influxdb/iox_query/src/exec.rs

538 lines
17 KiB
Rust

//! This module handles the manipulation / execution of storage
//! plans. This is currently implemented using DataFusion, and this
//! interface abstracts away many of the details
pub(crate) mod context;
pub mod field;
pub mod fieldlist;
mod non_null_checker;
mod query_tracing;
mod schema_pivot;
pub mod seriesset;
pub(crate) mod split;
pub mod stringset;
use executor::DedicatedExecutor;
use object_store::DynObjectStore;
use parquet_file::storage::StorageId;
use trace::span::{SpanExt, SpanRecorder};
use std::{collections::HashMap, sync::Arc};
use datafusion::{
self,
execution::{
context::SessionState,
runtime_env::{RuntimeConfig, RuntimeEnv},
},
logical_expr::{expr_rewriter::normalize_col, Extension},
logical_expr::{Expr, LogicalPlan},
prelude::SessionContext,
};
pub use context::{IOxSessionConfig, IOxSessionContext, SessionContextIOxExt};
use schema_pivot::SchemaPivotNode;
use self::{non_null_checker::NonNullCheckerNode, split::StreamSplitNode};
/// Configuration for an Executor
#[derive(Debug, Clone)]
pub struct ExecutorConfig {
/// Number of threads per thread pool
pub num_threads: usize,
/// Target parallelism for query execution
pub target_query_partitions: usize,
/// Object stores
pub object_stores: HashMap<StorageId, Arc<DynObjectStore>>,
}
#[derive(Debug)]
pub struct DedicatedExecutors {
/// Executor for running user queries
query_exec: DedicatedExecutor,
/// Executor for running system/reorganization tasks such as
/// compact
reorg_exec: DedicatedExecutor,
/// Number of threads per thread pool
num_threads: usize,
}
impl DedicatedExecutors {
pub fn new(num_threads: usize) -> Self {
let query_exec = DedicatedExecutor::new("IOx Query Executor Thread", num_threads);
let reorg_exec = DedicatedExecutor::new("IOx Reorg Executor Thread", num_threads);
Self {
query_exec,
reorg_exec,
num_threads,
}
}
pub fn num_threads(&self) -> usize {
self.num_threads
}
}
/// Handles executing DataFusion plans, and marshalling the results into rust
/// native structures.
///
/// TODO: Have a resource manager that would limit how many plans are
/// running, based on a policy
#[derive(Debug)]
pub struct Executor {
/// Executors
executors: Arc<DedicatedExecutors>,
/// The default configuration options with which to create contexts
config: ExecutorConfig,
/// The DataFusion [RuntimeEnv] (including memory manager and disk
/// manager) used for all executions
runtime: Arc<RuntimeEnv>,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ExecutorType {
/// Run using the pool for queries
Query,
/// Run using the pool for system / reorganization tasks
Reorg,
}
impl Executor {
/// Creates a new executor with a two dedicated thread pools, each
/// with num_threads
pub fn new(num_threads: usize) -> Self {
Self::new_with_config(ExecutorConfig {
num_threads,
target_query_partitions: num_threads,
object_stores: HashMap::default(),
})
}
pub fn new_with_config(config: ExecutorConfig) -> Self {
let executors = Arc::new(DedicatedExecutors::new(config.num_threads));
Self::new_with_config_and_executors(config, executors)
}
/// Low-level constructor.
///
/// This is mostly useful if you wanna keep the executors (because they are quiet expensive to create) but need a fresh IOx runtime.
///
/// # Panic
/// Panics if the number of threads in `executors` is different from `config`.
pub fn new_with_config_and_executors(
config: ExecutorConfig,
executors: Arc<DedicatedExecutors>,
) -> Self {
assert_eq!(config.num_threads, executors.num_threads);
let runtime_config = RuntimeConfig::new();
for (id, store) in &config.object_stores {
runtime_config
.object_store_registry
.register_store("iox", id, Arc::clone(store));
}
let runtime = Arc::new(RuntimeEnv::new(runtime_config).expect("creating runtime"));
Self {
executors,
config,
runtime,
}
}
/// Return a new execution config, suitable for executing a new query or system task.
///
/// Note that this context (and all its clones) will be shut down once `Executor` is dropped.
pub fn new_execution_config(&self, executor_type: ExecutorType) -> IOxSessionConfig {
let exec = self.executor(executor_type).clone();
IOxSessionConfig::new(exec, Arc::clone(&self.runtime))
.with_target_partitions(self.config.target_query_partitions)
}
/// Get IOx context from DataFusion state.
pub fn new_context_from_df(
&self,
executor_type: ExecutorType,
state: &SessionState,
) -> IOxSessionContext {
let inner = SessionContext::with_state(state.clone());
let exec = self.executor(executor_type).clone();
let recorder = SpanRecorder::new(state.span_ctx().child_span("Query Execution"));
IOxSessionContext::new(inner, Some(exec), recorder)
}
/// Create a new execution context, suitable for executing a new query or system task
///
/// Note that this context (and all its clones) will be shut down once `Executor` is dropped.
pub fn new_context(&self, executor_type: ExecutorType) -> IOxSessionContext {
self.new_execution_config(executor_type).build()
}
/// Return the execution pool of the specified type
fn executor(&self, executor_type: ExecutorType) -> &DedicatedExecutor {
match executor_type {
ExecutorType::Query => &self.executors.query_exec,
ExecutorType::Reorg => &self.executors.reorg_exec,
}
}
/// Initializes shutdown.
pub fn shutdown(&self) {
self.executors.query_exec.shutdown();
self.executors.reorg_exec.shutdown();
}
/// Stops all subsequent task executions, and waits for the worker
/// thread to complete. Note this will shutdown all created contexts.
///
/// Only the first all to `join` will actually wait for the
/// executing thread to complete. All other calls to join will
/// complete immediately.
pub async fn join(&self) {
self.executors.query_exec.join().await;
self.executors.reorg_exec.join().await;
}
}
// No need to implement `Drop` because this is done by DedicatedExecutor already
/// Create a SchemaPivot node which an arbitrary input like
/// ColA | ColB | ColC
/// ------+------+------
/// 1 | NULL | NULL
/// 2 | 2 | NULL
/// 3 | 2 | NULL
///
/// And pivots it to a table with a single string column for any
/// columns that had non null values.
///
/// non_null_column
/// -----------------
/// "ColA"
/// "ColB"
pub fn make_schema_pivot(input: LogicalPlan) -> LogicalPlan {
let node = Arc::new(SchemaPivotNode::new(input));
LogicalPlan::Extension(Extension { node })
}
/// Make a NonNullChecker node takes an arbitrary input array and
/// produces a single string output column that contains
///
/// 1. the single `table_name` string if any of the input columns are non-null
/// 2. zero rows if all of the input columns are null
///
/// For this input:
///
/// ColA | ColB | ColC
/// ------+------+------
/// 1 | NULL | NULL
/// 2 | 2 | NULL
/// 3 | 2 | NULL
///
/// The output would be (given 'the_table_name' was the table name)
///
/// non_null_column
/// -----------------
/// the_table_name
///
/// However, for this input (All NULL)
///
/// ColA | ColB | ColC
/// ------+------+------
/// NULL | NULL | NULL
/// NULL | NULL | NULL
/// NULL | NULL | NULL
///
/// There would be no output rows
///
/// non_null_column
/// -----------------
pub fn make_non_null_checker(table_name: &str, input: LogicalPlan) -> LogicalPlan {
let node = Arc::new(NonNullCheckerNode::new(table_name, input));
LogicalPlan::Extension(Extension { node })
}
/// Create a StreamSplit node which takes an input stream of record
/// batches and produces multiple output streams based on a list of `N` predicates.
/// The output will have `N+1` streams, and each row is sent to the stream
/// corresponding to the first predicate that evaluates to true, or the last stream if none do.
///
/// For example, if the input looks like:
/// ```text
/// X | time
/// ---+-----
/// a | 1000
/// b | 4000
/// c | 2000
/// ```
///
/// A StreamSplit with split_exprs = [`time <= 1000`, `1000 < time <=2000`] will produce the
/// following three output streams (output DataFusion Partitions):
///
///
/// ```text
/// X | time
/// ---+-----
/// a | 1000
/// ```
///
/// ```text
/// X | time
/// ---+-----
/// b | 2000
/// ```
/// and
/// ```text
/// X | time
/// ---+-----
/// b | 4000
/// ```
pub fn make_stream_split(input: LogicalPlan, split_exprs: Vec<Expr>) -> LogicalPlan {
// rewrite the input expression so that it is fully qualified with the input schema
let split_exprs = split_exprs
.into_iter()
.map(|split_expr| normalize_col(split_expr, &input).expect("normalize is infallable"))
.collect::<Vec<_>>();
let node = Arc::new(StreamSplitNode::new(input, split_exprs));
LogicalPlan::Extension(Extension { node })
}
/// A type that can provide `IOxSessionContext` for query
pub trait ExecutionContextProvider {
/// Returns a new execution context suitable for running queries
fn new_query_context(&self, span_ctx: Option<trace::ctx::SpanContext>) -> IOxSessionContext;
}
#[cfg(test)]
mod tests {
use arrow::{
array::{ArrayRef, Int64Array, StringArray},
datatypes::{DataType, Field, Schema, SchemaRef},
};
use datafusion::{
datasource::{provider_as_source, MemTable},
logical_expr::LogicalPlanBuilder,
};
use stringset::StringSet;
use super::*;
use crate::exec::stringset::StringSetRef;
use crate::plan::stringset::StringSetPlan;
use arrow::record_batch::RecordBatch;
#[tokio::test]
async fn executor_known_string_set_plan_ok() {
let expected_strings = to_set(&["Foo", "Bar"]);
let plan = StringSetPlan::Known(Arc::clone(&expected_strings));
let exec = Executor::new(1);
let ctx = exec.new_context(ExecutorType::Query);
let result_strings = ctx.to_string_set(plan).await.unwrap();
assert_eq!(result_strings, expected_strings);
exec.join().await;
}
#[tokio::test]
async fn executor_datafusion_string_set_single_plan_no_batches() {
// Test with a single plan that produces no batches
let schema = Arc::new(Schema::new(vec![Field::new("a", DataType::Utf8, true)]));
let scan = make_plan(schema, vec![]);
let plan: StringSetPlan = vec![scan].into();
let exec = Executor::new(1);
let ctx = exec.new_context(ExecutorType::Query);
let results = ctx.to_string_set(plan).await.unwrap();
assert_eq!(results, StringSetRef::new(StringSet::new()));
exec.join().await;
}
#[tokio::test]
async fn executor_datafusion_string_set_single_plan_one_batch() {
// Test with a single plan that produces one record batch
let data = to_string_array(&["foo", "bar", "baz", "foo"]);
let batch = RecordBatch::try_from_iter_with_nullable(vec![("a", data, true)])
.expect("created new record batch");
let scan = make_plan(batch.schema(), vec![batch]);
let plan: StringSetPlan = vec![scan].into();
let exec = Executor::new(1);
let ctx = exec.new_context(ExecutorType::Query);
let results = ctx.to_string_set(plan).await.unwrap();
assert_eq!(results, to_set(&["foo", "bar", "baz"]));
exec.join().await;
}
#[tokio::test]
async fn executor_datafusion_string_set_single_plan_two_batch() {
// Test with a single plan that produces multiple record batches
let schema = Arc::new(Schema::new(vec![Field::new("a", DataType::Utf8, true)]));
let data1 = to_string_array(&["foo", "bar"]);
let batch1 = RecordBatch::try_new(Arc::clone(&schema), vec![data1])
.expect("created new record batch");
let data2 = to_string_array(&["baz", "foo"]);
let batch2 = RecordBatch::try_new(Arc::clone(&schema), vec![data2])
.expect("created new record batch");
let scan = make_plan(schema, vec![batch1, batch2]);
let plan: StringSetPlan = vec![scan].into();
let exec = Executor::new(1);
let ctx = exec.new_context(ExecutorType::Query);
let results = ctx.to_string_set(plan).await.unwrap();
assert_eq!(results, to_set(&["foo", "bar", "baz"]));
exec.join().await;
}
#[tokio::test]
async fn executor_datafusion_string_set_multi_plan() {
// Test with multiple datafusion logical plans
let schema = Arc::new(Schema::new(vec![Field::new("a", DataType::Utf8, true)]));
let data1 = to_string_array(&["foo", "bar"]);
let batch1 = RecordBatch::try_new(Arc::clone(&schema), vec![data1])
.expect("created new record batch");
let scan1 = make_plan(Arc::clone(&schema), vec![batch1]);
let data2 = to_string_array(&["baz", "foo"]);
let batch2 = RecordBatch::try_new(Arc::clone(&schema), vec![data2])
.expect("created new record batch");
let scan2 = make_plan(schema, vec![batch2]);
let plan: StringSetPlan = vec![scan1, scan2].into();
let exec = Executor::new(1);
let ctx = exec.new_context(ExecutorType::Query);
let results = ctx.to_string_set(plan).await.unwrap();
assert_eq!(results, to_set(&["foo", "bar", "baz"]));
exec.join().await;
}
#[tokio::test]
async fn executor_datafusion_string_set_nulls() {
// Ensure that nulls in the output set are handled reasonably
// (error, rather than silently ignored)
let schema = Arc::new(Schema::new(vec![Field::new("a", DataType::Utf8, true)]));
let array = StringArray::from_iter(vec![Some("foo"), None]);
let data = Arc::new(array);
let batch = RecordBatch::try_new(Arc::clone(&schema), vec![data])
.expect("created new record batch");
let scan = make_plan(schema, vec![batch]);
let plan: StringSetPlan = vec![scan].into();
let exec = Executor::new(1);
let ctx = exec.new_context(ExecutorType::Query);
let results = ctx.to_string_set(plan).await;
let actual_error = match results {
Ok(_) => "Unexpected Ok".into(),
Err(e) => format!("{}", e),
};
let expected_error = "unexpected null value";
assert!(
actual_error.contains(expected_error),
"expected error '{}' not found in '{:?}'",
expected_error,
actual_error,
);
exec.join().await;
}
#[tokio::test]
async fn executor_datafusion_string_set_bad_schema() {
// Ensure that an incorect schema (an int) gives a reasonable error
let data: ArrayRef = Arc::new(Int64Array::from(vec![1]));
let batch =
RecordBatch::try_from_iter(vec![("a", data)]).expect("created new record batch");
let scan = make_plan(batch.schema(), vec![batch]);
let plan: StringSetPlan = vec![scan].into();
let exec = Executor::new(1);
let ctx = exec.new_context(ExecutorType::Query);
let results = ctx.to_string_set(plan).await;
let actual_error = match results {
Ok(_) => "Unexpected Ok".into(),
Err(e) => format!("{}", e),
};
let expected_error = "schema not a single Utf8";
assert!(
actual_error.contains(expected_error),
"expected error '{}' not found in '{:?}'",
expected_error,
actual_error
);
exec.join().await;
}
#[tokio::test]
async fn make_schema_pivot_is_planned() {
// Test that all the planning logic is wired up and that we
// can make a plan using a SchemaPivot node
let batch = RecordBatch::try_from_iter_with_nullable(vec![
("f1", to_string_array(&["foo", "bar"]), true),
("f2", to_string_array(&["baz", "bzz"]), true),
])
.expect("created new record batch");
let scan = make_plan(batch.schema(), vec![batch]);
let pivot = make_schema_pivot(scan);
let plan = vec![pivot].into();
let exec = Executor::new(1);
let ctx = exec.new_context(ExecutorType::Query);
let results = ctx.to_string_set(plan).await.expect("Executed plan");
assert_eq!(results, to_set(&["f1", "f2"]));
exec.join().await;
}
/// return a set for testing
fn to_set(strs: &[&str]) -> StringSetRef {
StringSetRef::new(strs.iter().map(|s| s.to_string()).collect::<StringSet>())
}
fn to_string_array(strs: &[&str]) -> ArrayRef {
let array: StringArray = strs.iter().map(|s| Some(*s)).collect();
Arc::new(array)
}
// creates a DataFusion plan that reads the RecordBatches into memory
fn make_plan(schema: SchemaRef, data: Vec<RecordBatch>) -> LogicalPlan {
let partitions = vec![data];
let projection = None;
// model one partition,
let table = MemTable::try_new(schema, partitions).unwrap();
let source = provider_as_source(Arc::new(table));
LogicalPlanBuilder::scan("memtable", source, projection)
.unwrap()
.build()
.unwrap()
}
}