diff --git a/pkg/col/coldata/bytes.go b/pkg/col/coldata/bytes.go
index 2616ad60dbd4..da4488a72d5a 100644
--- a/pkg/col/coldata/bytes.go
+++ b/pkg/col/coldata/bytes.go
@@ -244,6 +244,14 @@ func (b *Bytes) copyElements(srcElementsToCopy []element, src *Bytes, destIdx in
 	// Optimize copying of the elements by copying all of them directly into the
 	// destination. This way all inlined values become correctly set, and we
 	// only need to set the non-inlined values separately.
+	//
+	// Note that this behavior results in losing the references to the old
+	// non-inlined values, even if they could be reused. If Bytes is not Reset,
+	// then that unused space in Bytes.buffer can accumulate. However, checking
+	// whether there are old non-inlined values with non-zero capacity leads to
+	// performance regressions, and in the production code we do reset the Bytes
+	// in all cases, so we accept this poor behavior in such a hypothetical /
+	// test-only scenario. See #78703 for more details.
 	copy(destElements, srcElementsToCopy)
 	// Early bounds checks.
 	_ = destElements[len(srcElementsToCopy)-1]
diff --git a/pkg/sql/colexecop/testutils.go b/pkg/sql/colexecop/testutils.go
index 886f3b8df646..e050c2611715 100644
--- a/pkg/sql/colexecop/testutils.go
+++ b/pkg/sql/colexecop/testutils.go
@@ -108,8 +108,9 @@ func (s *RepeatableBatchSource) Next() coldata.Batch {
 	if s.batchesToReturn != 0 && s.batchesReturned > s.batchesToReturn {
 		return coldata.ZeroBatch
 	}
-	s.output.SetSelection(s.sel != nil)
+	s.output.ResetInternalBatch()
 	if s.sel != nil {
+		s.output.SetSelection(true)
 		copy(s.output.Selection()[:s.batchLen], s.sel[:s.batchLen])
 	}
 	for i, colVec := range s.colVecs {