse_retention_sql(
            &TableReference::bare("orders"),
            sql,
            schema,
        )
        .expect("should parse");

        let internal_name = "__data_orders_9876543210";
        let result = super::reconstruct_retention_sql_with_table_name(
            &parsed.delete_statement,
            internal_name,
        );

        assert!(
            result.is_ok(),
            "Should reconstruct complex SQL successfully"
        );
        let reconstructed = result.expect("reconstructed");

        // Verify the internal table name is used
        assert!(
            reconstructed.contains(internal_name),
            "Should contain internal table name"
        );

        // Basic sanity check - make sure it's still a valid DELETE statement structure
        assert!(
            reconstructed.to_lowercase().starts_with("delete from"),
            "Should start with DELETE FROM"
        );
    }

    #[tokio::test]
    async fn test_sort_columns_with_duckdb_accelerator() {
        use crate::datafusion::sort_columns::SortColumn;
        use tempfile::TempDir;

        // Create a temporary directory for the DuckDB file
        let temp_dir = TempDir::new().expect("to create temp dir");
        let db_path = temp_dir.path().join("test_sort.db");

        // Create schema
        let schema = Arc::new(Schema::new(vec![
            Field::new("id", DataType::Int64, false),
            Field::new("value", DataType::Int64, false),
        ]));
        let df_schema = ToDFSchema::to_dfschema_ref(Arc::clone(&schema))
            .expect("to convert Arrow schema to DataFusion schema");

        // Prepare the external table command with file path
        let mut external_table = CreateExternalTable {
            schema: df_schema,
            name: TableReference::bare("sort_test_dataset"),
            location: String::new(),
            file_type: String::new(),
            table_partition_cols: vec![],
            if_not_exists: true,
            or_replace: false,
            definition: None,
            order_exprs: vec![],
            unbounded: false,
            options: HashMap::new(),
            constraints: Constraints::new_unverified(vec![]),
            column_defaults: HashMap::default(),
            temporary: false,
        };
        external_table.options.insert(
            "open".to_string(),
            db_path
                .to_str()
                .expect("to convert path to string")
                .to_string(),
        );

        // Create sort columns configuration - sort by value DESC
        let sort_columns = vec![SortColumn::desc("value")];
        let handler = super::make_on_refresh_write_handler(
            "sort_test_dataset".to_string(),
            super::OnRefreshConfig::empty().with_sort(sort_columns),
        );

        // Create the accelerator and table
        let accelerator = DuckDBAccelerator::new();
        let table = super::create_table_provider(
            &accelerator.duckdb_factory,
            &external_table,
            Some(handler),
        )
        .await
        .expect("table should be created");

        // Insert data in unsorted order
        let input = RecordBatch::try_new(
            Arc::clone(&schema),
            vec![
                Arc::new(Int64Array::from(vec![1, 2, 3, 4, 5])),
                Arc::new(Int64Array::from(vec![30, 10, 50, 20, 40])), // unsorted values
            ],
        )
        .expect("to create RecordBatch");

        let exec = Arc::new(MockExec::new(vec![Ok(input.clone())], Arc::clone(&schema)));

        let write_ctx = SessionContext::new();
        let insert_plan = table
            .insert_into(
                &write_ctx.state(),
                Arc::<MockExec>::clone(&exec),
                InsertOp::Append,
            )
            .await
            .expect("to create insert plan");

        // Execute the insert - this should trigger the sort
        collect(insert_plan, write_ctx.task_ctx())
            .await
            .expect("insert should succeed");

        // Query the table to verify sorting was applied
        let read_ctx = SessionContext::new();
        let scan = table
            .scan(&read_ctx.state(), None, &[], None)
            .await
            .expect("scan should succeed");

        let results = collect(scan, read_ctx.task_ctx())
            .await
            .expect("collect should succeed");

        // Collect all values in order
        let mut values = Vec::new();
        for batch in &results {
            let value_array = batch
                .column(1)
                .as_any()
                .downcast_ref::<Int64Array>()
                .expect("value column should be Int64Array");

            for i in 0..value_array.len() {
                values.push(value_array.value(i));
            }
        }

        // Verify data is sorted in descending order
        assert_eq!(
            values,
            vec![50, 40, 30, 20, 10],
            "Data should be sorted by value DESC"
        );

        // cleanup
        drop(table);
        drop(temp_dir);
    }

    /// Regression test for <https://github.com/spiceai/spiceai/issues/2889>.
    ///
    /// Arrow Dictionary-encoded columns (enums) must be transparently unpacked
    /// to their value types before reaching the `DuckDB` accelerator.
    #[tokio::test]
    async fn test_duckdb_dictionary_type_round_trip() {
        use arrow::array::StringDictionaryBuilder;
        use arrow::datatypes::Int32Type;

        // Build a schema containing a Dictionary(Int32, Utf8) column.
        let source_schema = Arc::new(Schema::new(vec![
            Field::new("id", DataType::Int64, false),
            Field::new(
                "status",
                DataType::Dictionary(Box::new(DataType::Int32), Box::new(DataType::Utf8)),
                false,
            ),
        ]));

        // Normalize the schema (Dictionary -> Utf8).
        let accel_schema = Arc::new(arrow_tools::schema::normalize_dictionary_types(
            &source_schema,
        ));
        assert_eq!(accel_schema.field(1).data_type(), &DataType::Utf8);

        let df_schema = ToDFSchema::to_dfschema_ref(Arc::clone(&accel_schema)).expect("df schema");

        let external_table = CreateExternalTable {
            schema: df_schema,
            name: TableReference::bare("dict_test"),
            location: String::new(),
            file_type: String::new(),
            table_partition_cols: vec![],
            if_not_exists: true,
            or_replace: false,
            definition: None,
            order_exprs: vec![],
            unbounded: false,
            options: HashMap::new(),
            constraints: Constraints::new_unverified(vec![]),
            column_defaults: HashMap::default(),
            temporary: false,
        };

        let duckdb_accelerator = DuckDBAccelerator::new();
        let table = duckdb_accelerator
            .create_external_table(external_table, None, vec![], None)
            .await
            .expect("DuckDB table with normalized Dictionary types should be created");

        // Build a record batch with Dictionary-encoded data.
        let ids = Int64Array::from(vec![1, 2, 3]);
        let mut status_builder = StringDictionaryBuilder::<Int32Type>::new();
        status_builder.append_value("active");
        status_builder.append_value("inactive");
        status_builder.append_value("active");
        let statuses = status_builder.finish();

        let data = RecordBatch::try_new(
            Arc::clone(&source_schema),
            vec![Arc::new(ids), Arc::new(statuses)],
        )
        .expect("record batch");

        // Cast from Dictionary to the normalized schema before inserting.
        let casted = arrow_tools::record_batch::try_cast_to(data, Arc::clone(&accel_schema))
            .expect("cast Dictionary to Utf8");

        let exec = MockExec::new(vec![Ok(casted)], accel_schema);
        let ctx = SessionContext::new();

        let insertion = table
            .insert_into(&ctx.state(), Arc::new(exec), InsertOp::Append)
            .await
            .expect("insertion of Dictionary data into DuckDB should succeed");

        let result = collect(insertion, ctx.task_ctx())
            .await
            .expect("insert should succeed");

        assert!(!result.is_empty());

        // Verify the data can be read back.
        let scan = table
            .scan(&ctx.state(), None, &[], None)
            .await
            .expect("scan should succeed");

        let batches = collect(scan, ctx.task_ctx())
            .await
            .expect("should read back data");

        let total_rows: usize = batches.iter().map(RecordBatch::num_rows).sum();
        assert_eq!(total_rows, 3, "should have 3 rows");
    }

    /// Tests that the DROP TABLE SQL used by `drop_table` correctly removes a table.
    #[tokio::test]
    async fn test_drop_table_sql_removes_table() {
        use datafusion_table_providers::duckdb::DuckDB;
        use datafusion_table_providers::sql::db_connection_pool::duckdbpool::DuckDbConnectionPool;

        let pool =
            Arc::new(DuckDbConnectionPool::new_memory().expect("to create DuckDB connection pool"));
        let mut conn = pool.connect_sync().expect("to get connection from pool");
        let duckdb_conn = DuckDB::duckdb_conn(&mut conn).expect("to get DuckDB connection");

        let table_name = "drop_test_table";
        let underlying = duckdb_conn.get_underlying_conn_mut();

        // Create a table
        underlying
            .execute(
                &format!("CREATE TABLE \"{table_name}\" (id INTEGER, name VARCHAR)"),
                [],
            )
            .expect("create table should succeed");

        // Insert data
        underlying
            .execute(
                &format!("INSERT INTO \"{table_name}\" VALUES (1, 'alice')"),
                [],
            )
            .expect("insert should succeed");

        // Verify table exists
        let count: i32 = underlying
            .query_row(
                "SELECT count(*) FROM information_schema.tables WHERE table_name = ?",
                [table_name],
                |row| row.get(0),
            )
            .expect("to query table count");
        assert_eq!(count, 1, "table should exist before drop");

        // Execute the same DROP TABLE SQL that drop_table() uses
        underlying
            .execute(&format!("DROP TABLE IF EXISTS \"{table_name}\""), [])
            .expect("drop should succeed");

        // Verify the table is gone
        let count: i32 = underlying
            .query_row(
                "SELECT count(*) FROM information_schema.tables WHERE table_name = ?",
                [table_name],
                |row| row.get(0),
            )
            .expect("to query table count");
        assert_eq!(count, 0, "table should not exist after drop");

        // Verify DROP IF EXISTS on non-existent table doesn't error
        underlying
            .execute(&format!("DROP TABLE IF EXISTS \"{table_name}\""), [])
            .expect("drop of non-existent table should succeed");
    }

    #[test]
    fn storage_profile_drives_setup_pragmas() {
        use crate::dataaccelerator::storage::ResolvedAccelerationStorage;

        // EBS bumps the checkpoint threshold to amortize remote-disk writes.
        let ebs = DuckDBAccelerator::storage_setup_queries(ResolvedAccelerationStorage::Ebs);
        assert!(
            ebs.iter().any(|q| q.contains("checkpoint_threshold")),
            "EBS profile should tune checkpoint_threshold, got {ebs:?}"
        );

        // Tmpfs also raises checkpoint threshold (volatile, RAM-backed).
        let tmpfs = DuckDBAccelerator::storage_setup_queries(ResolvedAccelerationStorage::Tmpfs);
        assert!(
            tmpfs.iter().any(|q| q.contains("checkpoint_threshold")),
            "Tmpfs profile should tune checkpoint_threshold, got {tmpfs:?}"
        );

        // Local SSD and Unknown keep DuckDB defaults.
        assert!(
            DuckDBAccelerator::storage_setup_queries(ResolvedAccelerationStorage::LocalSsd)
                .is_empty()
        );
        assert!(
            DuckDBAccelerator::storage_setup_queries(ResolvedAccelerationStorage::Unknown)
                .is_empty()
        );
    }
}