open-consul/agent/rpc/peering/replication.go
Freddy 9eeb9e4ee3
Clean up imported nodes/services/checks as needed (#13367)
Previously, imported data would never be deleted. As
nodes/services/checks were registered and deregistered, resources
deleted from the exporting cluster would accumulate in the imported
cluster.

This commit makes updates to replication so that whenever an update is
received for a service name we reconcile what was present in the catalog
against what was received.

This handleUpdateService method can handle both updates and deletions.
2022-06-13 11:52:28 -06:00

480 lines
16 KiB
Go

package peering
import (
"errors"
"fmt"
"strings"
"github.com/golang/protobuf/proto"
"github.com/golang/protobuf/ptypes"
"github.com/hashicorp/consul/types"
"github.com/hashicorp/go-hclog"
"google.golang.org/genproto/googleapis/rpc/code"
"google.golang.org/protobuf/types/known/anypb"
"github.com/hashicorp/consul/agent/cache"
"github.com/hashicorp/consul/agent/structs"
"github.com/hashicorp/consul/proto/pbpeering"
"github.com/hashicorp/consul/proto/pbservice"
"github.com/hashicorp/consul/proto/pbstatus"
)
/*
TODO(peering):
At the start of each peering stream establishment (not initiation, but the
thing that reconnects) we need to do a little bit of light differential
snapshot correction to initially synchronize the local state store.
Then if we ever fail to apply a replication message we should either tear
down the entire connection (and thus force a resync on reconnect) or
request a resync operation.
*/
// makeServiceResponse handles preparing exported service instance updates to the peer cluster.
// Each cache.UpdateEvent will contain all instances for a service name.
// If there are no instances in the event, we consider that to be a de-registration.
func makeServiceResponse(
logger hclog.Logger,
update cache.UpdateEvent,
) *pbpeering.ReplicationMessage {
any, csn, err := marshalToProtoAny[*pbservice.IndexedCheckServiceNodes](update.Result)
if err != nil {
// Log the error and skip this response to avoid locking up peering due to a bad update event.
logger.Error("failed to marshal", "error", err)
return nil
}
serviceName := strings.TrimPrefix(update.CorrelationID, subExportedService)
// If no nodes are present then it's due to one of:
// 1. The service is newly registered or exported and yielded a transient empty update.
// 2. All instances of the service were de-registered.
// 3. The service was un-exported.
//
// We don't distinguish when these three things occurred, but it's safe to send a DELETE Op in all cases, so we do that.
// Case #1 is a no-op for the importing peer.
if len(csn.Nodes) == 0 {
resp := &pbpeering.ReplicationMessage{
Payload: &pbpeering.ReplicationMessage_Response_{
Response: &pbpeering.ReplicationMessage_Response{
ResourceURL: pbpeering.TypeURLService,
// TODO(peering): Nonce management
Nonce: "",
ResourceID: serviceName,
Operation: pbpeering.ReplicationMessage_Response_DELETE,
},
},
}
return resp
}
// If there are nodes in the response, we push them as an UPSERT operation.
resp := &pbpeering.ReplicationMessage{
Payload: &pbpeering.ReplicationMessage_Response_{
Response: &pbpeering.ReplicationMessage_Response{
ResourceURL: pbpeering.TypeURLService,
// TODO(peering): Nonce management
Nonce: "",
ResourceID: serviceName,
Operation: pbpeering.ReplicationMessage_Response_UPSERT,
Resource: any,
},
},
}
return resp
}
func makeCARootsResponse(
logger hclog.Logger,
update cache.UpdateEvent,
) *pbpeering.ReplicationMessage {
any, _, err := marshalToProtoAny[*pbpeering.PeeringTrustBundle](update.Result)
if err != nil {
// Log the error and skip this response to avoid locking up peering due to a bad update event.
logger.Error("failed to marshal", "error", err)
return nil
}
resp := &pbpeering.ReplicationMessage{
Payload: &pbpeering.ReplicationMessage_Response_{
Response: &pbpeering.ReplicationMessage_Response{
ResourceURL: pbpeering.TypeURLRoots,
// TODO(peering): Nonce management
Nonce: "",
ResourceID: "roots",
Operation: pbpeering.ReplicationMessage_Response_UPSERT,
Resource: any,
},
},
}
return resp
}
// marshalToProtoAny takes any input and returns:
// the protobuf.Any type, the asserted T type, and any errors
// during marshalling or type assertion.
// `in` MUST be of type T or it returns an error.
func marshalToProtoAny[T proto.Message](in any) (*anypb.Any, T, error) {
typ, ok := in.(T)
if !ok {
var outType T
return nil, typ, fmt.Errorf("input type is not %T: %T", outType, in)
}
any, err := ptypes.MarshalAny(typ)
if err != nil {
return nil, typ, err
}
return any, typ, nil
}
func (s *Service) processResponse(
peerName string,
partition string,
resp *pbpeering.ReplicationMessage_Response,
) (*pbpeering.ReplicationMessage, error) {
if !pbpeering.KnownTypeURL(resp.ResourceURL) {
err := fmt.Errorf("received response for unknown resource type %q", resp.ResourceURL)
return makeReply(
resp.ResourceURL,
resp.Nonce,
code.Code_INVALID_ARGUMENT,
err.Error(),
), err
}
switch resp.Operation {
case pbpeering.ReplicationMessage_Response_UPSERT:
if resp.Resource == nil {
err := fmt.Errorf("received upsert response with no content")
return makeReply(
resp.ResourceURL,
resp.Nonce,
code.Code_INVALID_ARGUMENT,
err.Error(),
), err
}
if err := s.handleUpsert(peerName, partition, resp.ResourceURL, resp.ResourceID, resp.Resource); err != nil {
return makeReply(
resp.ResourceURL,
resp.Nonce,
code.Code_INTERNAL,
fmt.Sprintf("upsert error, ResourceURL: %q, ResourceID: %q: %v", resp.ResourceURL, resp.ResourceID, err),
), fmt.Errorf("upsert error: %w", err)
}
return makeReply(resp.ResourceURL, resp.Nonce, code.Code_OK, ""), nil
case pbpeering.ReplicationMessage_Response_DELETE:
if err := s.handleDelete(peerName, partition, resp.ResourceURL, resp.ResourceID); err != nil {
return makeReply(
resp.ResourceURL,
resp.Nonce,
code.Code_INTERNAL,
fmt.Sprintf("delete error, ResourceURL: %q, ResourceID: %q: %v", resp.ResourceURL, resp.ResourceID, err),
), fmt.Errorf("delete error: %w", err)
}
return makeReply(resp.ResourceURL, resp.Nonce, code.Code_OK, ""), nil
default:
var errMsg string
if op := pbpeering.ReplicationMessage_Response_Operation_name[int32(resp.Operation)]; op != "" {
errMsg = fmt.Sprintf("unsupported operation: %q", op)
} else {
errMsg = fmt.Sprintf("unsupported operation: %d", resp.Operation)
}
return makeReply(
resp.ResourceURL,
resp.Nonce,
code.Code_INVALID_ARGUMENT,
errMsg,
), errors.New(errMsg)
}
}
func (s *Service) handleUpsert(
peerName string,
partition string,
resourceURL string,
resourceID string,
resource *anypb.Any,
) error {
switch resourceURL {
case pbpeering.TypeURLService:
sn := structs.ServiceNameFromString(resourceID)
sn.OverridePartition(partition)
csn := &pbservice.IndexedCheckServiceNodes{}
if err := ptypes.UnmarshalAny(resource, csn); err != nil {
return fmt.Errorf("failed to unmarshal resource: %w", err)
}
return s.handleUpdateService(peerName, partition, sn, csn)
case pbpeering.TypeURLRoots:
roots := &pbpeering.PeeringTrustBundle{}
if err := ptypes.UnmarshalAny(resource, roots); err != nil {
return fmt.Errorf("failed to unmarshal resource: %w", err)
}
return s.handleUpsertRoots(peerName, partition, roots)
default:
return fmt.Errorf("unexpected resourceURL: %s", resourceURL)
}
}
// handleUpdateService handles both deletion and upsert events for a service.
// On an UPSERT event:
// - All nodes, services, checks in the input pbNodes are re-applied through Raft.
// - Any nodes, services, or checks in the catalog that were not in the input pbNodes get deleted.
//
// On a DELETE event:
// - A reconciliation against nil or empty input pbNodes leads to deleting all stored catalog resources
// associated with the service name.
func (s *Service) handleUpdateService(
peerName string,
partition string,
sn structs.ServiceName,
pbNodes *pbservice.IndexedCheckServiceNodes,
) error {
// Capture instances in the state store for reconciliation later.
_, storedInstances, err := s.Backend.Store().CheckServiceNodes(nil, sn.Name, &sn.EnterpriseMeta, peerName)
if err != nil {
return fmt.Errorf("failed to read imported services: %w", err)
}
structsNodes, err := pbNodes.CheckServiceNodesToStruct()
if err != nil {
return fmt.Errorf("failed to convert protobuf instances to structs: %w", err)
}
// Normalize the data into a convenient form for operation.
snap := newHealthSnapshot(structsNodes, partition, peerName)
for _, nodeSnap := range snap.Nodes {
// First register the node
req := nodeSnap.Node.ToRegisterRequest()
if err := s.Backend.Apply().CatalogRegister(&req); err != nil {
return fmt.Errorf("failed to register node: %w", err)
}
// Then register all services on that node
for _, svcSnap := range nodeSnap.Services {
req.Service = svcSnap.Service
if err := s.Backend.Apply().CatalogRegister(&req); err != nil {
return fmt.Errorf("failed to register service: %w", err)
}
}
req.Service = nil
// Then register all checks on that node
var chks structs.HealthChecks
for _, svcSnap := range nodeSnap.Services {
for _, c := range svcSnap.Checks {
chks = append(chks, c)
}
}
req.Checks = chks
if err := s.Backend.Apply().CatalogRegister(&req); err != nil {
return fmt.Errorf("failed to register check: %w", err)
}
}
//
// Now that the data received has been stored in the state store, the rest of this
// function is responsible for cleaning up data in the catalog that wasn't in the snapshot.
//
// nodeCheckTuple uniquely identifies a node check in the catalog.
// The partition is not needed because we are only operating on one partition's catalog.
type nodeCheckTuple struct {
checkID types.CheckID
node string
}
var (
// unusedNodes tracks node names that were not present in the latest response.
// Missing nodes are not assumed to be deleted because there may be other service names
// registered on them.
// Inside we also track a map of node checks associated with the node.
unusedNodes = make(map[string]struct{})
// deletedNodeChecks tracks node checks that were not present in the latest response.
// A single node check will be attached to all service instances of a node, so this
// deduplication prevents issuing multiple deregistrations for a single check.
deletedNodeChecks = make(map[nodeCheckTuple]struct{})
)
for _, csn := range storedInstances {
if _, ok := snap.Nodes[csn.Node.ID]; !ok {
unusedNodes[string(csn.Node.ID)] = struct{}{}
// Since the node is not in the snapshot we can know the associated service
// instance is not in the snapshot either, since a service instance can't
// exist without a node.
// This will also delete all service checks.
err := s.Backend.Apply().CatalogDeregister(&structs.DeregisterRequest{
Node: csn.Node.Node,
ServiceID: csn.Service.ID,
EnterpriseMeta: csn.Service.EnterpriseMeta,
PeerName: peerName,
})
if err != nil {
return fmt.Errorf("failed to deregister service %q: %w", csn.Service.CompoundServiceID(), err)
}
// We can't know if a node check was deleted from the exporting cluster
// (but not the node itself) if the node wasn't in the snapshot,
// so we do not loop over checks here.
// If the unusedNode gets deleted below that will also delete node checks.
continue
}
// Delete the service instance if not in the snapshot.
sid := csn.Service.CompoundServiceID()
if _, ok := snap.Nodes[csn.Node.ID].Services[sid]; !ok {
err := s.Backend.Apply().CatalogDeregister(&structs.DeregisterRequest{
Node: csn.Node.Node,
ServiceID: csn.Service.ID,
EnterpriseMeta: csn.Service.EnterpriseMeta,
PeerName: peerName,
})
if err != nil {
ident := fmt.Sprintf("partition:%s/peer:%s/node:%s/ns:%s/service_id:%s",
csn.Service.PartitionOrDefault(), peerName, csn.Node.Node, csn.Service.NamespaceOrDefault(), csn.Service.ID)
return fmt.Errorf("failed to deregister service %q: %w", ident, err)
}
// When a service is deleted all associated checks also get deleted as a side effect.
continue
}
// Reconcile checks.
for _, chk := range csn.Checks {
if _, ok := snap.Nodes[csn.Node.ID].Services[sid].Checks[chk.CheckID]; !ok {
// Checks without a ServiceID are node checks.
// If the node exists but the check does not then the check was deleted.
if chk.ServiceID == "" {
// Deduplicate node checks to avoid deregistering a check multiple times.
tuple := nodeCheckTuple{
checkID: chk.CheckID,
node: chk.Node,
}
deletedNodeChecks[tuple] = struct{}{}
continue
}
// If the check isn't a node check then it's a service check.
// Service checks that were not present can be deleted immediately because
// checks for a given service ID will only be attached to a single CheckServiceNode.
err := s.Backend.Apply().CatalogDeregister(&structs.DeregisterRequest{
Node: chk.Node,
CheckID: chk.CheckID,
EnterpriseMeta: chk.EnterpriseMeta,
PeerName: peerName,
})
if err != nil {
ident := fmt.Sprintf("partition:%s/peer:%s/node:%s/ns:%s/check_id:%s",
chk.PartitionOrDefault(), peerName, chk.Node, chk.NamespaceOrDefault(), chk.CheckID)
return fmt.Errorf("failed to deregister check %q: %w", ident, err)
}
}
}
}
// Delete all deduplicated node checks.
for chk := range deletedNodeChecks {
nodeMeta := structs.NodeEnterpriseMetaInPartition(sn.PartitionOrDefault())
err := s.Backend.Apply().CatalogDeregister(&structs.DeregisterRequest{
Node: chk.node,
CheckID: chk.checkID,
EnterpriseMeta: *nodeMeta,
PeerName: peerName,
})
if err != nil {
ident := fmt.Sprintf("partition:%s/peer:%s/node:%s/check_id:%s", nodeMeta.PartitionOrDefault(), peerName, chk.node, chk.checkID)
return fmt.Errorf("failed to deregister node check %q: %w", ident, err)
}
}
// Delete any nodes that do not have any other services registered on them.
for node := range unusedNodes {
nodeMeta := structs.NodeEnterpriseMetaInPartition(sn.PartitionOrDefault())
_, ns, err := s.Backend.Store().NodeServices(nil, node, nodeMeta, peerName)
if err != nil {
return fmt.Errorf("failed to query services on node: %w", err)
}
if ns != nil && len(ns.Services) >= 1 {
// At least one service is still registered on this node, so we keep it.
continue
}
// All services on the node were deleted, so the node is also cleaned up.
err = s.Backend.Apply().CatalogDeregister(&structs.DeregisterRequest{
Node: node,
PeerName: peerName,
EnterpriseMeta: *nodeMeta,
})
if err != nil {
ident := fmt.Sprintf("partition:%s/peer:%s/node:%s", nodeMeta.PartitionOrDefault(), peerName, node)
return fmt.Errorf("failed to deregister node %q: %w", ident, err)
}
}
return nil
}
func (s *Service) handleUpsertRoots(
peerName string,
partition string,
trustBundle *pbpeering.PeeringTrustBundle,
) error {
// We override the partition and peer name so that the trust bundle gets stored
// in the importing partition with a reference to the peer it was imported from.
trustBundle.Partition = partition
trustBundle.PeerName = peerName
req := &pbpeering.PeeringTrustBundleWriteRequest{
PeeringTrustBundle: trustBundle,
}
return s.Backend.Apply().PeeringTrustBundleWrite(req)
}
func (s *Service) handleDelete(
peerName string,
partition string,
resourceURL string,
resourceID string,
) error {
switch resourceURL {
case pbpeering.TypeURLService:
sn := structs.ServiceNameFromString(resourceID)
sn.OverridePartition(partition)
return s.handleUpdateService(peerName, partition, sn, nil)
default:
return fmt.Errorf("unexpected resourceURL: %s", resourceURL)
}
}
func makeReply(resourceURL, nonce string, errCode code.Code, errMsg string) *pbpeering.ReplicationMessage {
var rpcErr *pbstatus.Status
if errCode != code.Code_OK || errMsg != "" {
rpcErr = &pbstatus.Status{
Code: int32(errCode),
Message: errMsg,
}
}
// TODO: shouldn't this be response?
return &pbpeering.ReplicationMessage{
Payload: &pbpeering.ReplicationMessage_Request_{
Request: &pbpeering.ReplicationMessage_Request{
ResourceURL: resourceURL,
Nonce: nonce,
Error: rpcErr,
},
},
}
}