cherry: cherry: feat: adding bulk call for alpha to inform zero about the tablets (#8… (#8100)

* feat: adding bulk call for alpha to inform zero about the tablets (#8089)

* adding bulk call for alpha to inform zero about the tablets (#8088)

(cherry picked from commit c204d0a)

* fix(zero): fix update membership to make bulk tablet proposal instead of multiple small (#8090)

Converting the smaller tablet proposals into a single bulk call makes the execution faster. Else, this update becomes of the order of O(n^2), where n is the number of tablet updates.

(cherry picked from commit 695c6d7)

Co-authored-by: aman bansal <[email protected]>

Author: NamanJain8

dgraph/cmd/zero/raft.go

@@ -298,12 +298,25 @@ func (n *node) regenerateChecksum() {
 	}
 }
 
-func (n *node) handleTabletProposal(tablet *pb.Tablet) error {
+func (n *node) handleBulkTabletProposal(tablets []*pb.Tablet) error {
 	n.server.AssertLock()
-	state := n.server.state
-
 	defer n.regenerateChecksum()
+	for _, tablet := range tablets {
+		if err := n.handleTablet(tablet); err != nil {
+			glog.Warningf("not able to handle tablet %s. Got err: %+v", tablet.GetPredicate(), err)
+		}
+	}
 
+	return nil
+}
+
+// handleTablet will check if the given tablet is served by any group.
+// If not the tablet will be added to the current group predicate list
+//
+// This function doesn't take any locks.
+// It is the calling functions responsibility to manage the concurrency.
+func (n *node) handleTablet(tablet *pb.Tablet) error {
+	state := n.server.state
 	if tablet.GroupId == 0 {
 		return errors.Errorf("Tablet group id is zero: %+v", tablet)
 	}
@@ -339,6 +352,12 @@ func (n *node) handleTabletProposal(tablet *pb.Tablet) error {
 	return nil
 }
 
+func (n *node) handleTabletProposal(tablet *pb.Tablet) error {
+	n.server.AssertLock()
+	defer n.regenerateChecksum()
+	return n.handleTablet(tablet)
+}
+
 func (n *node) deleteNamespace(delNs uint64) error {
 	n.server.AssertLock()
 	state := n.server.state
@@ -431,6 +450,15 @@ func (n *node) applyProposal(e raftpb.Entry) (uint64, error) {
 			return key, err
 		}
 	}
+
+	if p.Tablets != nil && len(p.Tablets) > 0 {
+		if err := n.handleBulkTabletProposal(p.Tablets); err != nil {
+			span.Annotatef(nil, "While applying bulk tablet proposal: %v", err)
+			glog.Errorf("While applying bulk tablet proposal: %v", err)
+			return key, err
+		}
+	}
+
 	if p.License != nil {
 		// Check that the number of nodes in the cluster should be less than MaxNodes, otherwise
 		// reject the proposal.

dgraph/cmd/zero/zero.go

@@ -316,10 +316,8 @@ func (s *Server) ServingTablet(tablet string) *pb.Tablet {
 	defer s.RUnlock()
 
 	for _, group := range s.state.Groups {
-		for key, tab := range group.Tablets {
-			if key == tablet {
-				return tab
-			}
+		if tab, ok := group.Tablets[tablet]; ok {
+			return tab
 		}
 	}
 	return nil
@@ -341,10 +339,8 @@ func (s *Server) servingTablet(tablet string) *pb.Tablet {
 	s.AssertRLock()
 
 	for _, group := range s.state.Groups {
-		for key, tab := range group.Tablets {
-			if key == tablet {
-				return tab
-			}
+		if tab, ok := group.Tablets[tablet]; ok {
+			return tab
 		}
 	}
 	return nil
@@ -386,6 +382,8 @@ func (s *Server) createProposals(dst *pb.Group) ([]*pb.ZeroProposal, error) {
 			})
 		}
 	}
+
+	var tablets []*pb.Tablet
 	for key, dstTablet := range dst.Tablets {
 		group, has := s.state.Groups[dstTablet.GroupId]
 		if !has {
@@ -401,15 +399,81 @@ func (s *Server) createProposals(dst *pb.Group) ([]*pb.ZeroProposal, error) {
 		d := float64(dstTablet.OnDiskBytes)
 		if dstTablet.Remove || (s == 0 && d > 0) || (s > 0 && math.Abs(d/s-1) > 0.1) {
 			dstTablet.Force = false
-			proposal := &pb.ZeroProposal{
-				Tablet: dstTablet,
-			}
-			res = append(res, proposal)
+			tablets = append(tablets, dstTablet)
 		}
 	}
+
+	if len(tablets) > 0 {
+		res = append(res, &pb.ZeroProposal{Tablets: tablets})
+	}
 	return res, nil
 }
 
+func (s *Server) Inform(ctx context.Context, req *pb.TabletRequest) (*pb.TabletResponse, error) {
+	ctx, span := otrace.StartSpan(ctx, "Zero.Inform")
+	defer span.End()
+	if req == nil || len(req.Tablets) == 0 {
+		return nil, errors.Errorf("Tablets are empty in %+v", req)
+	}
+
+	if req.GroupId == 0 {
+		return nil, errors.Errorf("Group ID is Zero in %+v", req)
+	}
+
+	tablets := make([]*pb.Tablet, 0)
+	unknownTablets := make([]*pb.Tablet, 0)
+	for _, t := range req.Tablets {
+		tab := s.ServingTablet(t.Predicate)
+		span.Annotatef(nil, "Tablet for %s: %+v", t.Predicate, tab)
+		switch {
+		case tab != nil && !t.Force:
+			tablets = append(tablets, t)
+		case t.ReadOnly:
+			tablets = append(tablets, &pb.Tablet{})
+		default:
+			unknownTablets = append(unknownTablets, t)
+		}
+	}
+
+	if len(unknownTablets) == 0 {
+		return &pb.TabletResponse{
+			Tablets: tablets,
+		}, nil
+	}
+
+	// Set the tablet to be served by this server's group.
+	var proposal pb.ZeroProposal
+	proposal.Tablets = make([]*pb.Tablet, 0)
+	for _, t := range unknownTablets {
+		if x.IsReservedPredicate(t.Predicate) {
+			// Force all the reserved predicates to be allocated to group 1.
+			// This is to make it easier to stream ACL updates to all alpha servers
+			// since they only need to open one pipeline to receive updates for all
+			// ACL predicates.
+			// This will also make it easier to restore the reserved predicates after
+			// a DropAll operation.
+			t.GroupId = 1
+		}
+		proposal.Tablets = append(proposal.Tablets, t)
+	}
+
+	if err := s.Node.proposeAndWait(ctx, &proposal); err != nil && err != errTabletAlreadyServed {
+		span.Annotatef(nil, "While proposing tablet: %v", err)
+		return nil, err
+	}
+
+	for _, t := range unknownTablets {
+		tab := s.ServingTablet(t.Predicate)
+		x.AssertTrue(tab != nil)
+		span.Annotatef(nil, "Now serving tablet for %s: %+v", t.Predicate, tab)
+		tablets = append(tablets, tab)
+	}
+
+	return &pb.TabletResponse{
+		Tablets: tablets,
+	}, nil
+}
+
 // RemoveNode removes the given node from the given group.
 // It's the user's responsibility to ensure that node doesn't come back again
 // before calling the api.

protos/pb.proto

@@ -176,6 +176,7 @@ message ZeroProposal {
   ZeroSnapshot snapshot = 11;  // Used to make Zeros take a snapshot.
   // 12 has already been used.
   DeleteNsRequest delete_ns = 13;  // Used to delete namespace.
+  repeated Tablet tablets = 14;
 }
 
 // MembershipState is used to pack together the current membership state of all
@@ -556,6 +557,7 @@ service Zero {
 
   rpc Oracle(api.Payload) returns (stream OracleDelta) {}
   rpc ShouldServe(Tablet) returns (Tablet) {}
+  rpc Inform(TabletRequest) returns (TabletResponse) {}
   rpc AssignIds(Num) returns (AssignedIds) {}
   rpc Timestamps(Num) returns (AssignedIds) {}
   rpc CommitOrAbort(api.TxnContext) returns (api.TxnContext) {}
@@ -585,6 +587,14 @@ service Worker {
   rpc TaskStatus(TaskStatusRequest) returns (TaskStatusResponse) {}
 }
 
+message TabletResponse {
+  repeated Tablet tablets = 1;
+}
+message TabletRequest {
+  repeated Tablet tablets = 1;
+  uint32 group_id = 2 [(gogoproto.jsontag) = "groupId,omitempty"];  // Served by which group.
+}
+
 message SubscriptionRequest {
   repeated bytes prefixes = 1;
   repeated badgerpb3.Match matches = 2;