grpc 1.60.0 → 1.62.0

data/src/core/load_balancing/xds/cds.cc

@@ -0,0 +1,757 @@
+//
+// Copyright 2019 gRPC authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+
+#include <grpc/support/port_platform.h>
+
+#include <algorithm>
+#include <map>
+#include <memory>
+#include <set>
+#include <string>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include "absl/status/status.h"
+#include "absl/status/statusor.h"
+#include "absl/strings/str_cat.h"
+#include "absl/strings/string_view.h"
+#include "absl/types/optional.h"
+#include "absl/types/variant.h"
+
+#include <grpc/grpc_security.h>
+#include <grpc/impl/connectivity_state.h>
+#include <grpc/support/json.h>
+#include <grpc/support/log.h>
+
+#include "src/core/load_balancing/address_filtering.h"
+#include "src/core/load_balancing/outlier_detection/outlier_detection.h"
+#include "src/core/load_balancing/xds/xds_channel_args.h"
+#include "src/core/ext/xds/xds_cluster.h"
+#include "src/core/ext/xds/xds_common_types.h"
+#include "src/core/ext/xds/xds_health_status.h"
+#include "src/core/lib/channel/channel_args.h"
+#include "src/core/lib/config/core_configuration.h"
+#include "src/core/lib/debug/trace.h"
+#include "src/core/lib/gprpp/debug_location.h"
+#include "src/core/lib/gprpp/env.h"
+#include "src/core/lib/gprpp/match.h"
+#include "src/core/lib/gprpp/orphanable.h"
+#include "src/core/lib/gprpp/ref_counted_ptr.h"
+#include "src/core/lib/gprpp/time.h"
+#include "src/core/lib/gprpp/unique_type_name.h"
+#include "src/core/lib/gprpp/work_serializer.h"
+#include "src/core/lib/iomgr/pollset_set.h"
+#include "src/core/lib/json/json.h"
+#include "src/core/lib/json/json_args.h"
+#include "src/core/lib/json/json_object_loader.h"
+#include "src/core/lib/json/json_writer.h"
+#include "src/core/load_balancing/delegating_helper.h"
+#include "src/core/load_balancing/lb_policy.h"
+#include "src/core/load_balancing/lb_policy_factory.h"
+#include "src/core/load_balancing/lb_policy_registry.h"
+#include "src/core/resolver/xds/xds_dependency_manager.h"
+
+namespace grpc_core {
+
+TraceFlag grpc_cds_lb_trace(false, "cds_lb");
+
+namespace {
+
+// TODO(roth): Remove this after the 1.63 release.
+bool XdsAggregateClusterBackwardCompatibilityEnabled() {
+  auto value = GetEnv("GRPC_XDS_AGGREGATE_CLUSTER_BACKWARD_COMPAT");
+  if (!value.has_value()) return false;
+  bool parsed_value;
+  bool parse_succeeded = gpr_parse_bool_value(value->c_str(), &parsed_value);
+  return parse_succeeded && parsed_value;
+}
+
+using XdsConfig = XdsDependencyManager::XdsConfig;
+
+constexpr absl::string_view kCds = "cds_experimental";
+
+// Config for this LB policy.
+class CdsLbConfig : public LoadBalancingPolicy::Config {
+ public:
+  CdsLbConfig() = default;
+
+  CdsLbConfig(const CdsLbConfig&) = delete;
+  CdsLbConfig& operator=(const CdsLbConfig&) = delete;
+
+  CdsLbConfig(CdsLbConfig&& other) = delete;
+  CdsLbConfig& operator=(CdsLbConfig&& other) = delete;
+
+  absl::string_view name() const override { return kCds; }
+
+  const std::string& cluster() const { return cluster_; }
+  bool is_dynamic() const { return is_dynamic_; }
+
+  static const JsonLoaderInterface* JsonLoader(const JsonArgs&) {
+    static const auto* loader =
+        JsonObjectLoader<CdsLbConfig>()
+            .Field("cluster", &CdsLbConfig::cluster_)
+            .OptionalField("isDynamic", &CdsLbConfig::is_dynamic_)
+            .Finish();
+    return loader;
+  }
+
+ private:
+  std::string cluster_;
+  bool is_dynamic_ = false;
+};
+
+// CDS LB policy.
+class CdsLb : public LoadBalancingPolicy {
+ public:
+  explicit CdsLb(Args args);
+
+  absl::string_view name() const override { return kCds; }
+
+  absl::Status UpdateLocked(UpdateArgs args) override;
+  void ResetBackoffLocked() override;
+  void ExitIdleLocked() override;
+
+ private:
+  // Delegating helper to be passed to child policy.
+  using Helper = ParentOwningDelegatingChannelControlHelper<CdsLb>;
+
+  // State used to retain child policy names for the priority policy.
+  struct ChildNameState {
+    std::vector<size_t /*child_number*/> priority_child_numbers;
+    size_t next_available_child_number = 0;
+
+    void Reset() {
+      priority_child_numbers.clear();
+      next_available_child_number = 0;
+    }
+  };
+
+  ~CdsLb() override;
+
+  void ShutdownLocked() override;
+
+  // Computes child numbers for new_cluster, reusing child numbers
+  // from old_cluster and child_name_state_ in an intelligent
+  // way to avoid unnecessary churn.
+  ChildNameState ComputeChildNames(
+      const XdsConfig::ClusterConfig* old_cluster,
+      const XdsConfig::ClusterConfig& new_cluster,
+      const XdsConfig::ClusterConfig::EndpointConfig& endpoint_config) const;
+
+  std::string GetChildPolicyName(const std::string& cluster, size_t priority);
+
+  Json CreateChildPolicyConfigForLeafCluster(
+      const XdsConfig::ClusterConfig& new_cluster,
+      const XdsConfig::ClusterConfig::EndpointConfig& endpoint_config,
+      const XdsClusterResource* aggregate_cluster_resource);
+  Json CreateChildPolicyConfigForAggregateCluster(
+      const XdsConfig::ClusterConfig::AggregateConfig& aggregate_config);
+
+  void ResetState();
+
+  void ReportTransientFailure(absl::Status status);
+
+  std::string cluster_name_;
+  RefCountedPtr<const XdsConfig> xds_config_;
+
+  // Cluster subscription, for dynamic clusters (e.g., RLS).
+  RefCountedPtr<XdsDependencyManager::ClusterSubscription> subscription_;
+
+  ChildNameState child_name_state_;
+
+  // Child LB policy.
+  OrphanablePtr<LoadBalancingPolicy> child_policy_;
+
+  // Internal state.
+  bool shutting_down_ = false;
+};
+
+//
+// CdsLb
+//
+
+CdsLb::CdsLb(Args args) : LoadBalancingPolicy(std::move(args)) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+    gpr_log(GPR_INFO, "[cdslb %p] created", this);
+  }
+}
+
+CdsLb::~CdsLb() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+    gpr_log(GPR_INFO, "[cdslb %p] destroying cds LB policy", this);
+  }
+}
+
+void CdsLb::ShutdownLocked() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+    gpr_log(GPR_INFO, "[cdslb %p] shutting down", this);
+  }
+  shutting_down_ = true;
+  ResetState();
+}
+
+void CdsLb::ResetBackoffLocked() {
+  if (child_policy_ != nullptr) child_policy_->ResetBackoffLocked();
+}
+
+void CdsLb::ExitIdleLocked() {
+  if (child_policy_ != nullptr) child_policy_->ExitIdleLocked();
+}
+
+// We need at least one priority for each discovery mechanism, just so that we
+// have a child in which to create the xds_cluster_impl policy.  This ensures
+// that we properly handle the case of a discovery mechanism dropping 100% of
+// calls, the OnError() case, and the OnResourceDoesNotExist() case.
+const XdsEndpointResource::PriorityList& GetUpdatePriorityList(
+    const XdsEndpointResource* update) {
+  static const NoDestruct<XdsEndpointResource::PriorityList>
+      kPriorityListWithEmptyPriority(1);
+  if (update == nullptr || update->priorities.empty()) {
+    return *kPriorityListWithEmptyPriority;
+  }
+  return update->priorities;
+}
+
+std::string MakeChildPolicyName(absl::string_view cluster,
+                                size_t child_number) {
+  return absl::StrCat("{cluster=", cluster, ", child_number=", child_number,
+                      "}");
+}
+
+class PriorityEndpointIterator : public EndpointAddressesIterator {
+ public:
+  PriorityEndpointIterator(
+      std::string cluster_name,
+      std::shared_ptr<const XdsEndpointResource> endpoints,
+      std::vector<size_t /*child_number*/> priority_child_numbers)
+      : cluster_name_(std::move(cluster_name)),
+        endpoints_(std::move(endpoints)),
+        priority_child_numbers_(std::move(priority_child_numbers)) {}
+
+  void ForEach(absl::FunctionRef<void(const EndpointAddresses&)> callback)
+      const override {
+    const auto& priority_list = GetUpdatePriorityList(endpoints_.get());
+    for (size_t priority = 0; priority < priority_list.size(); ++priority) {
+      const auto& priority_entry = priority_list[priority];
+      std::string priority_child_name =
+          MakeChildPolicyName(cluster_name_, priority_child_numbers_[priority]);
+      for (const auto& p : priority_entry.localities) {
+        const auto& locality_name = p.first;
+        const auto& locality = p.second;
+        std::vector<RefCountedStringValue> hierarchical_path = {
+            RefCountedStringValue(priority_child_name),
+            RefCountedStringValue(locality_name->AsHumanReadableString())};
+        auto hierarchical_path_attr =
+            MakeRefCounted<HierarchicalPathArg>(std::move(hierarchical_path));
+        for (const auto& endpoint : locality.endpoints) {
+          uint32_t endpoint_weight =
+              locality.lb_weight *
+              endpoint.args().GetInt(GRPC_ARG_ADDRESS_WEIGHT).value_or(1);
+          callback(EndpointAddresses(
+              endpoint.addresses(),
+              endpoint.args()
+                  .SetObject(hierarchical_path_attr)
+                  .Set(GRPC_ARG_ADDRESS_WEIGHT, endpoint_weight)
+                  .SetObject(locality_name->Ref())
+                  .Set(GRPC_ARG_XDS_LOCALITY_WEIGHT, locality.lb_weight)));
+        }
+      }
+    }
+  }
+
+ private:
+  std::string cluster_name_;
+  std::shared_ptr<const XdsEndpointResource> endpoints_;
+  std::vector<size_t /*child_number*/> priority_child_numbers_;
+};
+
+absl::Status CdsLb::UpdateLocked(UpdateArgs args) {
+  // Get new config.
+  auto new_config = args.config.TakeAsSubclass<CdsLbConfig>();
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+    gpr_log(GPR_INFO, "[cdslb %p] received update: cluster=%s is_dynamic=%d",
+            this, new_config->cluster().c_str(), new_config->is_dynamic());
+  }
+  GPR_ASSERT(new_config != nullptr);
+  // Cluster name should never change, because we should use a different
+  // child name in xds_cluster_manager in that case.
+  if (cluster_name_.empty()) {
+    cluster_name_ = new_config->cluster();
+  } else {
+    GPR_ASSERT(cluster_name_ == new_config->cluster());
+  }
+  // Start dynamic subscription if needed.
+  if (new_config->is_dynamic() && subscription_ == nullptr) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+      gpr_log(GPR_INFO,
+              "[cdslb %p] obtaining dynamic subscription for cluster %s", this,
+              cluster_name_.c_str());
+    }
+    auto* dependency_mgr = args.args.GetObject<XdsDependencyManager>();
+    if (dependency_mgr == nullptr) {
+      // Should never happen.
+      absl::Status status =
+          absl::InternalError("xDS dependency mgr not passed to CDS LB policy");
+      ReportTransientFailure(status);
+      return status;
+    }
+    subscription_ = dependency_mgr->GetClusterSubscription(cluster_name_);
+  }
+  // Get xDS config.
+  auto new_xds_config = args.args.GetObjectRef<XdsConfig>();
+  if (new_xds_config == nullptr) {
+    // Should never happen.
+    absl::Status status =
+        absl::InternalError("xDS config not passed to CDS LB policy");
+    ReportTransientFailure(status);
+    return status;
+  }
+  auto it = new_xds_config->clusters.find(cluster_name_);
+  if (it == new_xds_config->clusters.end()) {
+    // Cluster not present.
+    if (new_config->is_dynamic()) {
+      // If we are already subscribed, it's possible that we just
+      // recently subscribed but another update came through before we
+      // got the new cluster, in which case it will still be missing.
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+        gpr_log(GPR_INFO,
+                "[cdslb %p] xDS config has no entry for dynamic cluster %s, "
+                "waiting for subsequent update",
+                this, cluster_name_.c_str());
+      }
+      // Stay in CONNECTING until we get an update that has the cluster.
+      return absl::OkStatus();
+    }
+    // Not a dynamic cluster.  This should never happen.
+    absl::Status status = absl::UnavailableError(absl::StrCat(
+        "xDS config has no entry for static cluster ", cluster_name_));
+    ReportTransientFailure(status);
+    return status;
+  }
+  auto& new_cluster_config = it->second;
+  // If new list is not OK, report TRANSIENT_FAILURE.
+  if (!new_cluster_config.ok()) {
+    ReportTransientFailure(new_cluster_config.status());
+    return new_cluster_config.status();
+  }
+  GPR_ASSERT(new_cluster_config->cluster != nullptr);
+  // Find old cluster, if any.
+  const XdsConfig::ClusterConfig* old_cluster_config = nullptr;
+  if (xds_config_ != nullptr) {
+    auto it_old = xds_config_->clusters.find(cluster_name_);
+    if (it_old != xds_config_->clusters.end() && it_old->second.ok()) {
+      old_cluster_config = &*it_old->second;
+      // If nothing changed for a leaf cluster, then ignore the update.
+      // Can't do this for an aggregate cluster, because even if the aggregate
+      // cluster itself didn't change, the leaf clusters may have changed.
+      if (*new_cluster_config == *old_cluster_config &&
+          absl::holds_alternative<XdsConfig::ClusterConfig::EndpointConfig>(
+              new_cluster_config->children)) {
+        return absl::OkStatus();
+      }
+    }
+  }
+  // TODO(roth): Remove this after the 1.63 release.
+  const XdsClusterResource* aggregate_cluster_resource = nullptr;
+  static constexpr absl::string_view kArgXdsAggregateClusterName =
+      GRPC_ARG_NO_SUBCHANNEL_PREFIX "xds_aggregate_cluster_name";
+  if (XdsAggregateClusterBackwardCompatibilityEnabled()) {
+    if (absl::holds_alternative<XdsConfig::ClusterConfig::EndpointConfig>(
+            new_cluster_config->children)) {
+      auto aggregate_cluster = args.args.GetString(kArgXdsAggregateClusterName);
+      if (aggregate_cluster.has_value()) {
+        auto it = new_xds_config->clusters.find(*aggregate_cluster);
+        if (it == new_xds_config->clusters.end()) {
+          // Cluster not present.  This should never happen.
+          absl::Status status = absl::UnavailableError(
+              absl::StrCat("xDS config has no entry for aggregate cluster ",
+                           *aggregate_cluster));
+          ReportTransientFailure(status);
+          return status;
+        }
+        auto& aggregate_cluster_config = it->second;
+        if (!aggregate_cluster_config.ok()) {
+          ReportTransientFailure(aggregate_cluster_config.status());
+          return aggregate_cluster_config.status();
+        }
+        GPR_ASSERT(aggregate_cluster_config->cluster != nullptr);
+        aggregate_cluster_resource = aggregate_cluster_config->cluster.get();
+      }
+    } else {
+      args.args = args.args.Set(kArgXdsAggregateClusterName, cluster_name_);
+    }
+  }
+  // Construct child policy config and update state based on the cluster type.
+  Json child_policy_config_json;
+  UpdateArgs update_args;
+  Match(
+      new_cluster_config->children,
+      // Leaf cluster.
+      [&](const XdsConfig::ClusterConfig::EndpointConfig& endpoint_config) {
+        // Compute new child numbers.
+        child_name_state_ = ComputeChildNames(
+            old_cluster_config, *new_cluster_config, endpoint_config);
+        // Populate addresses and resolution_note for child policy.
+        update_args.addresses = std::make_shared<PriorityEndpointIterator>(
+            cluster_name_, endpoint_config.endpoints,
+            child_name_state_.priority_child_numbers);
+        update_args.resolution_note = endpoint_config.resolution_note;
+        // Construct child policy config.
+        child_policy_config_json = CreateChildPolicyConfigForLeafCluster(
+            *new_cluster_config, endpoint_config, aggregate_cluster_resource);
+      },
+      // Aggregate cluster.
+      [&](const XdsConfig::ClusterConfig::AggregateConfig& aggregate_config) {
+        child_name_state_.Reset();
+        // Construct child policy config.
+        child_policy_config_json =
+            CreateChildPolicyConfigForAggregateCluster(aggregate_config);
+      });
+  // Swap in new xDS config, now that we're done with the old one.
+  xds_config_ = std::move(new_xds_config);
+  // Validate child policy config.
+  auto child_config =
+      CoreConfiguration::Get().lb_policy_registry().ParseLoadBalancingConfig(
+          child_policy_config_json);
+  if (!child_config.ok()) {
+    // Should never happen.
+    absl::Status status = absl::InternalError(
+        absl::StrCat(cluster_name_, ": error parsing child policy config: ",
+                     child_config.status().message()));
+    ReportTransientFailure(status);
+    return status;
+  }
+  // Create child policy if not already present.
+  if (child_policy_ == nullptr) {
+    LoadBalancingPolicy::Args lb_args;
+    lb_args.work_serializer = work_serializer();
+    lb_args.args = args.args;
+    lb_args.channel_control_helper =
+        std::make_unique<Helper>(RefAsSubclass<CdsLb>());
+    child_policy_ =
+        CoreConfiguration::Get().lb_policy_registry().CreateLoadBalancingPolicy(
+            (*child_config)->name(), std::move(lb_args));
+    if (child_policy_ == nullptr) {
+      // Should never happen.
+      absl::Status status = absl::UnavailableError(
+          absl::StrCat(cluster_name_, ": failed to create child policy"));
+      ReportTransientFailure(status);
+      return status;
+    }
+    grpc_pollset_set_add_pollset_set(child_policy_->interested_parties(),
+                                     interested_parties());
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+      gpr_log(GPR_INFO, "[cdslb %p] created child policy %s (%p)", this,
+              std::string((*child_config)->name()).c_str(),
+              child_policy_.get());
+    }
+  }
+  // Update child policy.
+  update_args.config = std::move(*child_config);
+  update_args.args = args.args;
+  return child_policy_->UpdateLocked(std::move(update_args));
+}
+
+CdsLb::ChildNameState CdsLb::ComputeChildNames(
+    const XdsConfig::ClusterConfig* old_cluster,
+    const XdsConfig::ClusterConfig& new_cluster,
+    const XdsConfig::ClusterConfig::EndpointConfig& endpoint_config) const {
+  GPR_ASSERT(
+      !absl::holds_alternative<XdsConfig::ClusterConfig::AggregateConfig>(
+          new_cluster.children));
+  // First, build some maps from locality to child number and the reverse
+  // from old_cluster and child_name_state_.
+  std::map<XdsLocalityName*, size_t /*child_number*/, XdsLocalityName::Less>
+      locality_child_map;
+  std::map<size_t, std::set<XdsLocalityName*, XdsLocalityName::Less>>
+      child_locality_map;
+  if (old_cluster != nullptr) {
+    auto* old_endpoint_config =
+        absl::get_if<XdsConfig::ClusterConfig::EndpointConfig>(
+            &old_cluster->children);
+    if (old_endpoint_config != nullptr) {
+      const auto& prev_priority_list =
+          GetUpdatePriorityList(old_endpoint_config->endpoints.get());
+      for (size_t priority = 0; priority < prev_priority_list.size();
+           ++priority) {
+        size_t child_number =
+            child_name_state_.priority_child_numbers[priority];
+        const auto& localities = prev_priority_list[priority].localities;
+        for (const auto& p : localities) {
+          XdsLocalityName* locality_name = p.first;
+          locality_child_map[locality_name] = child_number;
+          child_locality_map[child_number].insert(locality_name);
+        }
+      }
+    }
+  }
+  // Now construct new state containing priority child numbers for the new
+  // cluster based on the maps constructed above.
+  ChildNameState new_child_name_state;
+  new_child_name_state.next_available_child_number =
+      child_name_state_.next_available_child_number;
+  const XdsEndpointResource::PriorityList& priority_list =
+      GetUpdatePriorityList(endpoint_config.endpoints.get());
+  for (size_t priority = 0; priority < priority_list.size(); ++priority) {
+    const auto& localities = priority_list[priority].localities;
+    absl::optional<size_t> child_number;
+    // If one of the localities in this priority already existed, reuse its
+    // child number.
+    for (const auto& p : localities) {
+      XdsLocalityName* locality_name = p.first;
+      if (!child_number.has_value()) {
+        auto it = locality_child_map.find(locality_name);
+        if (it != locality_child_map.end()) {
+          child_number = it->second;
+          locality_child_map.erase(it);
+          // Remove localities that *used* to be in this child number, so
+          // that we don't incorrectly reuse this child number for a
+          // subsequent priority.
+          for (XdsLocalityName* old_locality :
+               child_locality_map[*child_number]) {
+            locality_child_map.erase(old_locality);
+          }
+        }
+      } else {
+        // Remove all localities that are now in this child number, so
+        // that we don't accidentally reuse this child number for a
+        // subsequent priority.
+        locality_child_map.erase(locality_name);
+      }
+    }
+    // If we didn't find an existing child number, assign a new one.
+    if (!child_number.has_value()) {
+      for (child_number = new_child_name_state.next_available_child_number;
+           child_locality_map.find(*child_number) != child_locality_map.end();
+           ++(*child_number)) {
+      }
+      new_child_name_state.next_available_child_number = *child_number + 1;
+      // Add entry so we know that the child number is in use.
+      // (Don't need to add the list of localities, since we won't use them.)
+      child_locality_map[*child_number];
+    }
+    new_child_name_state.priority_child_numbers.push_back(*child_number);
+  }
+  return new_child_name_state;
+}
+
+Json CdsLb::CreateChildPolicyConfigForLeafCluster(
+    const XdsConfig::ClusterConfig& new_cluster,
+    const XdsConfig::ClusterConfig::EndpointConfig& endpoint_config,
+    const XdsClusterResource* aggregate_cluster_resource) {
+  const auto& cluster_resource = *new_cluster.cluster;
+  const bool is_logical_dns =
+      absl::holds_alternative<XdsClusterResource::LogicalDns>(
+          cluster_resource.type);
+  // Determine what xDS LB policy to use.
+  Json xds_lb_policy;
+  if (is_logical_dns) {
+    xds_lb_policy = Json::FromArray({
+        Json::FromObject({
+            {"pick_first", Json::FromObject({})},
+        }),
+    });
+  }
+  // TODO(roth): Remove this "else if" block after the 1.63 release.
+  else if (XdsAggregateClusterBackwardCompatibilityEnabled() &&
+           aggregate_cluster_resource != nullptr) {
+    xds_lb_policy =
+        Json::FromArray(aggregate_cluster_resource->lb_policy_config);
+  } else {
+    xds_lb_policy = Json::FromArray(new_cluster.cluster->lb_policy_config);
+  }
+  // Wrap it in the priority policy.
+  Json::Object priority_children;
+  Json::Array priority_priorities;
+  const auto& priority_list =
+      GetUpdatePriorityList(endpoint_config.endpoints.get());
+  for (size_t priority = 0; priority < priority_list.size(); ++priority) {
+    // Add priority entry, with the appropriate child name.
+    std::string child_name = MakeChildPolicyName(
+        cluster_name_, child_name_state_.priority_child_numbers[priority]);
+    priority_priorities.emplace_back(Json::FromString(child_name));
+    Json::Object child_config = {{"config", xds_lb_policy}};
+    if (!is_logical_dns) {
+      child_config["ignore_reresolution_requests"] = Json::FromBool(true);
+    }
+    priority_children[child_name] = Json::FromObject(std::move(child_config));
+  }
+  Json priority_policy = Json::FromArray({Json::FromObject({
+      {"priority_experimental",
+       Json::FromObject({
+           {"children", Json::FromObject(std::move(priority_children))},
+           {"priorities", Json::FromArray(std::move(priority_priorities))},
+       })},
+  })});
+  // Wrap the priority policy in the xds_override_host policy.
+  Json xds_override_host_policy = Json::FromArray({Json::FromObject({
+      {"xds_override_host_experimental",
+       Json::FromObject({
+           {"clusterName", Json::FromString(cluster_name_)},
+           {"childPolicy", std::move(priority_policy)},
+       })},
+  })});
+  // Wrap the xds_override_host policy in the xds_cluster_impl policy.
+  Json xds_cluster_impl_policy = Json::FromArray({Json::FromObject({
+      {"xds_cluster_impl_experimental",
+       Json::FromObject({
+           {"clusterName", Json::FromString(cluster_name_)},
+           {"childPolicy", std::move(xds_override_host_policy)},
+       })},
+  })});
+  // Wrap the xds_cluster_impl policy in the outlier_detection policy.
+  Json::Object outlier_detection_config = {
+      {"childPolicy", std::move(xds_cluster_impl_policy)},
+  };
+  if (cluster_resource.outlier_detection.has_value()) {
+    auto& outlier_detection_update = *cluster_resource.outlier_detection;
+    outlier_detection_config["interval"] =
+        Json::FromString(outlier_detection_update.interval.ToJsonString());
+    outlier_detection_config["baseEjectionTime"] = Json::FromString(
+        outlier_detection_update.base_ejection_time.ToJsonString());
+    outlier_detection_config["maxEjectionTime"] = Json::FromString(
+        outlier_detection_update.max_ejection_time.ToJsonString());
+    outlier_detection_config["maxEjectionPercent"] =
+        Json::FromNumber(outlier_detection_update.max_ejection_percent);
+    if (outlier_detection_update.success_rate_ejection.has_value()) {
+      outlier_detection_config["successRateEjection"] = Json::FromObject({
+          {"stdevFactor",
+           Json::FromNumber(
+               outlier_detection_update.success_rate_ejection->stdev_factor)},
+          {"enforcementPercentage",
+           Json::FromNumber(outlier_detection_update.success_rate_ejection
+                                ->enforcement_percentage)},
+          {"minimumHosts",
+           Json::FromNumber(
+               outlier_detection_update.success_rate_ejection->minimum_hosts)},
+          {"requestVolume",
+           Json::FromNumber(
+               outlier_detection_update.success_rate_ejection->request_volume)},
+      });
+    }
+    if (outlier_detection_update.failure_percentage_ejection.has_value()) {
+      outlier_detection_config["failurePercentageEjection"] = Json::FromObject({
+          {"threshold",
+           Json::FromNumber(outlier_detection_update
+                                .failure_percentage_ejection->threshold)},
+          {"enforcementPercentage",
+           Json::FromNumber(
+               outlier_detection_update.failure_percentage_ejection
+                   ->enforcement_percentage)},
+          {"minimumHosts",
+           Json::FromNumber(outlier_detection_update
+                                .failure_percentage_ejection->minimum_hosts)},
+          {"requestVolume",
+           Json::FromNumber(outlier_detection_update
+                                .failure_percentage_ejection->request_volume)},
+      });
+    }
+  }
+  Json outlier_detection_policy = Json::FromArray({Json::FromObject({
+      {"outlier_detection_experimental",
+       Json::FromObject(std::move(outlier_detection_config))},
+  })});
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+    gpr_log(GPR_INFO, "[cdslb %p] generated config for child policy: %s", this,
+            JsonDump(outlier_detection_policy, /*indent=*/1).c_str());
+  }
+  return outlier_detection_policy;
+}
+
+Json CdsLb::CreateChildPolicyConfigForAggregateCluster(
+    const XdsConfig::ClusterConfig::AggregateConfig& aggregate_config) {
+  Json::Object priority_children;
+  Json::Array priority_priorities;
+  for (const absl::string_view& leaf_cluster : aggregate_config.leaf_clusters) {
+    priority_children[std::string(leaf_cluster)] = Json::FromObject({
+        {"config",
+         Json::FromArray({
+             Json::FromObject({
+                 {"cds_experimental",
+                  Json::FromObject({
+                      {"cluster", Json::FromString(std::string(leaf_cluster))},
+                  })},
+             }),
+         })},
+    });
+    priority_priorities.emplace_back(
+        Json::FromString(std::string(leaf_cluster)));
+  }
+  Json json = Json::FromArray({Json::FromObject({
+      {"priority_experimental",
+       Json::FromObject({
+           {"children", Json::FromObject(std::move(priority_children))},
+           {"priorities", Json::FromArray(std::move(priority_priorities))},
+       })},
+  })});
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+    gpr_log(GPR_INFO, "[cdslb %p] generated config for child policy: %s", this,
+            JsonDump(json, /*indent=*/1).c_str());
+  }
+  return json;
+}
+
+void CdsLb::ResetState() {
+  cluster_name_.clear();
+  xds_config_.reset();
+  child_name_state_.Reset();
+  if (child_policy_ != nullptr) {
+    grpc_pollset_set_del_pollset_set(child_policy_->interested_parties(),
+                                     interested_parties());
+    child_policy_.reset();
+  }
+}
+
+void CdsLb::ReportTransientFailure(absl::Status status) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_cds_lb_trace)) {
+    gpr_log(GPR_INFO, "[cdslb %p] reporting TRANSIENT_FAILURE: %s", this,
+            status.ToString().c_str());
+  }
+  ResetState();
+  channel_control_helper()->UpdateState(
+      GRPC_CHANNEL_TRANSIENT_FAILURE, status,
+      MakeRefCounted<TransientFailurePicker>(status));
+}
+
+//
+// factory
+//
+
+class CdsLbFactory : public LoadBalancingPolicyFactory {
+ public:
+  OrphanablePtr<LoadBalancingPolicy> CreateLoadBalancingPolicy(
+      LoadBalancingPolicy::Args args) const override {
+    return MakeOrphanable<CdsLb>(std::move(args));
+  }
+
+  absl::string_view name() const override { return kCds; }
+
+  absl::StatusOr<RefCountedPtr<LoadBalancingPolicy::Config>>
+  ParseLoadBalancingConfig(const Json& json) const override {
+    return LoadFromJson<RefCountedPtr<CdsLbConfig>>(
+        json, JsonArgs(), "errors validating cds LB policy config");
+  }
+};
+
+}  // namespace
+
+void RegisterCdsLbPolicy(CoreConfiguration::Builder* builder) {
+  builder->lb_policy_registry()->RegisterLoadBalancingPolicyFactory(
+      std::make_unique<CdsLbFactory>());
+}
+
+}  // namespace grpc_core