workload/schemachange: make RSW more deterministic by using setseed

The recently added setseed builtin lets us make the ORDER BY random()
clauses used in this workload deterministic.

Each worker goroutine of the RSW now runs a determistic sequence of
operations given the same COCKROACH_RANDOM_SEED. However, since the
workload is concurrent, the overall RSW is not deterministic, since the
interleaving of goroutine execution is random.

This also makes the "deck" used to decide which operation to execute be
per-worker rather than global.

Release note: None

pkg/workload/schemachange/operation_generator.go

@@ -3256,6 +3256,9 @@ func (og *operationGenerator) getTableColumns(
 func (og *operationGenerator) randColumn(
 	ctx context.Context, tx pgx.Tx, tableName tree.TableName, pctExisting int,
 ) (string, error) {
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return "", err
+	}
 	if og.randIntn(100) >= pctExisting {
 		// We make a unique name for all columns by prefixing them with the table
 		// index to make it easier to reference columns from different tables.
@@ -3282,6 +3285,9 @@ ORDER BY random()
 func (og *operationGenerator) randColumnWithMeta(
 	ctx context.Context, tx pgx.Tx, tableName tree.TableName, pctExisting int,
 ) (column, error) {
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return column{}, err
+	}
 	if og.randIntn(100) >= pctExisting {
 		// We make a unique name for all columns by prefixing them with the table
 		// index to make it easier to reference columns from different tables.
@@ -3371,7 +3377,9 @@ func (og *operationGenerator) randChildColumnForFkRelation(
 func (og *operationGenerator) randParentColumnForFkRelation(
 	ctx context.Context, tx pgx.Tx, unique bool,
 ) (*tree.TableName, *column, error) {
-
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return nil, nil, err
+	}
 	subQuery := strings.Builder{}
 	subQuery.WriteString(`
 		SELECT table_schema, table_name, column_name, crdb_sql_type, is_nullable, contype, conkey
@@ -3444,6 +3452,9 @@ func (og *operationGenerator) randParentColumnForFkRelation(
 func (og *operationGenerator) randConstraint(
 	ctx context.Context, tx pgx.Tx, tableName string,
 ) (string, error) {
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return "", err
+	}
 	q := fmt.Sprintf(`
   SELECT constraint_name
     FROM [SHOW CONSTRAINTS FROM %s]
@@ -3461,6 +3472,9 @@ ORDER BY random()
 func (og *operationGenerator) randIndex(
 	ctx context.Context, tx pgx.Tx, tableName tree.TableName, pctExisting int,
 ) (string, error) {
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return "", err
+	}
 	if og.randIntn(100) >= pctExisting {
 		// We make a unique name for all indices by prefixing them with the table
 		// index to make it easier to reference columns from different tables.
@@ -3485,7 +3499,9 @@ ORDER BY random()
 func (og *operationGenerator) randSequence(
 	ctx context.Context, tx pgx.Tx, pctExisting int, desiredSchema string,
 ) (*tree.TableName, error) {
-
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return nil, err
+	}
 	if desiredSchema != "" {
 		if og.randIntn(100) >= pctExisting {
 			treeSeqName := tree.MakeTableNameFromPrefix(tree.ObjectNamePrefix{
@@ -3587,7 +3603,9 @@ ORDER BY random()
 func (og *operationGenerator) randTable(
 	ctx context.Context, tx pgx.Tx, pctExisting int, desiredSchema string,
 ) (*tree.TableName, error) {
-
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return nil, err
+	}
 	if desiredSchema != "" {
 		if og.randIntn(100) >= pctExisting {
 			treeTableName := tree.MakeTableNameFromPrefix(tree.ObjectNamePrefix{
@@ -3658,6 +3676,9 @@ ORDER BY random()
 func (og *operationGenerator) randView(
 	ctx context.Context, tx pgx.Tx, pctExisting int, desiredSchema string,
 ) (*tree.TableName, error) {
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return nil, err
+	}
 	if desiredSchema != "" {
 		if og.randIntn(100) >= pctExisting {
 			treeViewName := tree.MakeTableNameFromPrefix(tree.ObjectNamePrefix{
@@ -3702,6 +3723,9 @@ func (og *operationGenerator) randView(
 		}, tree.Name(fmt.Sprintf("view_%s", og.newUniqueSeqNumSuffix())))
 		return &treeViewName, nil
 	}
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return nil, err
+	}
 	const q = `
   SELECT schema_name, table_name
     FROM [SHOW TABLES]
@@ -3812,6 +3836,9 @@ func (og *operationGenerator) createSchema(ctx context.Context, tx pgx.Tx) (*opS
 func (og *operationGenerator) randSchema(
 	ctx context.Context, tx pgx.Tx, pctExisting int,
 ) (string, error) {
+	if err := og.setSeedInDB(ctx, tx); err != nil {
+		return "", err
+	}
 	if og.randIntn(100) >= pctExisting {
 		return fmt.Sprintf("schema_%s", og.newUniqueSeqNumSuffix()), nil
 	}
@@ -4309,6 +4336,11 @@ func (og *operationGenerator) randIntn(topBound int) int {
 	return og.params.rng.Intn(topBound)
 }
 
+// randFloat64 returns an float64 in the range [0,1.0).
+func (og *operationGenerator) randFloat64() float64 {
+	return og.params.rng.Float64()
+}
+
 func (og *operationGenerator) newUniqueSeqNumSuffix() string {
 	og.params.seqNum++
 	return fmt.Sprintf("w%d_%d", og.params.workerID, og.params.seqNum)
@@ -4335,7 +4367,7 @@ func (og *operationGenerator) typeFromTypeName(
 }
 
 // Check if the test is running with a mixed version cluster, with a version
-// less than or equal to the target version number. This can be used to detect
+// less than the target version number. This can be used to detect
 // in mixed version environments if certain errors should be encountered.
 func isClusterVersionLessThan(
 	ctx context.Context, tx pgx.Tx, targetVersion roachpb.Version,
@@ -4349,5 +4381,19 @@ func isClusterVersionLessThan(
 	if err != nil {
 		return false, err
 	}
-	return clusterVersion.LessEq(targetVersion), nil
+	return clusterVersion.Less(targetVersion), nil
+}
+
+func (og *operationGenerator) setSeedInDB(ctx context.Context, tx pgx.Tx) error {
+	if notSupported, err := isClusterVersionLessThan(ctx, tx, clusterversion.V24_1.Version()); err != nil {
+		return err
+	} else if notSupported {
+		// To allow the schemachange workload to work in a mixed-version state,
+		// this should not be treated as an error.
+		return nil
+	}
+	if _, err := tx.Exec(ctx, "SELECT setseed($1)", og.randFloat64()); err != nil {
+		return err
+	}
+	return nil
 }

pkg/workload/schemachange/schemachange.go

@@ -201,19 +201,17 @@ func (s *schemaChange) Ops(
 		return workload.QueryLoad{}, err
 	}
 	stdoutLog := makeAtomicLog(os.Stdout)
-	rng, seed := randutil.NewTestRand()
+	_, seed := randutil.NewTestRand()
 	stdoutLog.printLn(fmt.Sprintf("using random seed: %d", seed))
-	ops := newDeck(rng, opWeights...)
-	// A separate deck is constructed of only schema changes supported
-	// by the declarative schema changer. This deck has equal weights,
-	// only for supported schema changes.
+	// A separate weighting is constructed of only schema changes supported by the
+	// declarative schema changer. This will be used to make a per-worker deck
+	// that has equal weights, only for supported schema changes.
 	declarativeOpWeights := make([]int, len(opWeights))
 	for idx, weight := range opWeights {
 		if _, ok := opDeclarativeVersion[opType(idx)]; ok {
 			declarativeOpWeights[idx] = weight
 		}
 	}
-	declarativeOps := newDeck(rng, declarativeOpWeights...)
 
 	ql := workload.QueryLoad{
 		SQLDatabase: sqlDatabase,
@@ -251,12 +249,15 @@ func (s *schemaChange) Ops(
 		// operations.
 		workerRng := randutil.NewTestRandWithSeed(seed + int64(i))
 
-		// Each worker needs its own sequence number generator so that the names of
-		// generated objects are deterministic across runs.
+		// Each worker needs its own sequence number generator and operation deck so
+		// that the names of generated objects and operations are deterministic
+		// across runs.
 		seqNum, err := s.initSeqNum(ctx, pool, i)
 		if err != nil {
 			return workload.QueryLoad{}, err
 		}
+		ops := newDeck(workerRng, opWeights...)
+		declarativeOps := newDeck(workerRng, declarativeOpWeights...)
 
 		opGeneratorParams := operationGeneratorParams{
 			workerID:           i,