plan: Choose left table as build table for semi join when left one is smaller for MPPTask

planner/core/casetest/physical_plan_test.go

@@ -941,6 +941,62 @@ func TestMPPBCJModelOneTiFlash(t *testing.T) {
 	}
 }
 
+func TestMPPRightSemiJoin(t *testing.T) {
+	store := testkit.CreateMockStore(t, internal.WithMockTiFlash(3))
+	tk := testkit.NewTestKit(t, store)
+	tk.MustExec("use test")
+	tk.MustExec("drop table if exists t1")
+	tk.MustExec("create table t1 (a int)")
+	tk.MustExec("drop table if exists t2")
+	tk.MustExec("create table t2 (b int)")
+
+	tk.MustExec("insert into t1 values (1);")
+	tk.MustExec("insert into t2 values (1), (2), (3), (4), (5), (6), (7), (8);")
+
+	{
+		tk.MustExec("alter table t1 set tiflash replica 1")
+		tb := external.GetTableByName(t, tk, "test", "t1")
+		err := domain.GetDomain(tk.Session()).DDL().UpdateTableReplicaInfo(tk.Session(), tb.Meta().ID, true)
+		require.NoError(t, err)
+	}
+	{
+		tk.MustExec("alter table t2 set tiflash replica 1")
+		tb := external.GetTableByName(t, tk, "test", "t2")
+		err := domain.GetDomain(tk.Session()).DDL().UpdateTableReplicaInfo(tk.Session(), tb.Meta().ID, true)
+		require.NoError(t, err)
+	}
+	tk.MustExec("analyze table t1")
+	tk.MustExec("analyze table t2")
+	tk.MustExec("set @@tidb_allow_mpp=1; set @@tidb_enforce_mpp=1;")
+	{
+		var input []string
+		var output []struct {
+			SQL  string
+			Plan []string
+			Warn []string
+		}
+		planSuiteData := GetPlanSuiteData()
+		planSuiteData.LoadTestCases(t, &input, &output)
+		for i, tt := range input {
+			testdata.OnRecord(func() {
+				output[i].SQL = tt
+			})
+			if strings.HasPrefix(tt, "set") || strings.HasPrefix(tt, "insert") {
+				tk.MustExec(tt)
+				continue
+			}
+			testdata.OnRecord(func() {
+				output[i].SQL = tt
+				output[i].Plan = testdata.ConvertRowsToStrings(tk.MustQuery(tt).Rows())
+				output[i].Warn = testdata.ConvertSQLWarnToStrings(tk.Session().GetSessionVars().StmtCtx.GetWarnings())
+			})
+			res := tk.MustQuery(tt)
+			res.Check(testkit.Rows(output[i].Plan...))
+			require.Equal(t, output[i].Warn, testdata.ConvertSQLWarnToStrings(tk.Session().GetSessionVars().StmtCtx.GetWarnings()))
+		}
+	}
+}
+
 func TestHintScope(t *testing.T) {
 	store := testkit.CreateMockStore(t)
 	tk := testkit.NewTestKit(t, store)

planner/core/casetest/testdata/plan_suite_in.json

@@ -155,6 +155,13 @@
       "explain select * from t t1, t t2 where t1.a=t2.a"
     ]
   },
+  {
+    "name": "TestMPPRightSemiJoin",
+    "cases": [
+      "set @@session.tidb_allow_mpp=true",
+      "explain select * from t1 where exists (select * from t2 where t1.a=t2.b)"
+    ]
+  },
   {
     "name": "TestIssue37520",
     "cases": [

planner/core/casetest/testdata/plan_suite_out.json

@@ -1564,6 +1564,31 @@
       }
     ]
   },
+  {
+    "Name": "TestMPPRightSemiJoin",
+    "Cases": [
+      {
+        "SQL": "set @@session.tidb_allow_mpp=true",
+        "Plan": null,
+        "Warn": null
+      },
+      {
+        "SQL": "explain select * from t1 where exists (select * from t2 where t1.a=t2.b)",
+        "Plan": [
+          "TableReader_34 0.80 root  MppVersion: 1, data:ExchangeSender_33",
+          "└─ExchangeSender_33 0.80 mpp[tiflash]  ExchangeType: PassThrough",
+          "  └─HashJoin_32 0.80 mpp[tiflash]  semi join, equal:[eq(test.t1.a, test.t2.b)]",
+          "    ├─ExchangeReceiver_15(Build) 1.00 mpp[tiflash]  ",
+          "    │ └─ExchangeSender_14 1.00 mpp[tiflash]  ExchangeType: Broadcast, Compression: FAST",
+          "    │   └─Selection_13 1.00 mpp[tiflash]  not(isnull(test.t1.a))",
+          "    │     └─TableFullScan_12 1.00 mpp[tiflash] table:t1 pushed down filter:empty, keep order:false",
+          "    └─Selection_17(Probe) 8.00 mpp[tiflash]  not(isnull(test.t2.b))",
+          "      └─TableFullScan_16 8.00 mpp[tiflash] table:t2 pushed down filter:empty, keep order:false"
+        ],
+        "Warn": null
+      }
+    ]
+  },
   {
     "Name": "TestIssue37520",
     "Cases": [

planner/core/exhaust_physical_plans.go

@@ -2386,8 +2386,18 @@ func (p *LogicalJoin) tryToGetMppHashJoin(prop *property.PhysicalProperty, useBC
 			preferredBuildIndex = 1
 		}
 	} else if p.JoinType.IsSemiJoin() {
-		preferredBuildIndex = 1
-		fixedBuildSide = true
+		if !p.isNAAJ() && len(p.EqualConditions) > 0 && (p.JoinType == SemiJoin || p.JoinType == AntiSemiJoin) {
+			// TiFlash only supports Non-null_aware non-cross semi/anti_semi join to use both sides as build side
+			preferredBuildIndex = 1
+			// MPPOuterJoinFixedBuildSide default value is false
+			// use MPPOuterJoinFixedBuildSide here as a way to disable using left table as build side
+			if !p.ctx.GetSessionVars().MPPOuterJoinFixedBuildSide && p.children[1].statsInfo().Count() > p.children[0].statsInfo().Count() {
+				preferredBuildIndex = 0
+			}
+		} else {
+			preferredBuildIndex = 1
+			fixedBuildSide = true
+		}
 	}
 	if p.JoinType == LeftOuterJoin || p.JoinType == RightOuterJoin {
 		// TiFlash does not require that the build side must be the inner table for outer join.