Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Don't optimize AnalyzeExec (#6379) (try 2) #6494

Merged
merged 6 commits into from
May 31, 2023
Merged
Show file tree
Hide file tree
Changes from 2 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion datafusion/core/src/physical_optimizer/pipeline_checker.rs
Original file line number Diff line number Diff line change
Expand Up @@ -387,7 +387,7 @@ mod sql_tests {
};
let test2 = UnaryTestCase {
source_type: SourceType::Unbounded,
expect_fail: true,
expect_fail: false,
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

it works now!

};
let case = QueryCase {
sql: "EXPLAIN ANALYZE SELECT * FROM test".to_string(),
Expand Down
158 changes: 83 additions & 75 deletions datafusion/core/src/physical_plan/analyze.rs
Original file line number Diff line number Diff line change
Expand Up @@ -28,10 +28,12 @@ use crate::{
},
};
use arrow::{array::StringBuilder, datatypes::SchemaRef, record_batch::RecordBatch};
use futures::StreamExt;
use futures::{FutureExt, StreamExt, TryFutureExt};
use tokio::task::JoinSet;

use super::expressions::PhysicalSortExpr;
use super::{stream::RecordBatchReceiverStream, Distribution, SendableRecordBatchStream};
use super::stream::RecordBatchStreamAdapter;
use super::{Distribution, SendableRecordBatchStream};
use crate::execution::context::TaskContext;

/// `EXPLAIN ANALYZE` execution plan operator. This operator runs its input,
Expand Down Expand Up @@ -73,7 +75,7 @@ impl ExecutionPlan for AnalyzeExec {

/// Specifies we want the input as a single stream
fn required_input_distribution(&self) -> Vec<Distribution> {
vec![Distribution::SinglePartition]
vec![Distribution::UnspecifiedDistribution]
}

/// Specifies whether this plan generates an infinite stream of records.
Expand Down Expand Up @@ -121,96 +123,102 @@ impl ExecutionPlan for AnalyzeExec {
)));
}

// should be ensured by `SinglePartition` above
let input_partitions = self.input.output_partitioning().partition_count();
if input_partitions != 1 {
return Err(DataFusionError::Internal(format!(
"AnalyzeExec invalid number of input partitions. Expected 1, got {input_partitions}"
)));
// Gather futures that will run each input partition using a
// JoinSet to cancel outstanding futures on drop
let mut set = JoinSet::new();
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This uses the cool JoinSet I learned about from @nvartolomei and @Darksonn on #6449 ❤️

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this logic was just extracted into its own function

let num_input_partitions = self.input.output_partitioning().partition_count();

for input_partition in 0..num_input_partitions {
let input_stream = self.input.execute(input_partition, context.clone());

set.spawn(async move {
let mut total_rows = 0;
let mut input_stream = input_stream?;
while let Some(batch) = input_stream.next().await {
let batch = batch?;
total_rows += batch.num_rows();
}
Ok(total_rows)
});
}

let (tx, rx) = tokio::sync::mpsc::channel(input_partitions);
// Turn the tasks in the JoinSet into a stream of
// Result<usize> representing the counts of each output
// partition.
let counts_stream = futures::stream::unfold(set, |mut set| async {
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think looking at https://github.com/apache/arrow-datafusion/pull/6494/files?w=1 makes it clearer what I did -- which was to change the plumbing to use futures and stream fu rather than channels

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I confirmed that the tokio stream adapaters don't appear to have a JoinSet impl - https://docs.rs/tokio-stream/latest/tokio_stream/wrappers/index.html?search=

let next = set.join_next().await?; // returns Some when empty
// translate join errors (aka task panic's) into ExecutionErrors
let next = match next {
Ok(res) => res,
Err(e) => Err(DataFusionError::Execution(format!(
"Join error in AnalyzeExec: {e}"
))),
};
Some((next, set))
});

let start = Instant::now();
let captured_input = self.input.clone();
let mut input_stream = captured_input.execute(0, context)?;
let captured_schema = self.schema.clone();
let verbose = self.verbose;

// Task reads batches the input and when complete produce a
// RecordBatch with a report that is written to `tx` when done
let join_handle = tokio::task::spawn(async move {
let start = Instant::now();
let mut total_rows = 0;

// Note the code below ignores errors sending on tx. An
// error sending means the plan is being torn down and
// nothing is left that will handle the error (aka no one
// will hear us scream)
while let Some(b) = input_stream.next().await {
match b {
Ok(batch) => {
total_rows += batch.num_rows();
}
b @ Err(_) => {
// try and pass on errors from input
if tx.send(b).await.is_err() {
// receiver hung up, stop executing (no
// one will look at any further results we
// send)
return;
}
}
}
}
let end = Instant::now();

let mut type_builder = StringBuilder::with_capacity(1, 1024);
let mut plan_builder = StringBuilder::with_capacity(1, 1024);

// TODO use some sort of enum rather than strings?
type_builder.append_value("Plan with Metrics");
// future that gathers the input counts into an overall output
// count, and makes an output batch
let output = counts_stream
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

FWIW you could just use a regular async move here, instead of needing the futures adapters

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That is an excellent point -- I got so carried away with being clever I lost sight of that.

I rewrote this logic to use a single future and also moved the output record batch creation into a function to separate the business logic from the async orchestration: 5521a70

// combine results from all input stream into a total count
.fold(Ok(0), |total_rows: Result<usize>, input_rows| async move {
Ok(total_rows? + input_rows?)
})
// convert the total to a RecordBatch
.map(move |total_rows| {
let total_rows = total_rows?;
let end = Instant::now();

let annotated_plan =
DisplayableExecutionPlan::with_metrics(captured_input.as_ref())
.indent()
.to_string();
plan_builder.append_value(annotated_plan);
let mut type_builder = StringBuilder::with_capacity(1, 1024);
let mut plan_builder = StringBuilder::with_capacity(1, 1024);

// Verbose output
// TODO make this more sophisticated
if verbose {
type_builder.append_value("Plan with Full Metrics");
// TODO use some sort of enum rather than strings?
type_builder.append_value("Plan with Metrics");

let annotated_plan =
DisplayableExecutionPlan::with_full_metrics(captured_input.as_ref())
DisplayableExecutionPlan::with_metrics(captured_input.as_ref())
.indent()
.to_string();
plan_builder.append_value(annotated_plan);

type_builder.append_value("Output Rows");
plan_builder.append_value(total_rows.to_string());
// Verbose output
// TODO make this more sophisticated
if verbose {
type_builder.append_value("Plan with Full Metrics");

type_builder.append_value("Duration");
plan_builder.append_value(format!("{:?}", end - start));
}
let annotated_plan = DisplayableExecutionPlan::with_full_metrics(
captured_input.as_ref(),
)
.indent()
.to_string();
plan_builder.append_value(annotated_plan);

let maybe_batch = RecordBatch::try_new(
captured_schema,
vec![
Arc::new(type_builder.finish()),
Arc::new(plan_builder.finish()),
],
)
.map_err(Into::into);
// again ignore error
tx.send(maybe_batch).await.ok();
});
type_builder.append_value("Output Rows");
plan_builder.append_value(total_rows.to_string());

type_builder.append_value("Duration");
plan_builder.append_value(format!("{:?}", end - start));
}

Ok(RecordBatchReceiverStream::create(
&self.schema,
rx,
join_handle,
))
RecordBatch::try_new(
captured_schema,
vec![
Arc::new(type_builder.finish()),
Arc::new(plan_builder.finish()),
],
)
})
.map_err(DataFusionError::from);

Ok(Box::pin(RecordBatchStreamAdapter::new(
self.schema.clone(),
futures::stream::once(output),
)))
}

fn fmt_as(
Expand Down
12 changes: 7 additions & 5 deletions datafusion/core/src/physical_plan/planner.rs
Original file line number Diff line number Diff line change
Expand Up @@ -1192,11 +1192,9 @@ impl DefaultPhysicalPlanner {
"Unsupported logical plan: Distinct should be replaced to Aggregate".to_string(),
))
}
LogicalPlan::Analyze(a) => {
let input = self.create_initial_plan(&a.input, session_state).await?;
let schema = SchemaRef::new((*a.schema).clone().into());
Ok(Arc::new(AnalyzeExec::new(a.verbose, input, schema)))
}
LogicalPlan::Analyze(_) => Err(DataFusionError::Internal(
"Unsupported logical plan: Analyze must be root of the plan".to_string(),
)),
LogicalPlan::Extension(e) => {
let physical_inputs = self.create_initial_plan_multi(e.node.inputs(), session_state).await?;

Expand Down Expand Up @@ -1851,6 +1849,10 @@ impl DefaultPhysicalPlanner {
stringified_plans,
e.verbose,
))))
} else if let LogicalPlan::Analyze(a) = logical_plan {
let input = self.create_physical_plan(&a.input, session_state).await?;
let schema = SchemaRef::new((*a.schema).clone().into());
Ok(Some(Arc::new(AnalyzeExec::new(a.verbose, input, schema))))
} else {
Ok(None)
}
Expand Down
20 changes: 19 additions & 1 deletion datafusion/core/tests/sql/explain_analyze.rs
Original file line number Diff line number Diff line change
Expand Up @@ -687,13 +687,31 @@ async fn csv_explain_analyze() {
// Only test basic plumbing and try to avoid having to change too
// many things. explain_analyze_baseline_metrics covers the values
// in greater depth
let needle = "CoalescePartitionsExec, metrics=[output_rows=5, elapsed_compute=";
let needle = "AggregateExec: mode=FinalPartitioned, gby=[c1@0 as c1], aggr=[COUNT(UInt8(1))], metrics=[output_rows=5";
assert_contains!(&formatted, needle);

let verbose_needle = "Output Rows";
assert_not_contains!(formatted, verbose_needle);
}

#[tokio::test]
#[cfg_attr(tarpaulin, ignore)]
async fn csv_explain_analyze_order_by() {
let ctx = SessionContext::new();
register_aggregate_csv_by_sql(&ctx).await;
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

here is the new test

let sql = "EXPLAIN ANALYZE SELECT c1 FROM aggregate_test_100 order by c1";
let actual = execute_to_batches(&ctx, sql).await;
let formatted = arrow::util::pretty::pretty_format_batches(&actual)
.unwrap()
.to_string();

// Ensure that the ordering is not optimized away from the plan
// https://github.com/apache/arrow-datafusion/issues/6379
let needle =
"SortExec: expr=[c1@0 ASC NULLS LAST], metrics=[output_rows=100, elapsed_compute";
assert_contains!(&formatted, needle);
}

#[tokio::test]
#[cfg_attr(tarpaulin, ignore)]
async fn parquet_explain_analyze() {
Expand Down