diff --git a/pkg/sql/logictest/testdata/logic_test/upsert b/pkg/sql/logictest/testdata/logic_test/upsert
index ed779a55183a..ba0084b11550 100644
--- a/pkg/sql/logictest/testdata/logic_test/upsert
+++ b/pkg/sql/logictest/testdata/logic_test/upsert
@@ -1213,3 +1213,24 @@ DROP TABLE source
 
 statement ok
 DROP TABLE target
+
+# Regression test for UPSERT batching logic (#51391).
+statement ok
+SET CLUSTER SETTING kv.raft.command.max_size='4MiB';
+CREATE TABLE src (s STRING);
+CREATE TABLE dest (s STRING);
+INSERT INTO src
+SELECT
+	'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa'
+FROM
+	generate_series(1, 50000)
+
+# This statement produces a raft command of about 6.6 MiB in size, so if the
+# batching logic is incorrect, we'll encounter "command is too large" error.
+statement ok
+UPSERT INTO dest (s) (SELECT s FROM src)
+
+statement ok
+RESET CLUSTER SETTING kv.raft.command.max_size;
+DROP TABLE src;
+DROP TABLE dest
diff --git a/pkg/sql/opt_exec_factory.go b/pkg/sql/opt_exec_factory.go
index 5da25178bf0b..cf6b82acbb23 100644
--- a/pkg/sql/opt_exec_factory.go
+++ b/pkg/sql/opt_exec_factory.go
@@ -1514,7 +1514,6 @@ func (ef *execFactory) ConstructUpsert(
 			insertCols: ri.InsertCols,
 			tw: optTableUpserter{
 				ri:            ri,
-				alloc:         &ef.planner.alloc,
 				canaryOrdinal: int(canaryCol),
 				fkTables:      fkTables,
 				fetchCols:     fetchColDescs,
diff --git a/pkg/sql/tablewriter_upsert_opt.go b/pkg/sql/tablewriter_upsert_opt.go
index a7ce5f786bc3..9b399ea250e9 100644
--- a/pkg/sql/tablewriter_upsert_opt.go
+++ b/pkg/sql/tablewriter_upsert_opt.go
@@ -45,8 +45,7 @@ import (
 type optTableUpserter struct {
 	tableWriterBase
 
-	ri    row.Inserter
-	alloc *sqlbase.DatumAlloc
+	ri row.Inserter
 
 	// Should we collect the rows for a RETURNING clause?
 	collectRows bool
@@ -67,16 +66,6 @@ type optTableUpserter struct {
 	// collectRows is set, counted separately otherwise.
 	resultCount int
 
-	// Contains all the rows to be inserted.
-	insertRows rowcontainer.RowContainer
-
-	// existingRows is used to store rows in a batch when checking for conflicts
-	// with rows earlier in the batch. Is is reused per batch.
-	existingRows *rowcontainer.RowContainer
-
-	// For allocation avoidance.
-	indexKeyPrefix []byte
-
 	// fetchCols indicate which columns need to be fetched from the target table,
 	// in order to detect whether a conflict has occurred, as well as to provide
 	// existing values for updates.
@@ -119,18 +108,15 @@ func (tu *optTableUpserter) init(
 	ctx context.Context, txn *kv.Txn, evalCtx *tree.EvalContext,
 ) error {
 	tu.tableWriterBase.init(txn)
-	tableDesc := tu.tableDesc()
 
-	tu.insertRows.Init(
-		evalCtx.Mon.MakeBoundAccount(), sqlbase.ColTypeInfoFromColDescs(tu.ri.InsertCols), 0,
-	)
-
-	// collectRows, set upon initialization, indicates whether or not we want rows returned from the operation.
+	// collectRows, set upon initialization, indicates whether or not we want
+	// rows returned from the operation.
 	if tu.collectRows {
+		tu.resultRow = make(tree.Datums, len(tu.returnCols))
 		tu.rowsUpserted = rowcontainer.NewRowContainer(
 			evalCtx.Mon.MakeBoundAccount(),
-			sqlbase.ColTypeInfoFromColDescs(tableDesc.Columns),
-			tu.insertRows.Len(),
+			sqlbase.ColTypeInfoFromColDescs(tu.returnCols),
+			0, /* rowCapacity */
 		)
 
 		// Create the map from colIds to the expected columns.
@@ -138,6 +124,7 @@ func (tu *optTableUpserter) init(
 		// because even though we might insert values into mutation columns, we
 		// never return them back to the user.
 		tu.colIDToReturnIndex = map[sqlbase.ColumnID]int{}
+		tableDesc := tu.tableDesc()
 		for i := range tableDesc.Columns {
 			id := tableDesc.Columns[i].ID
 			tu.colIDToReturnIndex[id] = i
@@ -156,33 +143,14 @@ func (tu *optTableUpserter) init(
 		}
 	}
 
-	tu.insertRows.Init(
-		evalCtx.Mon.MakeBoundAccount(), sqlbase.ColTypeInfoFromColDescs(tu.ri.InsertCols), 0,
-	)
-
-	tu.indexKeyPrefix = sqlbase.MakeIndexKeyPrefix(tableDesc.TableDesc(), tableDesc.PrimaryIndex.ID)
-
-	if tu.collectRows {
-		tu.resultRow = make(tree.Datums, len(tu.returnCols))
-		tu.rowsUpserted = rowcontainer.NewRowContainer(
-			evalCtx.Mon.MakeBoundAccount(),
-			sqlbase.ColTypeInfoFromColDescs(tu.returnCols),
-			tu.insertRows.Len(),
-		)
-	}
-
 	return nil
 }
 
 // flushAndStartNewBatch is part of the tableWriter interface.
 func (tu *optTableUpserter) flushAndStartNewBatch(ctx context.Context) error {
-	tu.insertRows.Clear(ctx)
 	if tu.collectRows {
 		tu.rowsUpserted.Clear(ctx)
 	}
-	if tu.existingRows != nil {
-		tu.existingRows.Clear(ctx)
-	}
 	return tu.tableWriterBase.flushAndStartNewBatch(ctx, tu.tableDesc())
 }
 
@@ -197,14 +165,8 @@ func (tu *optTableUpserter) batchedValues(rowIdx int) tree.Datums {
 	return tu.rowsUpserted.At(rowIdx)
 }
 
-func (tu *optTableUpserter) curBatchSize() int { return tu.insertRows.Len() }
-
 // close is part of the tableWriter interface.
 func (tu *optTableUpserter) close(ctx context.Context) {
-	tu.insertRows.Close(ctx)
-	if tu.existingRows != nil {
-		tu.existingRows.Close(ctx)
-	}
 	if tu.rowsUpserted != nil {
 		tu.rowsUpserted.Close(ctx)
 	}
diff --git a/pkg/sql/upsert.go b/pkg/sql/upsert.go
index 5deebf632102..09a11ea6d8cd 100644
--- a/pkg/sql/upsert.go
+++ b/pkg/sql/upsert.go
@@ -174,7 +174,7 @@ func (n *upsertNode) processSourceRow(params runParams, rowVals tree.Datums) err
 // BatchedCount implements the batchedPlanNode interface.
 func (n *upsertNode) BatchedCount() int { return n.run.tw.batchedCount() }
 
-// BatchedCount implements the batchedPlanNode interface.
+// BatchedValues implements the batchedPlanNode interface.
 func (n *upsertNode) BatchedValues(rowIdx int) tree.Datums { return n.run.tw.batchedValues(rowIdx) }
 
 func (n *upsertNode) Close(ctx context.Context) {