grpc 1.64.0 → 1.68.1

data/src/core/xds/xds_client/lrs_client.cc

@@ -0,0 +1,1292 @@
+//
+// Copyright 2018 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 "src/core/xds/xds_client/lrs_client.h"
+
+#include <grpc/event_engine/event_engine.h>
+
+#include <memory>
+#include <set>
+#include <string>
+#include <vector>
+
+#include "absl/cleanup/cleanup.h"
+#include "absl/log/check.h"
+#include "absl/log/log.h"
+#include "absl/strings/string_view.h"
+#include "absl/types/optional.h"
+#include "envoy/config/core/v3/base.upb.h"
+#include "envoy/config/endpoint/v3/load_report.upb.h"
+#include "envoy/service/load_stats/v3/lrs.upb.h"
+#include "envoy/service/load_stats/v3/lrs.upbdefs.h"
+#include "google/protobuf/duration.upb.h"
+#include "src/core/lib/debug/trace.h"
+#include "src/core/lib/iomgr/exec_ctx.h"
+#include "src/core/util/backoff.h"
+#include "src/core/util/debug_location.h"
+#include "src/core/util/env.h"
+#include "src/core/util/orphanable.h"
+#include "src/core/util/ref_counted_ptr.h"
+#include "src/core/util/string.h"
+#include "src/core/util/sync.h"
+#include "src/core/util/upb_utils.h"
+#include "src/core/util/uri.h"
+#include "src/core/xds/xds_client/xds_api.h"
+#include "src/core/xds/xds_client/xds_bootstrap.h"
+#include "src/core/xds/xds_client/xds_locality.h"
+#include "upb/base/string_view.h"
+#include "upb/mem/arena.h"
+#include "upb/reflection/def.h"
+#include "upb/text/encode.h"
+
+#define GRPC_XDS_INITIAL_CONNECT_BACKOFF_SECONDS 1
+#define GRPC_XDS_RECONNECT_BACKOFF_MULTIPLIER 1.6
+#define GRPC_XDS_RECONNECT_MAX_BACKOFF_SECONDS 120
+#define GRPC_XDS_RECONNECT_JITTER 0.2
+#define GRPC_XDS_MIN_CLIENT_LOAD_REPORTING_INTERVAL_MS 1000
+
+namespace grpc_core {
+
+using ::grpc_event_engine::experimental::EventEngine;
+
+// TODO(roth): Remove this once the feature passes interop tests.
+bool XdsOrcaLrsPropagationChangesEnabled() {
+  auto value = GetEnv("GRPC_EXPERIMENTAL_XDS_ORCA_LRS_PROPAGATION");
+  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;
+}
+
+namespace {
+
+uint64_t GetAndResetCounter(std::atomic<uint64_t>* from) {
+  return from->exchange(0, std::memory_order_relaxed);
+}
+
+}  // namespace
+
+//
+// LrsClient::ClusterDropStats
+//
+
+LrsClient::ClusterDropStats::ClusterDropStats(
+    RefCountedPtr<LrsClient> lrs_client, absl::string_view lrs_server,
+    absl::string_view cluster_name, absl::string_view eds_service_name)
+    : RefCounted(GRPC_TRACE_FLAG_ENABLED(xds_client_refcount)
+                     ? "ClusterDropStats"
+                     : nullptr),
+      lrs_client_(std::move(lrs_client)),
+      lrs_server_(lrs_server),
+      cluster_name_(cluster_name),
+      eds_service_name_(eds_service_name) {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client_.get() << "] created drop stats " << this
+      << " for {" << lrs_server_ << ", " << cluster_name_ << ", "
+      << eds_service_name_ << "}";
+}
+
+LrsClient::ClusterDropStats::~ClusterDropStats() {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client_.get() << "] destroying drop stats "
+      << this << " for {" << lrs_server_ << ", " << cluster_name_ << ", "
+      << eds_service_name_ << "}";
+  lrs_client_->RemoveClusterDropStats(lrs_server_, cluster_name_,
+                                      eds_service_name_, this);
+  lrs_client_.reset(DEBUG_LOCATION, "ClusterDropStats");
+}
+
+LrsClient::ClusterDropStats::Snapshot
+LrsClient::ClusterDropStats::GetSnapshotAndReset() {
+  Snapshot snapshot;
+  snapshot.uncategorized_drops = GetAndResetCounter(&uncategorized_drops_);
+  MutexLock lock(&mu_);
+  snapshot.categorized_drops = std::move(categorized_drops_);
+  return snapshot;
+}
+
+void LrsClient::ClusterDropStats::AddUncategorizedDrops() {
+  uncategorized_drops_.fetch_add(1);
+}
+
+void LrsClient::ClusterDropStats::AddCallDropped(const std::string& category) {
+  MutexLock lock(&mu_);
+  ++categorized_drops_[category];
+}
+
+//
+// LrsClient::ClusterLocalityStats
+//
+
+LrsClient::ClusterLocalityStats::ClusterLocalityStats(
+    RefCountedPtr<LrsClient> lrs_client, absl::string_view lrs_server,
+    absl::string_view cluster_name, absl::string_view eds_service_name,
+    RefCountedPtr<XdsLocalityName> name,
+    RefCountedPtr<const BackendMetricPropagation> backend_metric_propagation)
+    : RefCounted(GRPC_TRACE_FLAG_ENABLED(xds_client_refcount)
+                     ? "ClusterLocalityStats"
+                     : nullptr),
+      lrs_client_(std::move(lrs_client)),
+      lrs_server_(lrs_server),
+      cluster_name_(cluster_name),
+      eds_service_name_(eds_service_name),
+      name_(std::move(name)),
+      backend_metric_propagation_(std::move(backend_metric_propagation)) {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client_.get() << "] created locality stats "
+      << this << " for {" << lrs_server_ << ", " << cluster_name_ << ", "
+      << eds_service_name_ << ", "
+      << (name_ == nullptr ? "<none>" : name_->human_readable_string().c_str())
+      << ", propagation=" << backend_metric_propagation_->AsString() << "}";
+}
+
+LrsClient::ClusterLocalityStats::~ClusterLocalityStats() {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client_.get() << "] destroying locality stats "
+      << this << " for {" << lrs_server_ << ", " << cluster_name_ << ", "
+      << eds_service_name_ << ", "
+      << (name_ == nullptr ? "<none>" : name_->human_readable_string().c_str())
+      << ", propagation=" << backend_metric_propagation_->AsString() << "}";
+  lrs_client_->RemoveClusterLocalityStats(lrs_server_, cluster_name_,
+                                          eds_service_name_, name_,
+                                          backend_metric_propagation_, this);
+  lrs_client_.reset(DEBUG_LOCATION, "ClusterLocalityStats");
+}
+
+LrsClient::ClusterLocalityStats::Snapshot
+LrsClient::ClusterLocalityStats::GetSnapshotAndReset() {
+  Snapshot snapshot;
+  for (auto& percpu_stats : stats_) {
+    Snapshot percpu_snapshot = {
+        GetAndResetCounter(&percpu_stats.total_successful_requests),
+        // Don't reset total_requests_in_progress because it's
+        // not related to a single reporting interval.
+        percpu_stats.total_requests_in_progress.load(std::memory_order_relaxed),
+        GetAndResetCounter(&percpu_stats.total_error_requests),
+        GetAndResetCounter(&percpu_stats.total_issued_requests),
+        {},
+        {},
+        {},
+        {}};
+    {
+      MutexLock lock(&percpu_stats.backend_metrics_mu);
+      percpu_snapshot.cpu_utilization = std::move(percpu_stats.cpu_utilization);
+      percpu_snapshot.mem_utilization = std::move(percpu_stats.mem_utilization);
+      percpu_snapshot.application_utilization =
+          std::move(percpu_stats.application_utilization);
+      percpu_snapshot.backend_metrics = std::move(percpu_stats.backend_metrics);
+    }
+    snapshot += percpu_snapshot;
+  }
+  return snapshot;
+}
+
+void LrsClient::ClusterLocalityStats::AddCallStarted() {
+  Stats& stats = stats_.this_cpu();
+  stats.total_issued_requests.fetch_add(1, std::memory_order_relaxed);
+  stats.total_requests_in_progress.fetch_add(1, std::memory_order_relaxed);
+}
+
+void LrsClient::ClusterLocalityStats::AddCallFinished(
+    const BackendMetricData* backend_metrics, bool fail) {
+  Stats& stats = stats_.this_cpu();
+  std::atomic<uint64_t>& to_increment =
+      fail ? stats.total_error_requests : stats.total_successful_requests;
+  to_increment.fetch_add(1, std::memory_order_relaxed);
+  stats.total_requests_in_progress.fetch_add(-1, std::memory_order_acq_rel);
+  if (backend_metrics == nullptr) return;
+  MutexLock lock(&stats.backend_metrics_mu);
+  if (!XdsOrcaLrsPropagationChangesEnabled()) {
+    for (const auto& m : backend_metrics->named_metrics) {
+      stats.backend_metrics[std::string(m.first)] += BackendMetric(1, m.second);
+    }
+    return;
+  }
+  if (backend_metric_propagation_->propagation_bits &
+      BackendMetricPropagation::kCpuUtilization) {
+    stats.cpu_utilization += BackendMetric(1, backend_metrics->cpu_utilization);
+  }
+  if (backend_metric_propagation_->propagation_bits &
+      BackendMetricPropagation::kMemUtilization) {
+    stats.mem_utilization += BackendMetric(1, backend_metrics->mem_utilization);
+  }
+  if (backend_metric_propagation_->propagation_bits &
+      BackendMetricPropagation::kApplicationUtilization) {
+    stats.application_utilization +=
+        BackendMetric(1, backend_metrics->application_utilization);
+  }
+  if (backend_metric_propagation_->propagation_bits &
+          BackendMetricPropagation::kNamedMetricsAll ||
+      !backend_metric_propagation_->named_metric_keys.empty()) {
+    for (const auto& m : backend_metrics->named_metrics) {
+      if (backend_metric_propagation_->propagation_bits &
+              BackendMetricPropagation::kNamedMetricsAll ||
+          backend_metric_propagation_->named_metric_keys.contains(m.first)) {
+        stats.backend_metrics[absl::StrCat("named_metrics.", m.first)] +=
+            BackendMetric(1, m.second);
+      }
+    }
+  }
+}
+
+//
+// Internal class declarations
+//
+
+// A call wrapper that can restart a call upon failure.
+// The template parameter is the kind of wrapped call.
+// TODO(roth): This is basically the same code as in XdsClient, and
+// probably very similar to many other places in the codebase.
+// Consider refactoring this into a common utility library somehow.
+template <typename T>
+class LrsClient::LrsChannel::RetryableCall final
+    : public InternallyRefCounted<RetryableCall<T>> {
+ public:
+  explicit RetryableCall(WeakRefCountedPtr<LrsChannel> lrs_channel);
+
+  // Disable thread-safety analysis because this method is called via
+  // OrphanablePtr<>, but there's no way to pass the lock annotation
+  // through there.
+  void Orphan() override ABSL_NO_THREAD_SAFETY_ANALYSIS;
+
+  void OnCallFinishedLocked() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&LrsClient::mu_);
+
+  T* call() const { return call_.get(); }
+  LrsChannel* lrs_channel() const { return lrs_channel_.get(); }
+
+  bool IsCurrentCallOnChannel() const;
+
+ private:
+  void StartNewCallLocked();
+  void StartRetryTimerLocked() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&LrsClient::mu_);
+
+  void OnRetryTimer();
+
+  // The wrapped xds call that talks to the xds server. It's instantiated
+  // every time we start a new call. It's null during call retry backoff.
+  OrphanablePtr<T> call_;
+  // The owning xds channel.
+  WeakRefCountedPtr<LrsChannel> lrs_channel_;
+
+  // Retry state.
+  BackOff backoff_;
+  absl::optional<EventEngine::TaskHandle> timer_handle_
+      ABSL_GUARDED_BY(&LrsClient::mu_);
+
+  bool shutting_down_ = false;
+};
+
+// An LRS call to the LRS server.
+class LrsClient::LrsChannel::LrsCall final
+    : public InternallyRefCounted<LrsCall> {
+ public:
+  // The ctor and dtor should not be used directly.
+  explicit LrsCall(RefCountedPtr<RetryableCall<LrsCall>> retryable_call);
+
+  void Orphan() override;
+
+  RetryableCall<LrsCall>* retryable_call() { return retryable_call_.get(); }
+  LrsChannel* lrs_channel() const { return retryable_call_->lrs_channel(); }
+  LrsClient* lrs_client() const { return lrs_channel()->lrs_client(); }
+  bool seen_response() const { return seen_response_; }
+
+ private:
+  class StreamEventHandler final
+      : public XdsTransportFactory::XdsTransport::StreamingCall::EventHandler {
+   public:
+    explicit StreamEventHandler(RefCountedPtr<LrsCall> lrs_call)
+        : lrs_call_(std::move(lrs_call)) {}
+
+    void OnRequestSent(bool /*ok*/) override { lrs_call_->OnRequestSent(); }
+    void OnRecvMessage(absl::string_view payload) override {
+      lrs_call_->OnRecvMessage(payload);
+    }
+    void OnStatusReceived(absl::Status status) override {
+      lrs_call_->OnStatusReceived(std::move(status));
+    }
+
+   private:
+    RefCountedPtr<LrsCall> lrs_call_;
+  };
+
+  // A repeating timer for a particular duration.
+  class Timer final : public InternallyRefCounted<Timer> {
+   public:
+    explicit Timer(RefCountedPtr<LrsCall> lrs_call)
+        : lrs_call_(std::move(lrs_call)) {}
+    ~Timer() override { lrs_call_.reset(DEBUG_LOCATION, "LRS timer"); }
+
+    // Disable thread-safety analysis because this method is called via
+    // OrphanablePtr<>, but there's no way to pass the lock annotation
+    // through there.
+    void Orphan() override ABSL_NO_THREAD_SAFETY_ANALYSIS;
+
+    void ScheduleNextReportLocked()
+        ABSL_EXCLUSIVE_LOCKS_REQUIRED(&LrsClient::mu_);
+
+   private:
+    bool IsCurrentTimerOnCall() const {
+      return this == lrs_call_->timer_.get();
+    }
+    LrsClient* lrs_client() const { return lrs_call_->lrs_client(); }
+
+    void OnNextReportTimer();
+
+    // The owning LRS call.
+    RefCountedPtr<LrsCall> lrs_call_;
+
+    absl::optional<EventEngine::TaskHandle> timer_handle_
+        ABSL_GUARDED_BY(&LrsClient::mu_);
+  };
+
+  void MaybeScheduleNextReportLocked()
+      ABSL_EXCLUSIVE_LOCKS_REQUIRED(&LrsClient::mu_);
+
+  void SendReportLocked() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&LrsClient::mu_);
+
+  void SendMessageLocked(std::string payload)
+      ABSL_EXCLUSIVE_LOCKS_REQUIRED(&LrsClient::mu_);
+
+  void OnRequestSent();
+  void OnRecvMessage(absl::string_view payload);
+  void OnStatusReceived(absl::Status status);
+
+  bool IsCurrentCallOnChannel() const;
+
+  // The owning RetryableCall<>.
+  RefCountedPtr<RetryableCall<LrsCall>> retryable_call_;
+
+  OrphanablePtr<XdsTransportFactory::XdsTransport::StreamingCall>
+      streaming_call_;
+
+  bool seen_response_ = false;
+  bool send_message_pending_ ABSL_GUARDED_BY(&LrsClient::mu_) = false;
+
+  // Load reporting state.
+  bool send_all_clusters_ = false;
+  std::set<std::string> cluster_names_;  // Asked for by the LRS server.
+  Duration load_reporting_interval_;
+  bool last_report_counters_were_zero_ = false;
+  OrphanablePtr<Timer> timer_;
+};
+
+//
+// LrsClient::LrsChannel
+//
+
+LrsClient::LrsChannel::LrsChannel(
+    WeakRefCountedPtr<LrsClient> lrs_client,
+    std::shared_ptr<const XdsBootstrap::XdsServer> server)
+    : DualRefCounted<LrsChannel>(GRPC_TRACE_FLAG_ENABLED(xds_client_refcount)
+                                     ? "LrsChannel"
+                                     : nullptr),
+      lrs_client_(std::move(lrs_client)),
+      server_(std::move(server)) {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client_.get() << "] creating channel " << this
+      << " for server " << server_->server_uri();
+  absl::Status status;
+  transport_ = lrs_client_->transport_factory_->GetTransport(*server_, &status);
+  CHECK(transport_ != nullptr);
+  if (!status.ok()) {
+    LOG(ERROR) << "Error creating LRS channel to " << server_->server_uri()
+               << ": " << status;
+  }
+}
+
+LrsClient::LrsChannel::~LrsChannel() {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client() << "] destroying lrs channel " << this
+      << " for server " << server_->server_uri();
+  lrs_client_.reset(DEBUG_LOCATION, "LrsChannel");
+}
+
+// This method should only ever be called when holding the lock, but we can't
+// use a ABSL_EXCLUSIVE_LOCKS_REQUIRED annotation, because Orphan() will be
+// called from DualRefCounted::Unref(), which cannot have a lock annotation for
+// a lock in this subclass.
+void LrsClient::LrsChannel::Orphaned() ABSL_NO_THREAD_SAFETY_ANALYSIS {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client() << "] orphaning lrs channel " << this
+      << " for server " << server_->server_uri();
+  transport_.reset();
+  // At this time, all strong refs are removed, remove from channel map to
+  // prevent subsequent subscription from trying to use this LrsChannel as
+  // it is shutting down.
+  lrs_client_->lrs_channel_map_.erase(server_->Key());
+  lrs_call_.reset();
+}
+
+void LrsClient::LrsChannel::ResetBackoff() { transport_->ResetBackoff(); }
+
+void LrsClient::LrsChannel::MaybeStartLrsCall() {
+  if (lrs_call_ != nullptr) return;
+  lrs_call_ = MakeOrphanable<RetryableCall<LrsCall>>(
+      WeakRef(DEBUG_LOCATION, "LrsCall"));
+}
+
+void LrsClient::LrsChannel::StopLrsCallLocked() {
+  lrs_client_->load_report_map_.erase(server_->Key());
+  lrs_call_.reset();
+}
+
+//
+// LrsClient::LrsChannel::RetryableCall<>
+//
+
+template <typename T>
+LrsClient::LrsChannel::RetryableCall<T>::RetryableCall(
+    WeakRefCountedPtr<LrsChannel> lrs_channel)
+    : lrs_channel_(std::move(lrs_channel)),
+      backoff_(BackOff::Options()
+                   .set_initial_backoff(Duration::Seconds(
+                       GRPC_XDS_INITIAL_CONNECT_BACKOFF_SECONDS))
+                   .set_multiplier(GRPC_XDS_RECONNECT_BACKOFF_MULTIPLIER)
+                   .set_jitter(GRPC_XDS_RECONNECT_JITTER)
+                   .set_max_backoff(Duration::Seconds(
+                       GRPC_XDS_RECONNECT_MAX_BACKOFF_SECONDS))) {
+  StartNewCallLocked();
+}
+
+template <typename T>
+void LrsClient::LrsChannel::RetryableCall<T>::Orphan() {
+  shutting_down_ = true;
+  call_.reset();
+  if (timer_handle_.has_value()) {
+    lrs_channel()->lrs_client()->engine()->Cancel(*timer_handle_);
+    timer_handle_.reset();
+  }
+  this->Unref(DEBUG_LOCATION, "RetryableCall+orphaned");
+}
+
+template <typename T>
+void LrsClient::LrsChannel::RetryableCall<T>::OnCallFinishedLocked() {
+  // If we saw a response on the current stream, reset backoff.
+  if (call_->seen_response()) backoff_.Reset();
+  call_.reset();
+  // Start retry timer.
+  StartRetryTimerLocked();
+}
+
+template <typename T>
+void LrsClient::LrsChannel::RetryableCall<T>::StartNewCallLocked() {
+  if (shutting_down_) return;
+  CHECK(lrs_channel_->transport_ != nullptr);
+  CHECK(call_ == nullptr);
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_channel()->lrs_client() << "] lrs server "
+      << lrs_channel()->server_->server_uri()
+      << ": start new call from retryable call " << this;
+  call_ = MakeOrphanable<T>(
+      this->Ref(DEBUG_LOCATION, "RetryableCall+start_new_call"));
+}
+
+template <typename T>
+void LrsClient::LrsChannel::RetryableCall<T>::StartRetryTimerLocked() {
+  if (shutting_down_) return;
+  const Duration delay = backoff_.NextAttemptDelay();
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_channel()->lrs_client() << "] lrs server "
+      << lrs_channel()->server_->server_uri()
+      << ": call attempt failed; retry timer will fire in " << delay.millis()
+      << "ms.";
+  timer_handle_ = lrs_channel()->lrs_client()->engine()->RunAfter(
+      delay,
+      [self = this->Ref(DEBUG_LOCATION, "RetryableCall+retry_timer_start")]() {
+        ApplicationCallbackExecCtx callback_exec_ctx;
+        ExecCtx exec_ctx;
+        self->OnRetryTimer();
+      });
+}
+
+template <typename T>
+void LrsClient::LrsChannel::RetryableCall<T>::OnRetryTimer() {
+  MutexLock lock(&lrs_channel_->lrs_client()->mu_);
+  if (timer_handle_.has_value()) {
+    timer_handle_.reset();
+    if (shutting_down_) return;
+    GRPC_TRACE_LOG(xds_client, INFO)
+        << "[lrs_client " << lrs_channel()->lrs_client() << "] lrs server "
+        << lrs_channel()->server_->server_uri()
+        << ": retry timer fired (retryable call: " << this << ")";
+    StartNewCallLocked();
+  }
+}
+
+//
+// LrsClient::LrsChannel::LrsCall::Timer
+//
+
+void LrsClient::LrsChannel::LrsCall::Timer::Orphan() {
+  if (timer_handle_.has_value()) {
+    lrs_client()->engine()->Cancel(*timer_handle_);
+    timer_handle_.reset();
+  }
+  Unref(DEBUG_LOCATION, "Orphan");
+}
+
+void LrsClient::LrsChannel::LrsCall::Timer::ScheduleNextReportLocked() {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client() << "] lrs server "
+      << lrs_call_->lrs_channel()->server_->server_uri()
+      << ": scheduling next load report in "
+      << lrs_call_->load_reporting_interval_;
+  timer_handle_ = lrs_client()->engine()->RunAfter(
+      lrs_call_->load_reporting_interval_,
+      [self = Ref(DEBUG_LOCATION, "timer")]() {
+        ApplicationCallbackExecCtx callback_exec_ctx;
+        ExecCtx exec_ctx;
+        self->OnNextReportTimer();
+      });
+}
+
+void LrsClient::LrsChannel::LrsCall::Timer::OnNextReportTimer() {
+  MutexLock lock(&lrs_client()->mu_);
+  timer_handle_.reset();
+  if (IsCurrentTimerOnCall()) lrs_call_->SendReportLocked();
+}
+
+//
+// LrsClient::LrsChannel::LrsCall
+//
+
+LrsClient::LrsChannel::LrsCall::LrsCall(
+    RefCountedPtr<RetryableCall<LrsCall>> retryable_call)
+    : InternallyRefCounted<LrsCall>(
+          GRPC_TRACE_FLAG_ENABLED(xds_client_refcount) ? "LrsCall" : nullptr),
+      retryable_call_(std::move(retryable_call)) {
+  // Init the LRS call. Note that the call will progress every time there's
+  // activity in lrs_client()->interested_parties_, which is comprised of
+  // the polling entities from client_channel.
+  CHECK_NE(lrs_client(), nullptr);
+  const char* method =
+      "/envoy.service.load_stats.v3.LoadReportingService/StreamLoadStats";
+  streaming_call_ = lrs_channel()->transport_->CreateStreamingCall(
+      method, std::make_unique<StreamEventHandler>(
+                  // Passing the initial ref here.  This ref will go away when
+                  // the StreamEventHandler is destroyed.
+                  RefCountedPtr<LrsCall>(this)));
+  CHECK(streaming_call_ != nullptr);
+  // Start the call.
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client() << "] lrs server "
+      << lrs_channel()->server_->server_uri()
+      << ": starting LRS call (lrs_call=" << this
+      << ", streaming_call=" << streaming_call_.get() << ")";
+  // Send the initial request.
+  std::string serialized_payload = lrs_client()->CreateLrsInitialRequest();
+  SendMessageLocked(std::move(serialized_payload));
+  // Read initial response.
+  streaming_call_->StartRecvMessage();
+}
+
+void LrsClient::LrsChannel::LrsCall::Orphan() {
+  timer_.reset();
+  // Note that the initial ref is held by the StreamEventHandler, which
+  // will be destroyed when streaming_call_ is destroyed, which may not happen
+  // here, since there may be other refs held to streaming_call_ by internal
+  // callbacks.
+  streaming_call_.reset();
+}
+
+void LrsClient::LrsChannel::LrsCall::MaybeScheduleNextReportLocked() {
+  // If there are no more registered stats to report, cancel the call.
+  auto it = lrs_client()->load_report_map_.find(lrs_channel()->server_->Key());
+  if (it == lrs_client()->load_report_map_.end() ||
+      it->second.load_report_map.empty()) {
+    it->second.lrs_channel->StopLrsCallLocked();
+    return;
+  }
+  // Don't start if the previous send_message op hasn't completed yet.
+  // If this happens, we'll be called again from OnRequestSent().
+  if (send_message_pending_) return;
+  // Don't start if no LRS response has arrived.
+  if (!seen_response()) return;
+  // If there is no timer, create one.
+  // This happens on the initial response and whenever the interval changes.
+  if (timer_ == nullptr) {
+    timer_ = MakeOrphanable<Timer>(Ref(DEBUG_LOCATION, "LRS timer"));
+  }
+  // Schedule the next load report.
+  timer_->ScheduleNextReportLocked();
+}
+
+bool LrsClient::LoadReportCountersAreZero(
+    const ClusterLoadReportMap& snapshot) {
+  for (const auto& p : snapshot) {
+    const ClusterLoadReport& cluster_snapshot = p.second;
+    if (!cluster_snapshot.dropped_requests.IsZero()) return false;
+    for (const auto& q : cluster_snapshot.locality_stats) {
+      const ClusterLocalityStats::Snapshot& locality_snapshot = q.second;
+      if (!locality_snapshot.IsZero()) return false;
+    }
+  }
+  return true;
+}
+
+void LrsClient::LrsChannel::LrsCall::SendReportLocked() {
+  // Construct snapshot from all reported stats.
+  ClusterLoadReportMap snapshot = lrs_client()->BuildLoadReportSnapshotLocked(
+      *lrs_channel()->server_, send_all_clusters_, cluster_names_);
+  // Skip client load report if the counters were all zero in the last
+  // report and they are still zero in this one.
+  const bool old_val = last_report_counters_were_zero_;
+  last_report_counters_were_zero_ = LoadReportCountersAreZero(snapshot);
+  if (old_val && last_report_counters_were_zero_) {
+    MaybeScheduleNextReportLocked();
+    return;
+  }
+  // Send a request that contains the snapshot.
+  std::string serialized_payload =
+      lrs_client()->CreateLrsRequest(std::move(snapshot));
+  SendMessageLocked(std::move(serialized_payload));
+}
+
+void LrsClient::LrsChannel::LrsCall::SendMessageLocked(std::string payload) {
+  send_message_pending_ = true;
+  streaming_call_->SendMessage(std::move(payload));
+}
+
+void LrsClient::LrsChannel::LrsCall::OnRequestSent() {
+  MutexLock lock(&lrs_client()->mu_);
+  send_message_pending_ = false;
+  if (IsCurrentCallOnChannel()) MaybeScheduleNextReportLocked();
+}
+
+void LrsClient::LrsChannel::LrsCall::OnRecvMessage(absl::string_view payload) {
+  MutexLock lock(&lrs_client()->mu_);
+  // If we're no longer the current call, ignore the result.
+  if (!IsCurrentCallOnChannel()) return;
+  // Start recv after any code branch
+  auto cleanup = absl::MakeCleanup(
+      [call = streaming_call_.get()]() { call->StartRecvMessage(); });
+  // Parse the response.
+  bool send_all_clusters = false;
+  std::set<std::string> new_cluster_names;
+  Duration new_load_reporting_interval;
+  absl::Status status = lrs_client()->ParseLrsResponse(
+      payload, &send_all_clusters, &new_cluster_names,
+      &new_load_reporting_interval);
+  if (!status.ok()) {
+    LOG(ERROR) << "[lrs_client " << lrs_client() << "] lrs server "
+               << lrs_channel()->server_->server_uri()
+               << ": LRS response parsing failed: " << status;
+    return;
+  }
+  seen_response_ = true;
+  if (GRPC_TRACE_FLAG_ENABLED(xds_client)) {
+    LOG(INFO) << "[lrs_client " << lrs_client() << "] lrs server "
+              << lrs_channel()->server_->server_uri()
+              << ": LRS response received, " << new_cluster_names.size()
+              << " cluster names, send_all_clusters=" << send_all_clusters
+              << ", load_report_interval="
+              << new_load_reporting_interval.millis() << "ms";
+    size_t i = 0;
+    for (const auto& name : new_cluster_names) {
+      LOG(INFO) << "[lrs_client " << lrs_client() << "] cluster_name " << i++
+                << ": " << name;
+    }
+  }
+  if (new_load_reporting_interval <
+      Duration::Milliseconds(GRPC_XDS_MIN_CLIENT_LOAD_REPORTING_INTERVAL_MS)) {
+    new_load_reporting_interval =
+        Duration::Milliseconds(GRPC_XDS_MIN_CLIENT_LOAD_REPORTING_INTERVAL_MS);
+    GRPC_TRACE_LOG(xds_client, INFO)
+        << "[lrs_client " << lrs_client() << "] lrs server "
+        << lrs_channel()->server_->server_uri()
+        << ": increased load_report_interval to minimum value "
+        << GRPC_XDS_MIN_CLIENT_LOAD_REPORTING_INTERVAL_MS << "ms";
+  }
+  // Ignore identical update.
+  if (send_all_clusters == send_all_clusters_ &&
+      cluster_names_ == new_cluster_names &&
+      load_reporting_interval_ == new_load_reporting_interval) {
+    GRPC_TRACE_LOG(xds_client, INFO)
+        << "[lrs_client " << lrs_client() << "] lrs server "
+        << lrs_channel()->server_->server_uri()
+        << ": incoming LRS response identical to current, ignoring.";
+    return;
+  }
+  // If the interval has changed, we'll need to restart the timer below.
+  const bool restart_timer =
+      load_reporting_interval_ != new_load_reporting_interval;
+  // Record the new config.
+  send_all_clusters_ = send_all_clusters;
+  cluster_names_ = std::move(new_cluster_names);
+  load_reporting_interval_ = new_load_reporting_interval;
+  // Restart timer if needed.
+  if (restart_timer) {
+    timer_.reset();
+    MaybeScheduleNextReportLocked();
+  }
+}
+
+void LrsClient::LrsChannel::LrsCall::OnStatusReceived(absl::Status status) {
+  MutexLock lock(&lrs_client()->mu_);
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << lrs_client() << "] lrs server "
+      << lrs_channel()->server_->server_uri()
+      << ": LRS call status received (lrs_channel=" << lrs_channel()
+      << ", lrs_call=" << this << ", streaming_call=" << streaming_call_.get()
+      << "): " << status;
+  // Ignore status from a stale call.
+  if (IsCurrentCallOnChannel()) {
+    // Try to restart the call.
+    retryable_call_->OnCallFinishedLocked();
+  }
+}
+
+bool LrsClient::LrsChannel::LrsCall::IsCurrentCallOnChannel() const {
+  // If the retryable LRS call is null (which only happens when the lrs
+  // channel is shutting down), all the LRS calls are stale.
+  if (lrs_channel()->lrs_call_ == nullptr) return false;
+  return this == lrs_channel()->lrs_call_->call();
+}
+
+//
+// LrsClient
+//
+
+LrsClient::LrsClient(
+    std::shared_ptr<XdsBootstrap> bootstrap, std::string user_agent_name,
+    std::string user_agent_version,
+    RefCountedPtr<XdsTransportFactory> transport_factory,
+    std::shared_ptr<grpc_event_engine::experimental::EventEngine> engine)
+    : DualRefCounted<LrsClient>(
+          GRPC_TRACE_FLAG_ENABLED(xds_client_refcount) ? "LrsClient" : nullptr),
+      bootstrap_(std::move(bootstrap)),
+      user_agent_name_(std::move(user_agent_name)),
+      user_agent_version_(std::move(user_agent_version)),
+      transport_factory_(std::move(transport_factory)),
+      engine_(std::move(engine)) {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << this << "] creating lrs client";
+}
+
+LrsClient::~LrsClient() {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << this << "] destroying lrs client";
+}
+
+void LrsClient::Orphaned() {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << this << "] shutting down lrs client";
+  MutexLock lock(&mu_);
+  // We may still be sending lingering queued load report data, so don't
+  // just clear the load reporting map, but we do want to clear the refs
+  // we're holding to the LrsChannel objects, to make sure that
+  // everything shuts down properly.
+  for (auto& p : load_report_map_) {
+    p.second.lrs_channel.reset(DEBUG_LOCATION, "LrsClient::Orphan()");
+  }
+}
+
+RefCountedPtr<LrsClient::LrsChannel> LrsClient::GetOrCreateLrsChannelLocked(
+    std::shared_ptr<const XdsBootstrap::XdsServer> server, const char* reason) {
+  std::string key = server->Key();
+  auto it = lrs_channel_map_.find(key);
+  if (it != lrs_channel_map_.end()) {
+    return it->second->Ref(DEBUG_LOCATION, reason);
+  }
+  // Channel not found, so create a new one.
+  auto lrs_channel = MakeRefCounted<LrsChannel>(
+      WeakRef(DEBUG_LOCATION, "LrsChannel"), std::move(server));
+  lrs_channel_map_[std::move(key)] = lrs_channel.get();
+  return lrs_channel;
+}
+
+RefCountedPtr<LrsClient::ClusterDropStats> LrsClient::AddClusterDropStats(
+    std::shared_ptr<const XdsBootstrap::XdsServer> lrs_server,
+    absl::string_view cluster_name, absl::string_view eds_service_name) {
+  auto key =
+      std::make_pair(std::string(cluster_name), std::string(eds_service_name));
+  RefCountedPtr<ClusterDropStats> cluster_drop_stats;
+  {
+    MutexLock lock(&mu_);
+    // We jump through some hoops here to make sure that the
+    // absl::string_views stored in the ClusterDropStats object point
+    // to the strings in the load_report_map_ keys, so that
+    // they have the same lifetime.
+    auto server_it =
+        load_report_map_.emplace(lrs_server->Key(), LoadReportServer()).first;
+    if (server_it->second.lrs_channel == nullptr) {
+      server_it->second.lrs_channel = GetOrCreateLrsChannelLocked(
+          lrs_server, "load report map (drop stats)");
+    }
+    auto load_report_it = server_it->second.load_report_map
+                              .emplace(std::move(key), LoadReportState())
+                              .first;
+    LoadReportState& load_report_state = load_report_it->second;
+    if (load_report_state.drop_stats != nullptr) {
+      cluster_drop_stats = load_report_state.drop_stats->RefIfNonZero();
+    }
+    if (cluster_drop_stats == nullptr) {
+      if (load_report_state.drop_stats != nullptr) {
+        load_report_state.deleted_drop_stats +=
+            load_report_state.drop_stats->GetSnapshotAndReset();
+      }
+      cluster_drop_stats = MakeRefCounted<ClusterDropStats>(
+          Ref(DEBUG_LOCATION, "DropStats"), server_it->first /*lrs_server*/,
+          load_report_it->first.first /*cluster_name*/,
+          load_report_it->first.second /*eds_service_name*/);
+      load_report_state.drop_stats = cluster_drop_stats.get();
+    }
+    server_it->second.lrs_channel->MaybeStartLrsCall();
+  }
+  return cluster_drop_stats;
+}
+
+void LrsClient::RemoveClusterDropStats(
+    absl::string_view lrs_server_key, absl::string_view cluster_name,
+    absl::string_view eds_service_name,
+    LrsClient::ClusterDropStats* cluster_drop_stats) {
+  MutexLock lock(&mu_);
+  auto server_it = load_report_map_.find(lrs_server_key);
+  if (server_it == load_report_map_.end()) return;
+  auto load_report_it = server_it->second.load_report_map.find(
+      std::make_pair(std::string(cluster_name), std::string(eds_service_name)));
+  if (load_report_it == server_it->second.load_report_map.end()) return;
+  LoadReportState& load_report_state = load_report_it->second;
+  if (load_report_state.drop_stats == cluster_drop_stats) {
+    // Record final snapshot in deleted_drop_stats, which will be
+    // added to the next load report.
+    load_report_state.deleted_drop_stats +=
+        load_report_state.drop_stats->GetSnapshotAndReset();
+    load_report_state.drop_stats = nullptr;
+  }
+}
+
+RefCountedPtr<LrsClient::ClusterLocalityStats>
+LrsClient::AddClusterLocalityStats(
+    std::shared_ptr<const XdsBootstrap::XdsServer> lrs_server,
+    absl::string_view cluster_name, absl::string_view eds_service_name,
+    RefCountedPtr<XdsLocalityName> locality,
+    RefCountedPtr<const BackendMetricPropagation> backend_metric_propagation) {
+  auto key =
+      std::make_pair(std::string(cluster_name), std::string(eds_service_name));
+  RefCountedPtr<ClusterLocalityStats> cluster_locality_stats;
+  {
+    MutexLock lock(&mu_);
+    // We jump through some hoops here to make sure that the
+    // absl::string_views stored in the ClusterLocalityStats object point
+    // to the strings in the load_report_map_ keys, so that
+    // they have the same lifetime.
+    auto server_it =
+        load_report_map_.emplace(lrs_server->Key(), LoadReportServer()).first;
+    if (server_it->second.lrs_channel == nullptr) {
+      server_it->second.lrs_channel = GetOrCreateLrsChannelLocked(
+          std::move(lrs_server), "load report map (locality stats)");
+    }
+    auto load_report_it = server_it->second.load_report_map
+                              .emplace(std::move(key), LoadReportState())
+                              .first;
+    LoadReportState& load_report_state = load_report_it->second;
+    LoadReportState::LocalityState& locality_state =
+        load_report_state.locality_stats[locality];
+    ClusterLocalityStats*& locality_stats =
+        locality_state.propagation_stats[backend_metric_propagation];
+    if (locality_stats != nullptr) {
+      cluster_locality_stats = locality_stats->RefIfNonZero();
+    }
+    if (cluster_locality_stats == nullptr) {
+      if (locality_stats != nullptr) {
+        locality_state.deleted_locality_stats +=
+            locality_stats->GetSnapshotAndReset();
+      }
+      cluster_locality_stats = MakeRefCounted<ClusterLocalityStats>(
+          Ref(DEBUG_LOCATION, "LocalityStats"), server_it->first /*lrs_server*/,
+          load_report_it->first.first /*cluster_name*/,
+          load_report_it->first.second /*eds_service_name*/,
+          std::move(locality), std::move(backend_metric_propagation));
+      locality_stats = cluster_locality_stats.get();
+    }
+    server_it->second.lrs_channel->MaybeStartLrsCall();
+  }
+  return cluster_locality_stats;
+}
+
+void LrsClient::RemoveClusterLocalityStats(
+    absl::string_view lrs_server_key, absl::string_view cluster_name,
+    absl::string_view eds_service_name,
+    const RefCountedPtr<XdsLocalityName>& locality,
+    const RefCountedPtr<const BackendMetricPropagation>&
+        backend_metric_propagation,
+    ClusterLocalityStats* cluster_locality_stats) {
+  MutexLock lock(&mu_);
+  auto server_it = load_report_map_.find(lrs_server_key);
+  if (server_it == load_report_map_.end()) return;
+  auto load_report_it = server_it->second.load_report_map.find(
+      std::make_pair(std::string(cluster_name), std::string(eds_service_name)));
+  if (load_report_it == server_it->second.load_report_map.end()) return;
+  LoadReportState& load_report_state = load_report_it->second;
+  auto locality_it = load_report_state.locality_stats.find(locality);
+  if (locality_it == load_report_state.locality_stats.end()) return;
+  LoadReportState::LocalityState& locality_state = locality_it->second;
+  auto propagation_it =
+      locality_state.propagation_stats.find(backend_metric_propagation);
+  if (propagation_it == locality_state.propagation_stats.end()) return;
+  ClusterLocalityStats* locality_stats = propagation_it->second;
+  if (locality_stats == cluster_locality_stats) {
+    // Record final snapshot in deleted_locality_stats, which will be
+    // added to the next load report.
+    locality_state.deleted_locality_stats +=
+        locality_stats->GetSnapshotAndReset();
+    locality_state.propagation_stats.erase(propagation_it);
+  }
+}
+
+void LrsClient::ResetBackoff() {
+  MutexLock lock(&mu_);
+  for (auto& p : lrs_channel_map_) {
+    p.second->ResetBackoff();
+  }
+}
+
+LrsClient::ClusterLoadReportMap LrsClient::BuildLoadReportSnapshotLocked(
+    const XdsBootstrap::XdsServer& lrs_server, bool send_all_clusters,
+    const std::set<std::string>& clusters) {
+  GRPC_TRACE_LOG(xds_client, INFO)
+      << "[lrs_client " << this << "] start building load report";
+  ClusterLoadReportMap snapshot_map;
+  auto server_it = load_report_map_.find(lrs_server.Key());
+  if (server_it == load_report_map_.end()) return snapshot_map;
+  auto& load_report_map = server_it->second.load_report_map;
+  for (auto load_report_it = load_report_map.begin();
+       load_report_it != load_report_map.end();) {
+    // Cluster key is cluster and EDS service name.
+    const auto& cluster_key = load_report_it->first;
+    LoadReportState& load_report = load_report_it->second;
+    // If the CDS response for a cluster indicates to use LRS but the
+    // LRS server does not say that it wants reports for this cluster,
+    // then we'll have stats objects here whose data we're not going to
+    // include in the load report.  However, we still need to clear out
+    // the data from the stats objects, so that if the LRS server starts
+    // asking for the data in the future, we don't incorrectly include
+    // data from previous reporting intervals in that future report.
+    const bool record_stats =
+        send_all_clusters || clusters.find(cluster_key.first) != clusters.end();
+    ClusterLoadReport snapshot;
+    // Aggregate drop stats.
+    snapshot.dropped_requests = std::move(load_report.deleted_drop_stats);
+    if (load_report.drop_stats != nullptr) {
+      snapshot.dropped_requests +=
+          load_report.drop_stats->GetSnapshotAndReset();
+      GRPC_TRACE_LOG(xds_client, INFO)
+          << "[lrs_client " << this << "] cluster=" << cluster_key.first
+          << " eds_service_name=" << cluster_key.second
+          << " drop_stats=" << load_report.drop_stats;
+    }
+    // Aggregate locality stats.
+    for (auto it = load_report.locality_stats.begin();
+         it != load_report.locality_stats.end();) {
+      const RefCountedPtr<XdsLocalityName>& locality_name = it->first;
+      auto& locality_state = it->second;
+      ClusterLocalityStats::Snapshot& locality_snapshot =
+          snapshot.locality_stats[locality_name];
+      locality_snapshot = std::move(locality_state.deleted_locality_stats);
+      for (const auto& p : locality_state.propagation_stats) {
+        ClusterLocalityStats* locality_stats = p.second;
+        if (locality_stats != nullptr) {
+          locality_snapshot += locality_stats->GetSnapshotAndReset();
+          GRPC_TRACE_LOG(xds_client, INFO)
+              << "[lrs_client " << this
+              << "] cluster=" << cluster_key.first.c_str()
+              << " eds_service_name=" << cluster_key.second.c_str()
+              << " locality=" << locality_name->human_readable_string().c_str()
+              << " propagation=" << p.first->AsString()
+              << " locality_stats=" << locality_stats;
+        }
+      }
+      // If the only thing left in this entry was final snapshots from
+      // deleted locality stats objects, remove the entry.
+      if (locality_state.propagation_stats.empty()) {
+        it = load_report.locality_stats.erase(it);
+      } else {
+        ++it;
+      }
+    }
+    // Compute load report interval.
+    const Timestamp now = Timestamp::Now();
+    snapshot.load_report_interval = now - load_report.last_report_time;
+    load_report.last_report_time = now;
+    // Record snapshot.
+    if (record_stats) {
+      snapshot_map[cluster_key] = std::move(snapshot);
+    }
+    // If the only thing left in this entry was final snapshots from
+    // deleted stats objects, remove the entry.
+    if (load_report.locality_stats.empty() &&
+        load_report.drop_stats == nullptr) {
+      load_report_it = load_report_map.erase(load_report_it);
+    } else {
+      ++load_report_it;
+    }
+  }
+  return snapshot_map;
+}
+
+namespace {
+
+struct LrsApiContext {
+  LrsClient* client;
+  upb_DefPool* def_pool;
+  upb_Arena* arena;
+};
+
+void MaybeLogLrsRequest(
+    const LrsApiContext& context,
+    const envoy_service_load_stats_v3_LoadStatsRequest* request) {
+  if (GRPC_TRACE_FLAG_ENABLED(xds_client) && ABSL_VLOG_IS_ON(2)) {
+    const upb_MessageDef* msg_type =
+        envoy_service_load_stats_v3_LoadStatsRequest_getmsgdef(
+            context.def_pool);
+    char buf[10240];
+    upb_TextEncode(reinterpret_cast<const upb_Message*>(request), msg_type,
+                   nullptr, 0, buf, sizeof(buf));
+    VLOG(2) << "[lrs_client " << context.client
+            << "] constructed LRS request: " << buf;
+  }
+}
+
+std::string SerializeLrsRequest(
+    const LrsApiContext& context,
+    const envoy_service_load_stats_v3_LoadStatsRequest* request) {
+  size_t output_length;
+  char* output = envoy_service_load_stats_v3_LoadStatsRequest_serialize(
+      request, context.arena, &output_length);
+  return std::string(output, output_length);
+}
+
+}  // namespace
+
+std::string LrsClient::CreateLrsInitialRequest() {
+  upb::Arena arena;
+  const LrsApiContext context = {this, def_pool_.ptr(), arena.ptr()};
+  // Create a request.
+  envoy_service_load_stats_v3_LoadStatsRequest* request =
+      envoy_service_load_stats_v3_LoadStatsRequest_new(arena.ptr());
+  // Populate node.
+  envoy_config_core_v3_Node* node_msg =
+      envoy_service_load_stats_v3_LoadStatsRequest_mutable_node(request,
+                                                                arena.ptr());
+  PopulateXdsNode(bootstrap_->node(), user_agent_name_, user_agent_version_,
+                  node_msg, arena.ptr());
+  envoy_config_core_v3_Node_add_client_features(
+      node_msg,
+      upb_StringView_FromString("envoy.lrs.supports_send_all_clusters"),
+      arena.ptr());
+  MaybeLogLrsRequest(context, request);
+  return SerializeLrsRequest(context, request);
+}
+
+namespace {
+
+void MaybeAddUnnamedMetric(
+    const LrsApiContext& context,
+    const LrsClient::ClusterLocalityStats::BackendMetric& backend_metric,
+    envoy_config_endpoint_v3_UnnamedEndpointLoadMetricStats* (*add_field)(
+        envoy_config_endpoint_v3_UpstreamLocalityStats*, upb_Arena*),
+    envoy_config_endpoint_v3_UpstreamLocalityStats* output) {
+  if (backend_metric.IsZero()) return;
+  auto* metric_proto = add_field(output, context.arena);
+  envoy_config_endpoint_v3_UnnamedEndpointLoadMetricStats_set_num_requests_finished_with_metric(
+      metric_proto, backend_metric.num_requests_finished_with_metric);
+  envoy_config_endpoint_v3_UnnamedEndpointLoadMetricStats_set_total_metric_value(
+      metric_proto, backend_metric.total_metric_value);
+}
+
+void LocalityStatsPopulate(
+    const LrsApiContext& context,
+    envoy_config_endpoint_v3_UpstreamLocalityStats* output,
+    const XdsLocalityName& locality_name,
+    const LrsClient::ClusterLocalityStats::Snapshot& snapshot) {
+  // Set locality.
+  envoy_config_core_v3_Locality* locality =
+      envoy_config_endpoint_v3_UpstreamLocalityStats_mutable_locality(
+          output, context.arena);
+  if (!locality_name.region().empty()) {
+    envoy_config_core_v3_Locality_set_region(
+        locality, StdStringToUpbString(locality_name.region()));
+  }
+  if (!locality_name.zone().empty()) {
+    envoy_config_core_v3_Locality_set_zone(
+        locality, StdStringToUpbString(locality_name.zone()));
+  }
+  if (!locality_name.sub_zone().empty()) {
+    envoy_config_core_v3_Locality_set_sub_zone(
+        locality, StdStringToUpbString(locality_name.sub_zone()));
+  }
+  // Set total counts.
+  envoy_config_endpoint_v3_UpstreamLocalityStats_set_total_successful_requests(
+      output, snapshot.total_successful_requests);
+  envoy_config_endpoint_v3_UpstreamLocalityStats_set_total_requests_in_progress(
+      output, snapshot.total_requests_in_progress);
+  envoy_config_endpoint_v3_UpstreamLocalityStats_set_total_error_requests(
+      output, snapshot.total_error_requests);
+  envoy_config_endpoint_v3_UpstreamLocalityStats_set_total_issued_requests(
+      output, snapshot.total_issued_requests);
+  // Add backend metrics.
+  MaybeAddUnnamedMetric(
+      context, snapshot.cpu_utilization,
+      envoy_config_endpoint_v3_UpstreamLocalityStats_mutable_cpu_utilization,
+      output);
+  MaybeAddUnnamedMetric(
+      context, snapshot.mem_utilization,
+      envoy_config_endpoint_v3_UpstreamLocalityStats_mutable_mem_utilization,
+      output);
+  MaybeAddUnnamedMetric(
+      context, snapshot.application_utilization,
+      envoy_config_endpoint_v3_UpstreamLocalityStats_mutable_application_utilization,
+      output);
+  for (const auto& p : snapshot.backend_metrics) {
+    const std::string& metric_name = p.first;
+    const LrsClient::ClusterLocalityStats::BackendMetric& metric_value =
+        p.second;
+    envoy_config_endpoint_v3_EndpointLoadMetricStats* load_metric =
+        envoy_config_endpoint_v3_UpstreamLocalityStats_add_load_metric_stats(
+            output, context.arena);
+    envoy_config_endpoint_v3_EndpointLoadMetricStats_set_metric_name(
+        load_metric, StdStringToUpbString(metric_name));
+    envoy_config_endpoint_v3_EndpointLoadMetricStats_set_num_requests_finished_with_metric(
+        load_metric, metric_value.num_requests_finished_with_metric);
+    envoy_config_endpoint_v3_EndpointLoadMetricStats_set_total_metric_value(
+        load_metric, metric_value.total_metric_value);
+  }
+}
+
+}  // namespace
+
+std::string LrsClient::CreateLrsRequest(
+    ClusterLoadReportMap cluster_load_report_map) {
+  upb::Arena arena;
+  const LrsApiContext context = {this, def_pool_.ptr(), arena.ptr()};
+  // Create a request.
+  envoy_service_load_stats_v3_LoadStatsRequest* request =
+      envoy_service_load_stats_v3_LoadStatsRequest_new(arena.ptr());
+  for (auto& p : cluster_load_report_map) {
+    const std::string& cluster_name = p.first.first;
+    const std::string& eds_service_name = p.first.second;
+    const ClusterLoadReport& load_report = p.second;
+    // Add cluster stats.
+    envoy_config_endpoint_v3_ClusterStats* cluster_stats =
+        envoy_service_load_stats_v3_LoadStatsRequest_add_cluster_stats(
+            request, arena.ptr());
+    // Set the cluster name.
+    envoy_config_endpoint_v3_ClusterStats_set_cluster_name(
+        cluster_stats, StdStringToUpbString(cluster_name));
+    // Set EDS service name, if non-empty.
+    if (!eds_service_name.empty()) {
+      envoy_config_endpoint_v3_ClusterStats_set_cluster_service_name(
+          cluster_stats, StdStringToUpbString(eds_service_name));
+    }
+    // Add locality stats.
+    for (const auto& p : load_report.locality_stats) {
+      const XdsLocalityName& locality_name = *p.first;
+      const auto& snapshot = p.second;
+      envoy_config_endpoint_v3_UpstreamLocalityStats* locality_stats =
+          envoy_config_endpoint_v3_ClusterStats_add_upstream_locality_stats(
+              cluster_stats, arena.ptr());
+      LocalityStatsPopulate(context, locality_stats, locality_name, snapshot);
+    }
+    // Add dropped requests.
+    uint64_t total_dropped_requests = 0;
+    for (const auto& p : load_report.dropped_requests.categorized_drops) {
+      const std::string& category = p.first;
+      const uint64_t count = p.second;
+      envoy_config_endpoint_v3_ClusterStats_DroppedRequests* dropped_requests =
+          envoy_config_endpoint_v3_ClusterStats_add_dropped_requests(
+              cluster_stats, arena.ptr());
+      envoy_config_endpoint_v3_ClusterStats_DroppedRequests_set_category(
+          dropped_requests, StdStringToUpbString(category));
+      envoy_config_endpoint_v3_ClusterStats_DroppedRequests_set_dropped_count(
+          dropped_requests, count);
+      total_dropped_requests += count;
+    }
+    total_dropped_requests += load_report.dropped_requests.uncategorized_drops;
+    // Set total dropped requests.
+    envoy_config_endpoint_v3_ClusterStats_set_total_dropped_requests(
+        cluster_stats, total_dropped_requests);
+    // Set real load report interval.
+    gpr_timespec timespec = load_report.load_report_interval.as_timespec();
+    google_protobuf_Duration* load_report_interval =
+        envoy_config_endpoint_v3_ClusterStats_mutable_load_report_interval(
+            cluster_stats, arena.ptr());
+    google_protobuf_Duration_set_seconds(load_report_interval, timespec.tv_sec);
+    google_protobuf_Duration_set_nanos(load_report_interval, timespec.tv_nsec);
+  }
+  MaybeLogLrsRequest(context, request);
+  return SerializeLrsRequest(context, request);
+}
+
+namespace {
+
+void MaybeLogLrsResponse(
+    const LrsApiContext& context,
+    const envoy_service_load_stats_v3_LoadStatsResponse* response) {
+  if (GRPC_TRACE_FLAG_ENABLED(xds_client) && ABSL_VLOG_IS_ON(2)) {
+    const upb_MessageDef* msg_type =
+        envoy_service_load_stats_v3_LoadStatsResponse_getmsgdef(
+            context.def_pool);
+    char buf[10240];
+    upb_TextEncode(reinterpret_cast<const upb_Message*>(response), msg_type,
+                   nullptr, 0, buf, sizeof(buf));
+    VLOG(2) << "[lrs_client " << context.client
+            << "] received LRS response: " << buf;
+  }
+}
+
+}  // namespace
+
+absl::Status LrsClient::ParseLrsResponse(absl::string_view encoded_response,
+                                         bool* send_all_clusters,
+                                         std::set<std::string>* cluster_names,
+                                         Duration* load_reporting_interval) {
+  upb::Arena arena;
+  // Decode the response.
+  const envoy_service_load_stats_v3_LoadStatsResponse* decoded_response =
+      envoy_service_load_stats_v3_LoadStatsResponse_parse(
+          encoded_response.data(), encoded_response.size(), arena.ptr());
+  // Parse the response.
+  if (decoded_response == nullptr) {
+    return absl::UnavailableError("Can't decode response.");
+  }
+  const LrsApiContext context = {this, def_pool_.ptr(), arena.ptr()};
+  MaybeLogLrsResponse(context, decoded_response);
+  // Check send_all_clusters.
+  if (envoy_service_load_stats_v3_LoadStatsResponse_send_all_clusters(
+          decoded_response)) {
+    *send_all_clusters = true;
+  } else {
+    // Store the cluster names.
+    size_t size;
+    const upb_StringView* clusters =
+        envoy_service_load_stats_v3_LoadStatsResponse_clusters(decoded_response,
+                                                               &size);
+    for (size_t i = 0; i < size; ++i) {
+      cluster_names->emplace(UpbStringToStdString(clusters[i]));
+    }
+  }
+  // Get the load report interval.
+  const google_protobuf_Duration* load_reporting_interval_duration =
+      envoy_service_load_stats_v3_LoadStatsResponse_load_reporting_interval(
+          decoded_response);
+  *load_reporting_interval = Duration::FromSecondsAndNanoseconds(
+      google_protobuf_Duration_seconds(load_reporting_interval_duration),
+      google_protobuf_Duration_nanos(load_reporting_interval_duration));
+  return absl::OkStatus();
+}
+
+}  // namespace grpc_core