*: use clustered PRIMARY KEY for stats system tables (#68324)

close #66751

Author: 0xPoe

pkg/executor/hot_regions_history_table_test.go

@@ -187,9 +187,6 @@ func TestTiDBHotRegionsHistory(t *testing.T) {
 		// mysql table_id = 21 ,index_id = 1, table_name = STATS_META, index_name = IDX_VER
 		{"2019-10-10 10:10:19", "MYSQL", "STATS_META", statsMetaTidStr, "IDX_VER", "1", "3", "3", "33333", "0", "1", "READ", "99", "99", "99", "99"},
 		{"2019-10-10 10:10:20", "MYSQL", "STATS_META", statsMetaTidStr, "IDX_VER", "1", "4", "4", "44444", "0", "0", "WRITE", "99", "99", "99", "99"},
-		// mysql table_id = 21 ,index_id = 2, table_name = STATS_META, index_name = TBL
-		{"2019-10-10 10:10:21", "MYSQL", "STATS_META", statsMetaTidStr, "TBL", "2", "5", "5", "55555", "0", "1", "READ", "99", "99", "99", "99"},
-		{"2019-10-10 10:10:22", "MYSQL", "STATS_META", statsMetaTidStr, "TBL", "2", "6", "6", "66666", "0", "0", "WRITE", "99", "99", "99", "99"},
 		// table_id = 1313, index_id = 1, deleted schema
 		{"2019-10-10 10:10:23", "UNKNOWN", "UNKNOWN", "1313", "UNKNOWN", "1", "7", "7", "77777", "0", "1", "READ", "99", "99", "99", "99"},
 		{"2019-10-10 10:10:24", "UNKNOWN", "UNKNOWN", "1313", "UNKNOWN", "1", "8", "8", "88888", "0", "0", "WRITE", "99", "99", "99", "99"},
@@ -275,12 +272,6 @@ func TestTiDBHotRegionsHistory(t *testing.T) {
 						StartKey: helper.NewIndexWithKeyRange(mockDB, &model.TableInfo{ID: statsMetaTid}, &model.IndexInfo{ID: 1}, storeCodec).StartKey,
 						EndKey:   helper.NewIndexWithKeyRange(mockDB, &model.TableInfo{ID: statsMetaTid}, &model.IndexInfo{ID: 1}, storeCodec).EndKey,
 					},
-					// mysql table_id = 21 ,index_id = 2, table_name = STATS_META, index_name = TBL
-					{UpdateTime: unixTimeMs("2019-10-10 10:10:21"), RegionID: 5, StoreID: 5, PeerID: 55555, IsLearner: false,
-						IsLeader: true, HotRegionType: "READ", HotDegree: 99, FlowBytes: 99, KeyRate: 99, QueryRate: 99,
-						StartKey: helper.NewIndexWithKeyRange(mockDB, &model.TableInfo{ID: statsMetaTid}, &model.IndexInfo{ID: 2}, storeCodec).StartKey,
-						EndKey:   helper.NewIndexWithKeyRange(mockDB, &model.TableInfo{ID: statsMetaTid}, &model.IndexInfo{ID: 2}, storeCodec).EndKey,
-					},
 					//      table_id = 1313, index_id = 1, deleted schema
 					{UpdateTime: unixTimeMs("2019-10-10 10:10:23"), RegionID: 7, StoreID: 7, PeerID: 77777, IsLeader: true,
 						HotRegionType: "READ", HotDegree: 99, FlowBytes: 99, KeyRate: 99, QueryRate: 99,
@@ -297,12 +288,6 @@ func TestTiDBHotRegionsHistory(t *testing.T) {
 						StartKey: helper.NewIndexWithKeyRange(mockDB, &model.TableInfo{ID: statsMetaTid}, &model.IndexInfo{ID: 1}, storeCodec).StartKey,
 						EndKey:   helper.NewIndexWithKeyRange(mockDB, &model.TableInfo{ID: statsMetaTid}, &model.IndexInfo{ID: 1}, storeCodec).EndKey,
 					},
-					// mysql table_id = 21 ,index_id = 2, table_name = STATS_META, index_name = TBL
-					{UpdateTime: unixTimeMs("2019-10-10 10:10:22"), RegionID: 6, StoreID: 6, PeerID: 66666, IsLearner: false,
-						IsLeader: false, HotRegionType: "WRITE", HotDegree: 99, FlowBytes: 99, KeyRate: 99, QueryRate: 99,
-						StartKey: helper.NewIndexWithKeyRange(mockDB, &model.TableInfo{ID: statsMetaTid}, &model.IndexInfo{ID: 2}, storeCodec).StartKey,
-						EndKey:   helper.NewIndexWithKeyRange(mockDB, &model.TableInfo{ID: statsMetaTid}, &model.IndexInfo{ID: 2}, storeCodec).EndKey,
-					},
 					//      table_id = 1313, index_id = 1, deleted schema
 					{UpdateTime: unixTimeMs("2019-10-10 10:10:24"), RegionID: 8, StoreID: 8, PeerID: 88888, IsLearner: false,
 						IsLeader: false, HotRegionType: "WRITE", HotDegree: 99, FlowBytes: 99, KeyRate: 99, QueryRate: 99,
@@ -328,7 +313,7 @@ func TestTiDBHotRegionsHistory(t *testing.T) {
 				fullHotRegions[0], fullHotRegions[1], fullHotRegions[2],
 				fullHotRegions[3],
 				fullHotRegions[6], fullHotRegions[7], fullHotRegions[8],
-				fullHotRegions[9], fullHotRegions[10], fullHotRegions[11],
+				fullHotRegions[9],
 			},
 		},
 		{
@@ -340,7 +325,7 @@ func TestTiDBHotRegionsHistory(t *testing.T) {
 				fullHotRegions[0], fullHotRegions[1], fullHotRegions[2],
 				fullHotRegions[3],
 				fullHotRegions[6], fullHotRegions[7], fullHotRegions[8],
-				fullHotRegions[9], fullHotRegions[10], fullHotRegions[11],
+				fullHotRegions[9],
 			},
 		},
 		{

pkg/meta/metadef/system_tables_def.go

@@ -173,8 +173,8 @@ const (
 		count 						BIGINT(64) UNSIGNED NOT NULL DEFAULT 0,
 		snapshot        			BIGINT(64) UNSIGNED NOT NULL DEFAULT 0,
 		last_stats_histograms_version 	BIGINT(64) UNSIGNED DEFAULT NULL,
-		INDEX idx_ver(version),
-		UNIQUE INDEX tbl(table_id)
+		PRIMARY KEY (table_id) CLUSTERED,
+		INDEX idx_ver(version)
 	);`
 
 	// CreateStatsHistogramsTable stores the statistics of table columns.
@@ -192,7 +192,7 @@ const (
 		flag 				BIGINT(64) NOT NULL DEFAULT 0,
 		correlation 		DOUBLE NOT NULL DEFAULT 0,
 		last_analyze_pos 	LONGBLOB DEFAULT NULL,
-		UNIQUE INDEX tbl(table_id, is_index, hist_id)
+		PRIMARY KEY (table_id, is_index, hist_id) CLUSTERED
 	);`
 
 	// CreateStatsBucketsTable stores the histogram info for every table columns.
@@ -206,7 +206,7 @@ const (
 		upper_bound LONGBLOB NOT NULL,
 		lower_bound LONGBLOB ,
 		ndv         BIGINT NOT NULL DEFAULT 0,
-		UNIQUE INDEX tbl(table_id, is_index, hist_id, bucket_id)
+		PRIMARY KEY (table_id, is_index, hist_id, bucket_id) CLUSTERED
 	);`
 
 	// CreateGCDeleteRangeTable stores schemas which can be deleted by DeleteRange.
@@ -293,7 +293,7 @@ const (
 		is_index 	TINYINT(2) NOT NULL,
 		hist_id 	BIGINT(64) NOT NULL,
 		value 		LONGBLOB,
-		INDEX tbl(table_id, is_index, hist_id)
+		PRIMARY KEY (table_id, is_index, hist_id) CLUSTERED
 	);`
 
 	// CreateExprPushdownBlacklistTable stores the expressions which are not allowed to be pushed down.
@@ -382,7 +382,7 @@ const (
 		seq_no bigint(64) NOT NULL comment 'sequence number of the gzipped data slice',
 		version bigint(64) NOT NULL comment 'stats version which corresponding to stats:version in EXPLAIN',
 		create_time datetime(6) NOT NULL,
-		UNIQUE KEY table_version_seq (table_id, version, seq_no),
+		PRIMARY KEY (table_id, version, seq_no) CLUSTERED,
 		KEY table_create_time (table_id, create_time, seq_no),
     	KEY idx_create_time (create_time)
 	);`
@@ -394,7 +394,7 @@ const (
 		version bigint(64) NOT NULL comment 'stats version which corresponding to stats:version in EXPLAIN',
     	source varchar(40) NOT NULL,
 		create_time datetime(6) NOT NULL,
-		UNIQUE KEY table_version (table_id, version),
+		PRIMARY KEY (table_id, version) CLUSTERED,
 		KEY table_create_time (table_id, create_time),
     	KEY idx_create_time (create_time)
 	);`
@@ -461,7 +461,7 @@ const (
 		modify_count bigint(64) NOT NULL DEFAULT 0,
 		count bigint(64) NOT NULL DEFAULT 0,
 		version bigint(64) UNSIGNED NOT NULL DEFAULT 0,
-		PRIMARY KEY (table_id));`
+		PRIMARY KEY (table_id) CLUSTERED);`
 
 	// CreatePasswordHistoryTable is a table save history passwd.
 	CreatePasswordHistoryTable = `CREATE TABLE  IF NOT EXISTS mysql.password_history (

pkg/session/test/bootstraptest/BUILD.bazel

@@ -9,7 +9,7 @@ go_test(
         "main_test.go",
     ],
     flaky = True,
-    shard_count = 43,
+    shard_count = 44,
     deps = [
         "//pkg/config",
         "//pkg/config/kerneltype",

pkg/session/test/bootstraptest/bootstrap_upgrade_test.go

@@ -102,6 +102,71 @@ func TestUpgradeVersion83AndVersion84(t *testing.T) {
 	}
 }
 
+// TestMysqlTablesWithoutClusteredPK pins the `mysql` base tables without a clustered
+// PRIMARY KEY. A diff flags a regression or a new table whose PK type was not
+// considered; the list also serves as a reference for future clustered-PK work.
+//
+// The classic and next-gen kernels bootstrap system tables through different paths
+// (DDL statements vs. direct meta KV writes). The next-gen path clusters every table
+// that has a PRIMARY KEY, while the classic path leaves composite and non-integer PKs
+// non-clustered, so the expected set differs by kernel.
+func TestMysqlTablesWithoutClusteredPK(t *testing.T) {
+	store := testkit.CreateMockStore(t)
+	tk := testkit.NewTestKit(t, store)
+
+	expected := []string{
+		"advisory_locks",
+		"bind_info",
+		"columns_priv",
+		"db",
+		"default_roles",
+		"dist_framework_meta",
+		"expr_pushdown_blacklist",
+		"gc_delete_range",
+		"gc_delete_range_done",
+		"global_grants",
+		"global_priv",
+		"global_variables",
+		"opt_rule_blacklist",
+		"password_history",
+		"plan_replayer_status",
+		"plan_replayer_task",
+		"request_unit_by_group",
+		"role_edges",
+		"stats_extended",
+		"stats_feedback",
+		"stats_top_n",
+		"tables_priv",
+		"tidb",
+		"tidb_ddl_reorg",
+		"tidb_kernel_options",
+		"tidb_pitr_id_map",
+		"tidb_runaway_queries",
+		"tidb_ttl_job_history",
+		"tidb_ttl_task",
+		"user",
+	}
+	if kerneltype.IsNextGen() {
+		expected = []string{
+			"bind_info",
+			"expr_pushdown_blacklist",
+			"gc_delete_range",
+			"gc_delete_range_done",
+			"opt_rule_blacklist",
+			"plan_replayer_status",
+			"stats_feedback",
+			"stats_top_n",
+			"tidb_ddl_reorg",
+			"tidb_runaway_queries",
+		}
+	}
+	tk.MustQuery(
+		"SELECT table_name FROM information_schema.tables " +
+			"WHERE table_schema = 'mysql' AND table_type = 'BASE TABLE' " +
+			"AND tidb_pk_type <> 'CLUSTERED' ORDER BY table_name",
+	).Check(testkit.Rows(expected...))
+}
+
 func revertVersionAndVariables(t *testing.T, se sessionapi.Session, ver int) {
 	session.MustExec(t, se, fmt.Sprintf("update mysql.tidb set variable_value='%d' where variable_name='tidb_server_version'", ver))
 	if ver <= 195 {

pkg/statistics/handle/bootstrap.go

@@ -352,6 +352,10 @@ func (o genHistSQLOptions) assert() {
 // We need to read all the records since we need to do initialization of table.ColAndIdxExistenceMap.
 func genInitStatsHistogramsSQL(options genHistSQLOptions) string {
 	options.assert()
+	// Keep the ORDER_INDEX(tbl) hint: `tbl` still exists on clusters bootstrapped
+	// before stats_histograms moved to a clustered PRIMARY KEY. On fresh clusters
+	// it is inapplicable but harmless: the clustered PK scan is already ordered by
+	// table_id.
 	selectPrefix := "select /*+ ORDER_INDEX(mysql.stats_histograms,tbl) */ HIGH_PRIORITY table_id, is_index, hist_id, distinct_count, version, null_count, cm_sketch, tot_col_size, stats_ver, correlation from mysql.stats_histograms"
 	orderSuffix := " order by table_id"
 
@@ -588,6 +592,7 @@ func genInitStatsTopNSQLForIndexes(isPaging bool, tableRange [2]int64) string {
 // Returns a map where keys are table IDs that have bucket entries.
 func getTablesWithBucketsInRange(sctx sessionctx.Context, tableRange [2]int64) (map[int64]struct{}, error) {
 	// Query to find table_ids that have buckets in the given range
+	// Keep the USE_INDEX(tbl) hint for upgraded clusters; see genInitStatsHistogramsSQL.
 	// TODO: Figure out if HIGH_PRIORITY is working as intended here.
 	sql := "select /*+ USE_INDEX(stats_buckets, tbl) */ HIGH_PRIORITY distinct table_id from mysql.stats_buckets" +
 		" where is_index = 1" +
@@ -732,6 +737,7 @@ func (*Handle) initStatsBuckets4Chunk(cache statstypes.StatsCache, iter *chunk.I
 // We only need to load the indexes' since we only record the existence of columns in ColAndIdxExistenceMap.
 // The stats of the column is not loaded during the bootstrap process.
 func genInitStatsBucketsSQLForIndexes(isPaging bool, tableRange [2]int64) string {
+	// Keep the ORDER_INDEX(tbl) hint for upgraded clusters; see genInitStatsHistogramsSQL.
 	selectPrefix := "select /*+ ORDER_INDEX(mysql.stats_buckets,tbl) */ HIGH_PRIORITY table_id, hist_id, count, repeats, lower_bound, upper_bound, ndv from mysql.stats_buckets where is_index=1"
 	orderSuffix := " order by table_id"
 	if !isPaging {

pkg/statistics/handle/history/history_stats.go

@@ -158,7 +158,11 @@ func RecordHistoricalStatsMeta(
 	}
 
 	modifyCount, count := rows[0].GetInt64(0), rows[0].GetInt64(1)
-	const sql = "REPLACE INTO mysql.stats_meta_history(table_id, modify_count, count, version, source, create_time) VALUES (%?, %?, %?, %?, %?, NOW())"
+	// Use NOW(6) to match the datetime(6) precision of the create_time column.
+	// Plain NOW() truncates to seconds, so rows written within the same second
+	// share an indistinguishable create_time. In-tree readers key off `version`,
+	// but full precision keeps create_time usable for ordering and auditing.
+	const sql = "REPLACE INTO mysql.stats_meta_history(table_id, modify_count, count, version, source, create_time) VALUES (%?, %?, %?, %?, %?, NOW(6))"
 	if _, err := handleutil.ExecWithCtx(
 		ctx,
 		sctx,

tests/integrationtest/r/executor/infoschema_reader.result

@@ -179,7 +179,7 @@ set global tidb_ddl_enable_fast_reorg = default;
 set global tidb_enable_dist_task = default;
 select * from information_schema.KEY_COLUMN_USAGE where TABLE_NAME='stats_meta' and COLUMN_NAME='table_id';
 CONSTRAINT_CATALOG	CONSTRAINT_SCHEMA	CONSTRAINT_NAME	TABLE_CATALOG	TABLE_SCHEMA	TABLE_NAME	COLUMN_NAME	ORDINAL_POSITION	POSITION_IN_UNIQUE_CONSTRAINT	REFERENCED_TABLE_SCHEMA	REFERENCED_TABLE_NAME	REFERENCED_COLUMN_NAME
-def	mysql	tbl	def	mysql	stats_meta	table_id	1	NULL	NULL	NULL	NULL
+def	mysql	PRIMARY	def	mysql	stats_meta	table_id	1	1	NULL	NULL	NULL
 create user key_column_tester;
 select * from information_schema.KEY_COLUMN_USAGE where TABLE_NAME != 'CLUSTER_SLOW_QUERY';
 CONSTRAINT_CATALOG	CONSTRAINT_SCHEMA	CONSTRAINT_NAME	TABLE_CATALOG	TABLE_SCHEMA	TABLE_NAME	COLUMN_NAME	ORDINAL_POSITION	POSITION_IN_UNIQUE_CONSTRAINT	REFERENCED_TABLE_SCHEMA	REFERENCED_TABLE_NAME	REFERENCED_COLUMN_NAME