Minor: Improve documentation on execution error handling

datafusion/execution/src/stream.rs

@@ -20,7 +20,9 @@ use datafusion_common::Result;
 use futures::Stream;
 use std::pin::Pin;
 
-/// Trait for types that stream [arrow::record_batch::RecordBatch]
+/// Trait for types that stream [RecordBatch]
+///
+/// See [`SendableRecordBatchStream`] for more details.
 pub trait RecordBatchStream: Stream<Item = Result<RecordBatch>> {
     /// Returns the schema of this `RecordBatchStream`.
     ///
@@ -29,5 +31,23 @@ pub trait RecordBatchStream: Stream<Item = Result<RecordBatch>> {
     fn schema(&self) -> SchemaRef;
 }
 
-/// Trait for a [`Stream`] of [`RecordBatch`]es
+/// Trait for a [`Stream`] of [`RecordBatch`]es that can be passed between threads
+///
+/// This trait is used to retrieve the results of DataFusion execution plans.
+///
+/// The trait is a specialized Rust Async [`Stream`] that also knows the schema
+/// of the data it will return (even if the stream has no data). Every
+/// `RecordBatch` returned by the stream should have the same schema as returned
+/// by [`schema`](`RecordBatchStream::schema`).
+///
+/// # Error Handling
+///
+/// One a stream returns an error, it should not be polled again (the caller
+/// should stop calling `next`) and handle the error.
+///
+/// However, returning `Ready(None)` (end of stream) is likely the safest
+/// behavior after an error. Like [`Stream`]s, `RecordBatchStream`s should not
+/// be polled after end of stream or returning an error. However, also like
+/// [`Stream`]s there is no mechanism to prevent callers polling  so returning
+/// `Ready(None)` is recommended.
 pub type SendableRecordBatchStream = Pin<Box<dyn RecordBatchStream + Send>>;

datafusion/physical-plan/src/execution_plan.rs

@@ -228,6 +228,16 @@ pub trait ExecutionPlan: Debug + DisplayAs + Send + Sync {
     /// [`TryStreamExt`]: futures::stream::TryStreamExt
     /// [`RecordBatchStreamAdapter`]: crate::stream::RecordBatchStreamAdapter
     ///
+    /// # Error handling
+    ///
+    /// Any error that occurs during execution is sent as an `Err` in the output
+    /// stream.
+    ///
+    /// `ExecutionPlan` implementations in DataFusion cancel additional work
+    /// immediately once an error occurs. The rationale is that if the overall
+    /// query will return an error,  any additional work such as continued
+    /// polling of inputs will be wasted as it will be thrown away.
+    ///
     /// # Cancellation / Aborting Execution
     ///
     /// The [`Stream`] that is returned must ensure that any allocated resources

datafusion/physical-plan/src/repartition/mod.rs

@@ -385,6 +385,11 @@ impl BatchPartitioner {
 ///         `───────'                   `───────'
 ///```
 ///
+/// # Error Handling
+///
+/// If any of the input partitions return an error, the error is propagated to
+/// all output partitions and inputs are not polled again.
+///
 /// # Output Ordering
 ///
 /// If more than one stream is being repartitioned, the output will be some

datafusion/physical-plan/src/sorts/merge.rs

@@ -39,6 +39,7 @@ use futures::Stream;
 /// A fallible [`PartitionedStream`] of [`Cursor`] and [`RecordBatch`]
 type CursorStream<C> = Box<dyn PartitionedStream<Output = Result<(C, RecordBatch)>>>;
 
+/// Merges a stream of sorted cursors and record batches into a single sorted stream
 #[derive(Debug)]
 pub(crate) struct SortPreservingMergeStream<C: CursorValues> {
     in_progress: BatchBuilder,

datafusion/physical-plan/src/sorts/sort_preserving_merge.rs

@@ -65,6 +65,11 @@ use log::{debug, trace};
 ///  Input Streams                                             Output stream
 ///    (sorted)                                                  (sorted)
 /// ```
+///
+/// # Error Handling
+///
+/// If any of the input partitions return an error, the error is propagated to
+/// the output and inputs are not polled again.
 #[derive(Debug)]
 pub struct SortPreservingMergeExec {
     /// Input plan