// _ _ // __ _____ __ ___ ___ __ _| |_ ___ // \ \ /\ / / _ \/ _` \ \ / / |/ _` | __/ _ \ // \ V V / __/ (_| |\ V /| | (_| | || __/ // \_/\_/ \___|\__,_| \_/ |_|\__,_|\__\___| // // Copyright © 2016 - 2025 Weaviate B.V. All rights reserved. // // CONTACT: hello@weaviate.io // package schema import ( "bytes" "context" "encoding/json" "errors" "fmt" "slices" "github.com/prometheus/client_golang/prometheus" "github.com/sirupsen/logrus" command "github.com/weaviate/weaviate/cluster/proto/api" "github.com/weaviate/weaviate/entities/models" "github.com/weaviate/weaviate/usecases/sharding" gproto "google.golang.org/protobuf/proto" ) var ( ErrBadRequest = errors.New("bad request") errDB = errors.New("updating db") ErrSchema = errors.New("updating schema") ) type replicationFSM interface { HasOngoingReplication(collection string, shard string, replica string) bool DeleteReplicationsByCollection(collection string) error DeleteReplicationsByTenants(collection string, tenants []string) error SetUnCancellable(id uint64) error } type SchemaManager struct { schema *schema db Indexer parser Parser log *logrus.Logger replicationFSM replicationFSM } func NewSchemaManager(nodeId string, db Indexer, parser Parser, reg prometheus.Registerer, log *logrus.Logger) *SchemaManager { return &SchemaManager{ schema: NewSchema(nodeId, db, reg), db: db, parser: parser, log: log, } } func (s *SchemaManager) NewSchemaReader() SchemaReader { return NewSchemaReader( s.schema, // Pass a versioned reader that will ignore all version and always return valid, we want to read the latest // state and not have to wait on a version VersionedSchemaReader{ schema: s.schema, WaitForUpdate: func(context.Context, uint64) error { return nil }, }, ) } func (s *SchemaManager) NewSchemaReaderWithWaitFunc(f func(context.Context, uint64) error) SchemaReader { return NewSchemaReader( s.schema, VersionedSchemaReader{ schema: s.schema, WaitForUpdate: f, }, ) } func (s *SchemaManager) SetIndexer(idx Indexer) { s.db = idx s.schema.shardReader = idx } func (s *SchemaManager) SetReplicationFSM(fsm replicationFSM) { s.replicationFSM = fsm } func (s *SchemaManager) SchemaSnapshot() ([]byte, error) { var buf bytes.Buffer err := json.NewEncoder(&buf).Encode(s.schema.MetaClasses()) return buf.Bytes(), err } func (s *SchemaManager) AliasSnapshot() ([]byte, error) { var buf bytes.Buffer err := json.NewEncoder(&buf).Encode(s.schema.aliases) return buf.Bytes(), err } func (s *SchemaManager) Restore(data []byte, parser Parser) error { return s.schema.Restore(data, parser) } func (s *SchemaManager) RestoreAliases(data []byte) error { return s.schema.RestoreAlias(data) } func (s *SchemaManager) RestoreLegacy(data []byte, parser Parser) error { return s.schema.RestoreLegacy(data, parser) } func (s *SchemaManager) PreApplyFilter(req *command.ApplyRequest) error { classInfo := s.schema.ClassInfo(req.Class) // Discard restoring a class if it already exists if req.Type == command.ApplyRequest_TYPE_RESTORE_CLASS && classInfo.Exists { s.log.WithField("class", req.Class).Info("class already restored") return fmt.Errorf("class name %s already exists", req.Class) } // Discard adding class if the name already exists or a similar one exists if req.Type == command.ApplyRequest_TYPE_ADD_CLASS { other, isAlias := s.schema.ClassEqual(req.Class) item := "class" if isAlias { item = "alias" } if other == req.Class { return fmt.Errorf("%s name %s already exists", item, req.Class) } else if other != "" { return fmt.Errorf("%w: found similar %s %q", ErrClassExists, item, other) } } return nil } func (s *SchemaManager) Load(ctx context.Context, nodeID string) error { if err := s.db.Open(ctx); err != nil { return err } return nil } func (s *SchemaManager) ReloadDBFromSchema() { classes := s.schema.MetaClasses() cs := make([]command.UpdateClassRequest, len(classes)) i := 0 for _, v := range classes { migratePropertiesIfNecessary(&v.Class) cs[i] = command.UpdateClassRequest{Class: &v.Class, State: &v.Sharding} i++ } s.db.TriggerSchemaUpdateCallbacks() s.log.Info("reload local db: update schema ...") s.db.ReloadLocalDB(context.Background(), cs) } func (s *SchemaManager) Close(ctx context.Context) (err error) { return s.db.Close(ctx) } func (s *SchemaManager) AddClass(cmd *command.ApplyRequest, nodeID string, schemaOnly bool, enableSchemaCallback bool) error { req := command.AddClassRequest{} // dupa if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } if req.State == nil { return fmt.Errorf("%w: nil sharding state", ErrBadRequest) } if err := s.parser.ParseClass(req.Class); err != nil { return fmt.Errorf("%w: parsing class: %w", ErrBadRequest, err) } req.State.SetLocalName(nodeID) // We need to make a copy of the sharding state to ensure that the state stored in the internal schema has no // references to. As we will make modification to it to reflect change in the sharding state (adding/removing // tenant) we don't want another goroutine holding a pointer to it and finding issues with concurrent read/writes. shardingStateCopy := req.State.DeepCopy() return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.addClass(req.Class, &shardingStateCopy, cmd.Version) }, updateStore: func() error { return s.db.AddClass(req) }, schemaOnly: schemaOnly, enableSchemaCallback: enableSchemaCallback, }, ) } func (s *SchemaManager) RestoreClass(cmd *command.ApplyRequest, nodeID string, schemaOnly bool, enableSchemaCallback bool) error { req := command.AddClassRequest{} if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } if req.State == nil { return fmt.Errorf("%w: nil sharding state", ErrBadRequest) } if err := s.parser.ParseClass(req.Class); err != nil { return fmt.Errorf("%w: parsing class: %w", ErrBadRequest, err) } req.State.SetLocalName(nodeID) if err := s.db.RestoreClassDir(cmd.Class); err != nil { s.log.WithField("class", cmd.Class).WithError(err). Error("restore class directory from backup") // continue since we need to add class to the schema anyway } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.addClass(req.Class, req.State, cmd.Version) }, updateStore: func() error { return s.db.AddClass(req) }, schemaOnly: schemaOnly, enableSchemaCallback: enableSchemaCallback, }, ) } // ReplaceStatesNodeName it update the node name inside sharding states. // WARNING: this shall be used in one node cluster environments only. // because it will replace the shard node name if the node name got updated // only if the replication factor is 1, otherwise it's no-op func (s *SchemaManager) ReplaceStatesNodeName(new string) { s.schema.replaceStatesNodeName(new) } // UpdateClass modifies the vectors and inverted indexes associated with a class // Other class properties are handled by separate functions func (s *SchemaManager) UpdateClass(cmd *command.ApplyRequest, nodeID string, schemaOnly bool, enableSchemaCallback bool) error { req := command.UpdateClassRequest{} if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } if req.State != nil { req.State.SetLocalName(nodeID) } update := func(meta *metaClass) error { // Ensure that if non-default values for properties is stored in raft we fix them before processing an update to // avoid triggering diff on properties and therefore discarding a legitimate update. migratePropertiesIfNecessary(&meta.Class) u, err := s.parser.ParseClassUpdate(&meta.Class, req.Class) if err != nil { return fmt.Errorf("%w :parse class update: %w", ErrBadRequest, err) } meta.Class.VectorIndexConfig = u.VectorIndexConfig meta.Class.InvertedIndexConfig = u.InvertedIndexConfig meta.Class.VectorConfig = u.VectorConfig meta.Class.ReplicationConfig = u.ReplicationConfig meta.Class.MultiTenancyConfig = u.MultiTenancyConfig meta.Class.Description = u.Description meta.Class.Properties = u.Properties meta.ClassVersion = cmd.Version if req.State != nil { meta.Sharding = *req.State } return nil } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.updateClass(req.Class.Class, update) }, updateStore: func() error { return s.db.UpdateClass(req) }, schemaOnly: schemaOnly, enableSchemaCallback: enableSchemaCallback, }, ) } func (s *SchemaManager) DeleteClass(cmd *command.ApplyRequest, schemaOnly bool, enableSchemaCallback bool) error { var hasFrozen bool tenants, err := s.schema.getTenants(cmd.Class, nil) if err != nil { hasFrozen = false } for _, t := range tenants { if t.ActivityStatus == models.TenantActivityStatusFROZEN || t.ActivityStatus == models.TenantActivityStatusFREEZING { hasFrozen = true break } } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { s.schema.deleteClass(cmd.Class); return nil }, updateStore: func() error { if s.replicationFSM == nil { return fmt.Errorf("replication deleter is not set, this should never happen") } else if err := s.replicationFSM.DeleteReplicationsByCollection(cmd.Class); err != nil { // If there is an error deleting the replications then we log it but make sure not to block the deletion of the class from a UX PoV s.log.WithField("error", err).WithField("class", cmd.Class).Error("could not delete replication operations for deleted class") } return s.db.DeleteClass(cmd.Class, hasFrozen) }, schemaOnly: schemaOnly, enableSchemaCallback: enableSchemaCallback, }, ) } func (s *SchemaManager) AddProperty(cmd *command.ApplyRequest, schemaOnly bool, enableSchemaCallback bool) error { req := command.AddPropertyRequest{} if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } if len(req.Properties) == 0 { return fmt.Errorf("%w: empty property", ErrBadRequest) } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.addProperty(cmd.Class, cmd.Version, req.Properties...) }, updateStore: func() error { return s.db.AddProperty(cmd.Class, req) }, schemaOnly: schemaOnly, enableSchemaCallback: enableSchemaCallback, }, ) } func (s *SchemaManager) UpdateShardStatus(cmd *command.ApplyRequest, schemaOnly bool) error { req := command.UpdateShardStatusRequest{} if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return nil }, updateStore: func() error { return s.db.UpdateShardStatus(&req) }, schemaOnly: schemaOnly, }, ) } func (s *SchemaManager) AddReplicaToShard(cmd *command.ApplyRequest, schemaOnly bool) error { req := command.AddReplicaToShard{} if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.addReplicaToShard(cmd.Class, cmd.Version, req.Shard, req.TargetNode) }, updateStore: func() error { if req.TargetNode == s.schema.nodeID { return s.db.AddReplicaToShard(req.Class, req.Shard, req.TargetNode) } return nil }, schemaOnly: schemaOnly, }, ) } func (s *SchemaManager) DeleteReplicaFromShard(cmd *command.ApplyRequest, schemaOnly bool) error { req := command.DeleteReplicaFromShard{} if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.deleteReplicaFromShard(cmd.Class, cmd.Version, req.Shard, req.TargetNode) }, updateStore: func() error { if req.TargetNode == s.schema.nodeID { return s.db.DeleteReplicaFromShard(req.Class, req.Shard, req.TargetNode) } return nil }, schemaOnly: schemaOnly, }, ) } func (s *SchemaManager) AddTenants(cmd *command.ApplyRequest, schemaOnly bool) error { req := &command.AddTenantsRequest{} if err := gproto.Unmarshal(cmd.SubCommand, req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.addTenants(cmd.Class, cmd.Version, req) }, updateStore: func() error { return s.db.AddTenants(cmd.Class, req) }, schemaOnly: schemaOnly, }, ) } func (s *SchemaManager) UpdateTenants(cmd *command.ApplyRequest, schemaOnly bool) error { req := &command.UpdateTenantsRequest{} if err := gproto.Unmarshal(cmd.SubCommand, req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } return s.apply( applyOp{ op: cmd.GetType().String(), // updateSchema func will update the request's tenants and therefore we use it as a filter that is then sent // to the updateStore function. This allows us to effectively use the schema update to narrow down work for // the DB update. updateSchema: func() error { return s.schema.updateTenants(cmd.Class, cmd.Version, req, s.replicationFSM) }, updateStore: func() error { return s.db.UpdateTenants(cmd.Class, req) }, schemaOnly: schemaOnly, }, ) } func (s *SchemaManager) DeleteTenants(cmd *command.ApplyRequest, schemaOnly bool) error { req := &command.DeleteTenantsRequest{} if err := gproto.Unmarshal(cmd.SubCommand, req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } tenants, err := s.schema.getTenants(cmd.Class, req.Tenants) if err != nil { // error are handled by the updateSchema, so they are ignored here. // Instead, we log the error to detect tenant status before deleting // them from the schema. this allows the database layer to decide whether // to send the delete request to the cloud provider. s.log.WithFields(logrus.Fields{ "class": cmd.Class, "tenants": req.Tenants, "error": err.Error(), }).Error("error getting tenants") } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.deleteTenants(cmd.Class, cmd.Version, req) }, updateStore: func() error { if s.replicationFSM == nil { return fmt.Errorf("replication deleter is not set, this should never happen") } else if err := s.replicationFSM.DeleteReplicationsByTenants(cmd.Class, req.Tenants); err != nil { // If there is an error deleting the replications then we log it but make sure not to block the deletion of the class from a UX PoV s.log.WithField("error", err).WithField("class", cmd.Class).WithField("tenants", tenants).Error("could not delete replication operations for deleted tenants") } return s.db.DeleteTenants(cmd.Class, tenants) }, schemaOnly: schemaOnly, }, ) } func (s *SchemaManager) UpdateTenantsProcess(cmd *command.ApplyRequest, schemaOnly bool) error { req := &command.TenantProcessRequest{} if err := gproto.Unmarshal(cmd.SubCommand, req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return s.schema.updateTenantsProcess(cmd.Class, cmd.Version, req) }, updateStore: func() error { return s.db.UpdateTenantsProcess(cmd.Class, req) }, schemaOnly: schemaOnly, }, ) } func (s *SchemaManager) SyncShard(cmd *command.ApplyRequest, schemaOnly bool) error { req := command.SyncShardRequest{} if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } if req.NodeId != s.schema.nodeID { return nil } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { return nil }, updateStore: func() error { return s.schema.Read(req.Collection, func(class *models.Class, state *sharding.State) error { physical, ok := state.Physical[req.Shard] // shard does not exist in the sharding state if !ok { // TODO: can we guarantee that the shard is not in use? // If so we should call s.db.DropShard(cmd.Class, req.Shard) here instead // For now, to be safe and avoid data loss, we just shut it down s.db.ShutdownShard(cmd.Class, req.Shard) // return early return nil } // if shard doesn't belong to this node if !slices.Contains(physical.BelongsToNodes, req.NodeId) { // shut it down s.db.ShutdownShard(cmd.Class, req.Shard) // return early return nil } // collection is single-tenant, shard is present, replica belongs to node if !state.PartitioningEnabled { // load it s.db.LoadShard(cmd.Class, req.Shard) // return early return nil } // collection is multi-tenant, shard is present, replica belongs to node switch physical.ActivityStatus() { // tenant is active case models.TenantActivityStatusACTIVE: // load it s.db.LoadShard(cmd.Class, req.Shard) // tenant is inactive case models.TenantActivityStatusINACTIVE: // shut it down s.db.ShutdownShard(cmd.Class, req.Shard) // tenant is in some other state default: // do nothing } return nil }) }, schemaOnly: schemaOnly, }, ) } func (s *SchemaManager) ReplicationAddReplicaToShard(cmd *command.ApplyRequest, schemaOnly bool) error { req := command.ReplicationAddReplicaToShard{} if err := json.Unmarshal(cmd.SubCommand, &req); err != nil { return fmt.Errorf("%w: %w", ErrBadRequest, err) } return s.apply( applyOp{ op: cmd.GetType().String(), updateSchema: func() error { err := s.replicationFSM.SetUnCancellable(req.OpId) if err != nil { return fmt.Errorf("set un-cancellable: %w", err) } return s.schema.addReplicaToShard(cmd.Class, cmd.Version, req.Shard, req.TargetNode) }, updateStore: func() error { if req.TargetNode == s.schema.nodeID { return s.db.AddReplicaToShard(req.Class, req.Shard, req.TargetNode) } return nil }, schemaOnly: schemaOnly, }, ) } type applyOp struct { op string updateSchema func() error updateStore func() error schemaOnly bool enableSchemaCallback bool } func (op applyOp) validate() error { if op.op == "" { return fmt.Errorf("op is not specified") } if op.updateSchema == nil { return fmt.Errorf("updateSchema func is nil") } if op.updateStore == nil { return fmt.Errorf("updateStore func is nil") } return nil } // apply does apply commands from RAFT to schema 1st and then db func (s *SchemaManager) apply(op applyOp) error { if err := op.validate(); err != nil { return fmt.Errorf("could not validate raft apply op: %w", err) } // schema applied 1st to make sure any validation happen before applying it to db if err := op.updateSchema(); err != nil { return fmt.Errorf("%w: %s: %w", ErrSchema, op.op, err) } if op.enableSchemaCallback && s.db != nil { // TriggerSchemaUpdateCallbacks is concurrent and at // this point of time schema shall be up to date. s.db.TriggerSchemaUpdateCallbacks() } if !op.schemaOnly { if err := op.updateStore(); err != nil { return fmt.Errorf("%w: %s: %w", errDB, op.op, err) } } return nil } // migratePropertiesIfNecessary migrate properties and set default values for them. // This is useful when adding new properties to ensure that their default value is properly set. // Current migrated properties: // IndexRangeFilters was introduced with 1.26, so objects which were created // on an older version, will have this value set to nil when the instance is // upgraded. If we come across a property with nil IndexRangeFilters, it // needs to be set as false, to avoid false positive class differences on // comparison during class updates. func migratePropertiesIfNecessary(class *models.Class) { for _, prop := range class.Properties { if prop.IndexRangeFilters == nil { prop.IndexRangeFilters = func() *bool { f := false; return &f }() } // Ensure we also migrate nested properties for _, nprop := range prop.NestedProperties { migrateNestedPropertiesIfNecessary(nprop) } } } func migrateNestedPropertiesIfNecessary(nprop *models.NestedProperty) { // migrate this nested property nprop.IndexRangeFilters = func() *bool { f := false; return &f }() // Recurse on all nested properties this one has for _, recurseNestedProperty := range nprop.NestedProperties { migrateNestedPropertiesIfNecessary(recurseNestedProperty) } }