diff --git a/datafusion/physical-plan/src/coalesce_partitions.rs b/datafusion/physical-plan/src/coalesce_partitions.rs
index 5ea3589f22b3e..c887c29083617 100644
--- a/datafusion/physical-plan/src/coalesce_partitions.rs
+++ b/datafusion/physical-plan/src/coalesce_partitions.rs
@@ -19,13 +19,15 @@
 //! into a single partition
 
 use std::any::Any;
+use std::pin::Pin;
 use std::sync::Arc;
+use std::task::{Context, Poll};
 
 use super::metrics::{BaselineMetrics, ExecutionPlanMetricsSet, MetricsSet};
 use super::stream::{ObservedStream, RecordBatchReceiverStream};
 use super::{
-    DisplayAs, ExecutionPlanProperties, PlanProperties, SendableRecordBatchStream,
-    Statistics,
+    DisplayAs, ExecutionPlanProperties, PlanProperties, RecordBatchStream,
+    SendableRecordBatchStream, Statistics,
 };
 use crate::execution_plan::{CardinalityEffect, EvaluationType, SchedulingType};
 use crate::filter_pushdown::{FilterDescription, FilterPushdownPhase};
@@ -34,11 +36,14 @@ use crate::sort_pushdown::SortOrderPushdownResult;
 use crate::{DisplayFormatType, ExecutionPlan, Partitioning, check_if_same_properties};
 use datafusion_physical_expr_common::sort_expr::PhysicalSortExpr;
 
+use arrow::datatypes::SchemaRef;
 use datafusion_common::config::ConfigOptions;
 use datafusion_common::tree_node::TreeNodeRecursion;
 use datafusion_common::{Result, assert_eq_or_internal_err, internal_err};
 use datafusion_execution::TaskContext;
 use datafusion_physical_expr::PhysicalExpr;
+use futures::Stream;
+use futures::StreamExt;
 
 /// Merge execution plan executes partitions in parallel and combines them into a single
 /// partition. No guarantees are made about the order of the resulting partition.
@@ -209,33 +214,14 @@ impl ExecutionPlan for CoalescePartitionsExec {
             }
             _ => {
                 let baseline_metrics = BaselineMetrics::new(&self.metrics, partition);
-                // record the (very) minimal work done so that
-                // elapsed_compute is not reported as 0
-                let elapsed_compute = baseline_metrics.elapsed_compute().clone();
-                let _timer = elapsed_compute.timer();
-
-                // use a stream that allows each sender to put in at
-                // least one result in an attempt to maximize
-                // parallelism.
-                let mut builder =
-                    RecordBatchReceiverStream::builder(self.schema(), input_partitions);
-
-                // spawn independent tasks whose resulting streams (of batches)
-                // are sent to the channel for consumption.
-                for part_i in 0..input_partitions {
-                    builder.run_input(
-                        Arc::clone(&self.input),
-                        part_i,
-                        Arc::clone(&context),
-                    );
-                }
-
-                let stream = builder.build();
-                Ok(Box::pin(ObservedStream::new(
-                    stream,
+                Ok(Box::pin(CoalescePartitionsStream {
+                    schema: self.schema(),
+                    input: Arc::clone(&self.input),
+                    context,
                     baseline_metrics,
-                    self.fetch,
-                )))
+                    fetch: self.fetch,
+                    state: CoalescePartitionsStreamState::Pending,
+                }))
             }
         }
     }
@@ -352,6 +338,81 @@ impl ExecutionPlan for CoalescePartitionsExec {
     }
 }
 
+/// A stream that lazily spawns input partition tasks on first poll,
+/// rather than eagerly in `execute()`.
+struct CoalescePartitionsStream {
+    schema: SchemaRef,
+    input: Arc<dyn ExecutionPlan>,
+    context: Arc<TaskContext>,
+    baseline_metrics: BaselineMetrics,
+    fetch: Option<usize>,
+    state: CoalescePartitionsStreamState,
+}
+
+enum CoalescePartitionsStreamState {
+    /// Tasks have not been spawned yet.
+    Pending,
+    /// Tasks have been spawned, polling the merged stream.
+    Running(Pin<Box<ObservedStream>>),
+}
+
+impl CoalescePartitionsStream {
+    fn start(&mut self) -> &mut Pin<Box<ObservedStream>> {
+        let input_partitions = self.input.output_partitioning().partition_count();
+
+        // record the (very) minimal work done so that
+        // elapsed_compute is not reported as 0
+        let elapsed_compute = self.baseline_metrics.elapsed_compute().clone();
+        let _timer = elapsed_compute.timer();
+
+        // use a stream that allows each sender to put in at
+        // least one result in an attempt to maximize
+        // parallelism.
+        let mut builder =
+            RecordBatchReceiverStream::builder(self.schema.clone(), input_partitions);
+
+        // spawn independent tasks whose resulting streams (of batches)
+        // are sent to the channel for consumption.
+        for part_i in 0..input_partitions {
+            builder.run_input(
+                Arc::clone(&self.input),
+                part_i,
+                Arc::clone(&self.context),
+            );
+        }
+
+        let stream = builder.build();
+        self.state = CoalescePartitionsStreamState::Running(Box::pin(
+            ObservedStream::new(stream, self.baseline_metrics.clone(), self.fetch),
+        ));
+        match &mut self.state {
+            CoalescePartitionsStreamState::Running(s) => s,
+            _ => unreachable!(),
+        }
+    }
+}
+
+impl RecordBatchStream for CoalescePartitionsStream {
+    fn schema(&self) -> SchemaRef {
+        self.schema.clone()
+    }
+}
+
+impl Stream for CoalescePartitionsStream {
+    type Item = Result<arrow::array::RecordBatch>;
+
+    fn poll_next(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+    ) -> Poll<Option<Self::Item>> {
+        let stream = match &mut self.state {
+            CoalescePartitionsStreamState::Running(s) => s,
+            CoalescePartitionsStreamState::Pending => self.start(),
+        };
+        stream.poll_next_unpin(cx)
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;