grpc 1.57.0 → 1.62.0

data/src/core/load_balancing/rls/rls.cc

@@ -0,0 +1,2477 @@
+//
+// Copyright 2020 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.
+//
+
+// Implementation of the Route Lookup Service (RLS) LB policy
+//
+// The policy queries a route lookup service for the name of the actual service
+// to use. A child policy that recognizes the name as a field of its
+// configuration will take further load balancing action on the request.
+
+#include <grpc/support/port_platform.h>
+
+#include <inttypes.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <algorithm>
+#include <deque>
+#include <list>
+#include <map>
+#include <memory>
+#include <random>
+#include <set>
+#include <string>
+#include <type_traits>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include "absl/base/thread_annotations.h"
+#include "absl/hash/hash.h"
+#include "absl/status/status.h"
+#include "absl/status/statusor.h"
+#include "absl/strings/str_cat.h"
+#include "absl/strings/str_format.h"
+#include "absl/strings/str_join.h"
+#include "absl/strings/string_view.h"
+#include "absl/types/optional.h"
+#include "upb/base/string_view.h"
+#include "upb/mem/arena.hpp"
+
+#include <grpc/byte_buffer.h>
+#include <grpc/byte_buffer_reader.h>
+#include <grpc/event_engine/event_engine.h>
+#include <grpc/grpc.h>
+#include <grpc/impl/channel_arg_names.h>
+#include <grpc/impl/connectivity_state.h>
+#include <grpc/impl/propagation_bits.h>
+#include <grpc/slice.h>
+#include <grpc/status.h>
+#include <grpc/support/json.h>
+#include <grpc/support/log.h>
+
+#include "src/core/client_channel/client_channel_filter.h"
+#include "src/core/load_balancing/child_policy_handler.h"
+#include "src/core/lib/backoff/backoff.h"
+#include "src/core/lib/channel/channel_args.h"
+#include "src/core/lib/channel/channelz.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/dual_ref_counted.h"
+#include "src/core/lib/gprpp/orphanable.h"
+#include "src/core/lib/gprpp/ref_counted_ptr.h"
+#include "src/core/lib/gprpp/status_helper.h"
+#include "src/core/lib/gprpp/sync.h"
+#include "src/core/lib/gprpp/time.h"
+#include "src/core/lib/gprpp/validation_errors.h"
+#include "src/core/lib/gprpp/work_serializer.h"
+#include "src/core/lib/iomgr/closure.h"
+#include "src/core/lib/iomgr/error.h"
+#include "src/core/lib/iomgr/exec_ctx.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/lib/security/credentials/fake/fake_credentials.h"
+#include "src/core/service_config/service_config_impl.h"
+#include "src/core/lib/slice/slice.h"
+#include "src/core/lib/slice/slice_internal.h"
+#include "src/core/lib/surface/call.h"
+#include "src/core/lib/surface/channel.h"
+#include "src/core/lib/transport/connectivity_state.h"
+#include "src/core/lib/transport/error_utils.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/endpoint_addresses.h"
+#include "src/core/resolver/resolver_registry.h"
+#include "src/proto/grpc/lookup/v1/rls.upb.h"
+
+namespace grpc_core {
+
+TraceFlag grpc_lb_rls_trace(false, "rls_lb");
+
+namespace {
+
+using ::grpc_event_engine::experimental::EventEngine;
+
+constexpr absl::string_view kRls = "rls_experimental";
+const char kGrpc[] = "grpc";
+const char* kRlsRequestPath = "/grpc.lookup.v1.RouteLookupService/RouteLookup";
+const char* kFakeTargetFieldValue = "fake_target_field_value";
+const char* kRlsHeaderKey = "x-google-rls-data";
+
+const Duration kDefaultLookupServiceTimeout = Duration::Seconds(10);
+const Duration kMaxMaxAge = Duration::Minutes(5);
+const Duration kMinExpirationTime = Duration::Seconds(5);
+const Duration kCacheBackoffInitial = Duration::Seconds(1);
+const double kCacheBackoffMultiplier = 1.6;
+const double kCacheBackoffJitter = 0.2;
+const Duration kCacheBackoffMax = Duration::Minutes(2);
+const Duration kDefaultThrottleWindowSize = Duration::Seconds(30);
+const double kDefaultThrottleRatioForSuccesses = 2.0;
+const int kDefaultThrottlePadding = 8;
+const Duration kCacheCleanupTimerInterval = Duration::Minutes(1);
+const int64_t kMaxCacheSizeBytes = 5 * 1024 * 1024;
+
+// Parsed RLS LB policy configuration.
+class RlsLbConfig : public LoadBalancingPolicy::Config {
+ public:
+  struct KeyBuilder {
+    std::map<std::string /*key*/, std::vector<std::string /*header*/>>
+        header_keys;
+    std::string host_key;
+    std::string service_key;
+    std::string method_key;
+    std::map<std::string /*key*/, std::string /*value*/> constant_keys;
+  };
+  using KeyBuilderMap = std::unordered_map<std::string /*path*/, KeyBuilder>;
+
+  struct RouteLookupConfig {
+    KeyBuilderMap key_builder_map;
+    std::string lookup_service;
+    Duration lookup_service_timeout = kDefaultLookupServiceTimeout;
+    Duration max_age = kMaxMaxAge;
+    Duration stale_age = kMaxMaxAge;
+    int64_t cache_size_bytes = 0;
+    std::string default_target;
+
+    static const JsonLoaderInterface* JsonLoader(const JsonArgs&);
+    void JsonPostLoad(const Json& json, const JsonArgs& args,
+                      ValidationErrors* errors);
+  };
+
+  RlsLbConfig() = default;
+
+  RlsLbConfig(const RlsLbConfig&) = delete;
+  RlsLbConfig& operator=(const RlsLbConfig&) = delete;
+
+  RlsLbConfig(RlsLbConfig&& other) = delete;
+  RlsLbConfig& operator=(RlsLbConfig&& other) = delete;
+
+  absl::string_view name() const override { return kRls; }
+
+  const KeyBuilderMap& key_builder_map() const {
+    return route_lookup_config_.key_builder_map;
+  }
+  const std::string& lookup_service() const {
+    return route_lookup_config_.lookup_service;
+  }
+  Duration lookup_service_timeout() const {
+    return route_lookup_config_.lookup_service_timeout;
+  }
+  Duration max_age() const { return route_lookup_config_.max_age; }
+  Duration stale_age() const { return route_lookup_config_.stale_age; }
+  int64_t cache_size_bytes() const {
+    return route_lookup_config_.cache_size_bytes;
+  }
+  const std::string& default_target() const {
+    return route_lookup_config_.default_target;
+  }
+  const std::string& rls_channel_service_config() const {
+    return rls_channel_service_config_;
+  }
+  const Json& child_policy_config() const { return child_policy_config_; }
+  const std::string& child_policy_config_target_field_name() const {
+    return child_policy_config_target_field_name_;
+  }
+  RefCountedPtr<LoadBalancingPolicy::Config>
+  default_child_policy_parsed_config() const {
+    return default_child_policy_parsed_config_;
+  }
+
+  static const JsonLoaderInterface* JsonLoader(const JsonArgs&);
+  void JsonPostLoad(const Json& json, const JsonArgs&,
+                    ValidationErrors* errors);
+
+ private:
+  RouteLookupConfig route_lookup_config_;
+  std::string rls_channel_service_config_;
+  Json child_policy_config_;
+  std::string child_policy_config_target_field_name_;
+  RefCountedPtr<LoadBalancingPolicy::Config>
+      default_child_policy_parsed_config_;
+};
+
+// RLS LB policy.
+class RlsLb : public LoadBalancingPolicy {
+ public:
+  explicit RlsLb(Args args);
+
+  absl::string_view name() const override { return kRls; }
+  absl::Status UpdateLocked(UpdateArgs args) override;
+  void ExitIdleLocked() override;
+  void ResetBackoffLocked() override;
+
+ private:
+  // Key to access entries in the cache and the request map.
+  struct RequestKey {
+    std::map<std::string, std::string> key_map;
+
+    bool operator==(const RequestKey& rhs) const {
+      return key_map == rhs.key_map;
+    }
+
+    template <typename H>
+    friend H AbslHashValue(H h, const RequestKey& key) {
+      std::hash<std::string> string_hasher;
+      for (auto& kv : key.key_map) {
+        h = H::combine(std::move(h), string_hasher(kv.first),
+                       string_hasher(kv.second));
+      }
+      return h;
+    }
+
+    size_t Size() const {
+      size_t size = sizeof(RequestKey);
+      for (auto& kv : key_map) {
+        size += kv.first.length() + kv.second.length();
+      }
+      return size;
+    }
+
+    std::string ToString() const {
+      return absl::StrCat(
+          "{", absl::StrJoin(key_map, ",", absl::PairFormatter("=")), "}");
+    }
+  };
+
+  // Data from an RLS response.
+  struct ResponseInfo {
+    absl::Status status;
+    std::vector<std::string> targets;
+    std::string header_data;
+
+    std::string ToString() const {
+      return absl::StrFormat("{status=%s, targets=[%s], header_data=\"%s\"}",
+                             status.ToString(), absl::StrJoin(targets, ","),
+                             header_data);
+    }
+  };
+
+  // Wraps a child policy for a given RLS target.
+  class ChildPolicyWrapper : public DualRefCounted<ChildPolicyWrapper> {
+   public:
+    ChildPolicyWrapper(RefCountedPtr<RlsLb> lb_policy, std::string target);
+
+    // Note: We are forced to disable lock analysis here because
+    // Orphan() is called by OrphanablePtr<>, which cannot have lock
+    // annotations for this particular caller.
+    void Orphan() override ABSL_NO_THREAD_SAFETY_ANALYSIS;
+
+    const std::string& target() const { return target_; }
+
+    PickResult Pick(PickArgs args) ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+      return picker_->Pick(args);
+    }
+
+    // Updates for the child policy are handled in two phases:
+    // 1. In StartUpdate(), we parse and validate the new child policy
+    //    config and store the parsed config.
+    // 2. In MaybeFinishUpdate(), we actually pass the parsed config to the
+    //    child policy's UpdateLocked() method.
+    //
+    // The reason we do this is to avoid deadlocks.  In StartUpdate(),
+    // if the new config fails to validate, then we need to set
+    // picker_ to an instance that will fail all requests, which
+    // requires holding the lock.  However, we cannot call the child
+    // policy's UpdateLocked() method from MaybeFinishUpdate() while
+    // holding the lock, since that would cause a deadlock: the child's
+    // UpdateLocked() will call the helper's UpdateState() method, which
+    // will try to acquire the lock to set picker_.  So StartUpdate() is
+    // called while we are still holding the lock, but MaybeFinishUpdate()
+    // is called after releasing it.
+    //
+    // Both methods grab the data they need from the parent object.
+    void StartUpdate() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+    absl::Status MaybeFinishUpdate() ABSL_LOCKS_EXCLUDED(&RlsLb::mu_);
+
+    void ExitIdleLocked() {
+      if (child_policy_ != nullptr) child_policy_->ExitIdleLocked();
+    }
+
+    void ResetBackoffLocked() {
+      if (child_policy_ != nullptr) child_policy_->ResetBackoffLocked();
+    }
+
+    // Gets the connectivity state of the child policy. Once the child policy
+    // reports TRANSIENT_FAILURE, the function will always return
+    // TRANSIENT_FAILURE state instead of the actual state of the child policy
+    // until the child policy reports another READY state.
+    grpc_connectivity_state connectivity_state() const
+        ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+      return connectivity_state_;
+    }
+
+   private:
+    // ChannelControlHelper object that allows the child policy to update state
+    // with the wrapper.
+    class ChildPolicyHelper : public DelegatingChannelControlHelper {
+     public:
+      explicit ChildPolicyHelper(WeakRefCountedPtr<ChildPolicyWrapper> wrapper)
+          : wrapper_(std::move(wrapper)) {}
+      ~ChildPolicyHelper() override {
+        wrapper_.reset(DEBUG_LOCATION, "ChildPolicyHelper");
+      }
+
+      void UpdateState(grpc_connectivity_state state,
+                       const absl::Status& status,
+                       RefCountedPtr<SubchannelPicker> picker) override;
+
+     private:
+      ChannelControlHelper* parent_helper() const override {
+        return wrapper_->lb_policy_->channel_control_helper();
+      }
+
+      WeakRefCountedPtr<ChildPolicyWrapper> wrapper_;
+    };
+
+    RefCountedPtr<RlsLb> lb_policy_;
+    std::string target_;
+
+    bool is_shutdown_ = false;
+
+    OrphanablePtr<ChildPolicyHandler> child_policy_;
+    RefCountedPtr<LoadBalancingPolicy::Config> pending_config_;
+
+    grpc_connectivity_state connectivity_state_ ABSL_GUARDED_BY(&RlsLb::mu_) =
+        GRPC_CHANNEL_CONNECTING;
+    RefCountedPtr<LoadBalancingPolicy::SubchannelPicker> picker_
+        ABSL_GUARDED_BY(&RlsLb::mu_);
+  };
+
+  // A picker that uses the cache and the request map in the LB policy
+  // (synchronized via a mutex) to determine how to route requests.
+  class Picker : public LoadBalancingPolicy::SubchannelPicker {
+   public:
+    explicit Picker(RefCountedPtr<RlsLb> lb_policy);
+
+    PickResult Pick(PickArgs args) override;
+
+   private:
+    RefCountedPtr<RlsLb> lb_policy_;
+    RefCountedPtr<RlsLbConfig> config_;
+    RefCountedPtr<ChildPolicyWrapper> default_child_policy_;
+  };
+
+  // An LRU cache with adjustable size.
+  class Cache {
+   public:
+    using Iterator = std::list<RequestKey>::iterator;
+
+    class Entry : public InternallyRefCounted<Entry> {
+     public:
+      Entry(RefCountedPtr<RlsLb> lb_policy, const RequestKey& key);
+
+      // Notify the entry when it's evicted from the cache. Performs shut down.
+      // Note: We are forced to disable lock analysis here because
+      // Orphan() is called by OrphanablePtr<>, which cannot have lock
+      // annotations for this particular caller.
+      void Orphan() override ABSL_NO_THREAD_SAFETY_ANALYSIS;
+
+      const absl::Status& status() const
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+        return status_;
+      }
+      Timestamp backoff_time() const
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+        return backoff_time_;
+      }
+      Timestamp backoff_expiration_time() const
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+        return backoff_expiration_time_;
+      }
+      Timestamp data_expiration_time() const
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+        return data_expiration_time_;
+      }
+      const std::string& header_data() const
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+        return header_data_;
+      }
+      Timestamp stale_time() const ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+        return stale_time_;
+      }
+      Timestamp min_expiration_time() const
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+        return min_expiration_time_;
+      }
+
+      std::unique_ptr<BackOff> TakeBackoffState()
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+        return std::move(backoff_state_);
+      }
+
+      // Cache size of entry.
+      size_t Size() const ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+      // Pick subchannel for request based on the entry's state.
+      PickResult Pick(PickArgs args) ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+      // If the cache entry is in backoff state, resets the backoff and, if
+      // applicable, its backoff timer. The method does not update the LB
+      // policy's picker; the caller is responsible for that if necessary.
+      void ResetBackoff() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+      // Check if the entry should be removed by the clean-up timer.
+      bool ShouldRemove() const ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+      // Check if the entry can be evicted from the cache, i.e. the
+      // min_expiration_time_ has passed.
+      bool CanEvict() const ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+      // Updates the entry upon reception of a new RLS response.
+      // Returns a list of child policy wrappers on which FinishUpdate()
+      // needs to be called after releasing the lock.
+      std::vector<ChildPolicyWrapper*> OnRlsResponseLocked(
+          ResponseInfo response, std::unique_ptr<BackOff> backoff_state)
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+      // Moves entry to the end of the LRU list.
+      void MarkUsed() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+     private:
+      class BackoffTimer : public InternallyRefCounted<BackoffTimer> {
+       public:
+        BackoffTimer(RefCountedPtr<Entry> entry, Timestamp backoff_time);
+
+        // Note: We are forced to disable lock analysis here because
+        // Orphan() is called by OrphanablePtr<>, which cannot have lock
+        // annotations for this particular caller.
+        void Orphan() override ABSL_NO_THREAD_SAFETY_ANALYSIS;
+
+       private:
+        void OnBackoffTimerLocked();
+
+        RefCountedPtr<Entry> entry_;
+        absl::optional<EventEngine::TaskHandle> backoff_timer_task_handle_
+            ABSL_GUARDED_BY(&RlsLb::mu_);
+      };
+
+      RefCountedPtr<RlsLb> lb_policy_;
+
+      bool is_shutdown_ ABSL_GUARDED_BY(&RlsLb::mu_) = false;
+
+      // Backoff states
+      absl::Status status_ ABSL_GUARDED_BY(&RlsLb::mu_);
+      std::unique_ptr<BackOff> backoff_state_ ABSL_GUARDED_BY(&RlsLb::mu_);
+      Timestamp backoff_time_ ABSL_GUARDED_BY(&RlsLb::mu_) =
+          Timestamp::InfPast();
+      Timestamp backoff_expiration_time_ ABSL_GUARDED_BY(&RlsLb::mu_) =
+          Timestamp::InfPast();
+      OrphanablePtr<BackoffTimer> backoff_timer_;
+
+      // RLS response states
+      std::vector<RefCountedPtr<ChildPolicyWrapper>> child_policy_wrappers_
+          ABSL_GUARDED_BY(&RlsLb::mu_);
+      std::string header_data_ ABSL_GUARDED_BY(&RlsLb::mu_);
+      Timestamp data_expiration_time_ ABSL_GUARDED_BY(&RlsLb::mu_) =
+          Timestamp::InfPast();
+      Timestamp stale_time_ ABSL_GUARDED_BY(&RlsLb::mu_) = Timestamp::InfPast();
+
+      Timestamp min_expiration_time_ ABSL_GUARDED_BY(&RlsLb::mu_);
+      Cache::Iterator lru_iterator_ ABSL_GUARDED_BY(&RlsLb::mu_);
+    };
+
+    explicit Cache(RlsLb* lb_policy);
+
+    // Finds an entry from the cache that corresponds to a key. If an entry is
+    // not found, nullptr is returned. Otherwise, the entry is considered
+    // recently used and its order in the LRU list of the cache is updated.
+    Entry* Find(const RequestKey& key)
+        ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+    // Finds an entry from the cache that corresponds to a key. If an entry is
+    // not found, an entry is created, inserted in the cache, and returned to
+    // the caller. Otherwise, the entry found is returned to the caller. The
+    // entry returned to the user is considered recently used and its order in
+    // the LRU list of the cache is updated.
+    Entry* FindOrInsert(const RequestKey& key)
+        ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+    // Resizes the cache. If the new cache size is greater than the current size
+    // of the cache, do nothing. Otherwise, evict the oldest entries that
+    // exceed the new size limit of the cache.
+    void Resize(size_t bytes) ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+    // Resets backoff of all the cache entries.
+    void ResetAllBackoff() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+    // Shutdown the cache; clean-up and orphan all the stored cache entries.
+    void Shutdown() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+   private:
+    // Shared logic for starting the cleanup timer
+    void StartCleanupTimer() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+    void OnCleanupTimer();
+
+    // Returns the entry size for a given key.
+    static size_t EntrySizeForKey(const RequestKey& key);
+
+    // Evicts oversized cache elements when the current size is greater than
+    // the specified limit.
+    void MaybeShrinkSize(size_t bytes)
+        ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+    RlsLb* lb_policy_;
+
+    size_t size_limit_ ABSL_GUARDED_BY(&RlsLb::mu_) = 0;
+    size_t size_ ABSL_GUARDED_BY(&RlsLb::mu_) = 0;
+
+    std::list<RequestKey> lru_list_ ABSL_GUARDED_BY(&RlsLb::mu_);
+    std::unordered_map<RequestKey, OrphanablePtr<Entry>, absl::Hash<RequestKey>>
+        map_ ABSL_GUARDED_BY(&RlsLb::mu_);
+    absl::optional<EventEngine::TaskHandle> cleanup_timer_handle_;
+  };
+
+  // Channel for communicating with the RLS server.
+  // Contains throttling logic for RLS requests.
+  class RlsChannel : public InternallyRefCounted<RlsChannel> {
+   public:
+    explicit RlsChannel(RefCountedPtr<RlsLb> lb_policy);
+
+    // Shuts down the channel.
+    void Orphan() override;
+
+    // Starts an RLS call.
+    // If stale_entry is non-null, it points to the entry containing
+    // stale data for the key.
+    void StartRlsCall(const RequestKey& key, Cache::Entry* stale_entry)
+        ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+    // Reports the result of an RLS call to the throttle.
+    void ReportResponseLocked(bool response_succeeded)
+        ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+    // Checks if a proposed RLS call should be throttled.
+    bool ShouldThrottle() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+      return throttle_.ShouldThrottle();
+    }
+
+    // Resets the channel's backoff.
+    void ResetBackoff();
+
+    grpc_channel* channel() const { return channel_; }
+
+   private:
+    // Watches the state of the RLS channel. Notifies the LB policy when
+    // the channel was previously in TRANSIENT_FAILURE and then becomes READY.
+    class StateWatcher : public AsyncConnectivityStateWatcherInterface {
+     public:
+      explicit StateWatcher(RefCountedPtr<RlsChannel> rls_channel)
+          : AsyncConnectivityStateWatcherInterface(
+                rls_channel->lb_policy_->work_serializer()),
+            rls_channel_(std::move(rls_channel)) {}
+
+     private:
+      void OnConnectivityStateChange(grpc_connectivity_state new_state,
+                                     const absl::Status& status) override;
+
+      RefCountedPtr<RlsChannel> rls_channel_;
+      bool was_transient_failure_ = false;
+    };
+
+    // Throttle state for RLS requests.
+    class Throttle {
+     public:
+      explicit Throttle(
+          Duration window_size = kDefaultThrottleWindowSize,
+          float ratio_for_successes = kDefaultThrottleRatioForSuccesses,
+          int padding = kDefaultThrottlePadding)
+          : window_size_(window_size),
+            ratio_for_successes_(ratio_for_successes),
+            padding_(padding) {}
+
+      bool ShouldThrottle() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+      void RegisterResponse(bool success)
+          ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_);
+
+     private:
+      Duration window_size_;
+      double ratio_for_successes_;
+      int padding_;
+      std::mt19937 rng_{std::random_device()()};
+
+      // Logged timestamp of requests.
+      std::deque<Timestamp> requests_ ABSL_GUARDED_BY(&RlsLb::mu_);
+
+      // Logged timestamps of failures.
+      std::deque<Timestamp> failures_ ABSL_GUARDED_BY(&RlsLb::mu_);
+    };
+
+    RefCountedPtr<RlsLb> lb_policy_;
+    bool is_shutdown_ = false;
+
+    grpc_channel* channel_ = nullptr;
+    RefCountedPtr<channelz::ChannelNode> parent_channelz_node_;
+    StateWatcher* watcher_ = nullptr;
+    Throttle throttle_ ABSL_GUARDED_BY(&RlsLb::mu_);
+  };
+
+  // A pending RLS request.  Instances will be tracked in request_map_.
+  class RlsRequest : public InternallyRefCounted<RlsRequest> {
+   public:
+    // Asynchronously starts a call on rls_channel for key.
+    // Stores backoff_state, which will be transferred to the data cache
+    // if the RLS request fails.
+    RlsRequest(RefCountedPtr<RlsLb> lb_policy, RlsLb::RequestKey key,
+               RefCountedPtr<RlsChannel> rls_channel,
+               std::unique_ptr<BackOff> backoff_state,
+               grpc_lookup_v1_RouteLookupRequest_Reason reason,
+               std::string stale_header_data);
+    ~RlsRequest() override;
+
+    // Shuts down the request.  If the request is still in flight, it is
+    // cancelled, in which case no response will be added to the cache.
+    void Orphan() override;
+
+   private:
+    // Callback to be invoked to start the call.
+    static void StartCall(void* arg, grpc_error_handle error);
+
+    // Helper for StartCall() that runs within the WorkSerializer.
+    void StartCallLocked();
+
+    // Callback to be invoked when the call is completed.
+    static void OnRlsCallComplete(void* arg, grpc_error_handle error);
+
+    // Call completion callback running on LB policy WorkSerializer.
+    void OnRlsCallCompleteLocked(grpc_error_handle error);
+
+    grpc_byte_buffer* MakeRequestProto();
+    ResponseInfo ParseResponseProto();
+
+    RefCountedPtr<RlsLb> lb_policy_;
+    RlsLb::RequestKey key_;
+    RefCountedPtr<RlsChannel> rls_channel_;
+    std::unique_ptr<BackOff> backoff_state_;
+    grpc_lookup_v1_RouteLookupRequest_Reason reason_;
+    std::string stale_header_data_;
+
+    // RLS call state.
+    Timestamp deadline_;
+    grpc_closure call_start_cb_;
+    grpc_closure call_complete_cb_;
+    grpc_call* call_ = nullptr;
+    grpc_byte_buffer* send_message_ = nullptr;
+    grpc_metadata_array recv_initial_metadata_;
+    grpc_byte_buffer* recv_message_ = nullptr;
+    grpc_metadata_array recv_trailing_metadata_;
+    grpc_status_code status_recv_;
+    grpc_slice status_details_recv_;
+  };
+
+  void ShutdownLocked() override;
+
+  // Returns a new picker to the channel to trigger reprocessing of
+  // pending picks.  Schedules the actual picker update on the ExecCtx
+  // to be run later, so it's safe to invoke this while holding the lock.
+  void UpdatePickerAsync();
+  // Hops into work serializer and calls UpdatePickerLocked().
+  static void UpdatePickerCallback(void* arg, grpc_error_handle error);
+  // Updates the picker in the work serializer.
+  void UpdatePickerLocked() ABSL_LOCKS_EXCLUDED(&mu_);
+
+  // Mutex to guard LB policy state that is accessed by the picker.
+  Mutex mu_;
+  bool is_shutdown_ ABSL_GUARDED_BY(mu_) = false;
+  bool update_in_progress_ = false;
+  Cache cache_ ABSL_GUARDED_BY(mu_);
+  // Maps an RLS request key to an RlsRequest object that represents a pending
+  // RLS request.
+  std::unordered_map<RequestKey, OrphanablePtr<RlsRequest>,
+                     absl::Hash<RequestKey>>
+      request_map_ ABSL_GUARDED_BY(mu_);
+  // The channel on which RLS requests are sent.
+  // Note that this channel may be swapped out when the RLS policy gets
+  // an update.  However, when that happens, any existing entries in
+  // request_map_ will continue to use the previous channel.
+  OrphanablePtr<RlsChannel> rls_channel_ ABSL_GUARDED_BY(mu_);
+
+  // Accessed only from within WorkSerializer.
+  absl::StatusOr<std::shared_ptr<EndpointAddressesIterator>> addresses_;
+  ChannelArgs channel_args_;
+  RefCountedPtr<RlsLbConfig> config_;
+  RefCountedPtr<ChildPolicyWrapper> default_child_policy_;
+  std::map<std::string /*target*/, ChildPolicyWrapper*> child_policy_map_;
+};
+
+//
+// RlsLb::ChildPolicyWrapper
+//
+
+RlsLb::ChildPolicyWrapper::ChildPolicyWrapper(RefCountedPtr<RlsLb> lb_policy,
+                                              std::string target)
+    : DualRefCounted<ChildPolicyWrapper>(
+          GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace) ? "ChildPolicyWrapper"
+                                                     : nullptr),
+      lb_policy_(std::move(lb_policy)),
+      target_(std::move(target)),
+      picker_(MakeRefCounted<QueuePicker>(nullptr)) {
+  lb_policy_->child_policy_map_.emplace(target_, this);
+}
+
+void RlsLb::ChildPolicyWrapper::Orphan() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] ChildPolicyWrapper=%p [%s]: shutdown",
+            lb_policy_.get(), this, target_.c_str());
+  }
+  is_shutdown_ = true;
+  lb_policy_->child_policy_map_.erase(target_);
+  if (child_policy_ != nullptr) {
+    grpc_pollset_set_del_pollset_set(child_policy_->interested_parties(),
+                                     lb_policy_->interested_parties());
+    child_policy_.reset();
+  }
+  picker_.reset();
+}
+
+absl::optional<Json> InsertOrUpdateChildPolicyField(const std::string& field,
+                                                    const std::string& value,
+                                                    const Json& config,
+                                                    ValidationErrors* errors) {
+  if (config.type() != Json::Type::kArray) {
+    errors->AddError("is not an array");
+    return absl::nullopt;
+  }
+  const size_t original_num_errors = errors->size();
+  Json::Array array;
+  for (size_t i = 0; i < config.array().size(); ++i) {
+    const Json& child_json = config.array()[i];
+    ValidationErrors::ScopedField json_field(errors, absl::StrCat("[", i, "]"));
+    if (child_json.type() != Json::Type::kObject) {
+      errors->AddError("is not an object");
+    } else {
+      const Json::Object& child = child_json.object();
+      if (child.size() != 1) {
+        errors->AddError("child policy object contains more than one field");
+      } else {
+        const std::string& child_name = child.begin()->first;
+        ValidationErrors::ScopedField json_field(
+            errors, absl::StrCat("[\"", child_name, "\"]"));
+        const Json& child_config_json = child.begin()->second;
+        if (child_config_json.type() != Json::Type::kObject) {
+          errors->AddError("child policy config is not an object");
+        } else {
+          Json::Object child_config = child_config_json.object();
+          child_config[field] = Json::FromString(value);
+          array.emplace_back(Json::FromObject(
+              {{child_name, Json::FromObject(std::move(child_config))}}));
+        }
+      }
+    }
+  }
+  if (errors->size() != original_num_errors) return absl::nullopt;
+  return Json::FromArray(std::move(array));
+}
+
+void RlsLb::ChildPolicyWrapper::StartUpdate() {
+  ValidationErrors errors;
+  auto child_policy_config = InsertOrUpdateChildPolicyField(
+      lb_policy_->config_->child_policy_config_target_field_name(), target_,
+      lb_policy_->config_->child_policy_config(), &errors);
+  GPR_ASSERT(child_policy_config.has_value());
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(
+        GPR_INFO,
+        "[rlslb %p] ChildPolicyWrapper=%p [%s]: validating update, config: %s",
+        lb_policy_.get(), this, target_.c_str(),
+        JsonDump(*child_policy_config).c_str());
+  }
+  auto config =
+      CoreConfiguration::Get().lb_policy_registry().ParseLoadBalancingConfig(
+          *child_policy_config);
+  // Returned RLS target fails the validation.
+  if (!config.ok()) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO,
+              "[rlslb %p] ChildPolicyWrapper=%p [%s]: config failed to parse: "
+              "%s",
+              lb_policy_.get(), this, target_.c_str(),
+              config.status().ToString().c_str());
+    }
+    pending_config_.reset();
+    picker_ = MakeRefCounted<TransientFailurePicker>(
+        absl::UnavailableError(config.status().message()));
+    child_policy_.reset();
+  } else {
+    pending_config_ = std::move(*config);
+  }
+}
+
+absl::Status RlsLb::ChildPolicyWrapper::MaybeFinishUpdate() {
+  // If pending_config_ is not set, that means StartUpdate() failed, so
+  // there's nothing to do here.
+  if (pending_config_ == nullptr) return absl::OkStatus();
+  // If child policy doesn't yet exist, create it.
+  if (child_policy_ == nullptr) {
+    Args create_args;
+    create_args.work_serializer = lb_policy_->work_serializer();
+    create_args.channel_control_helper = std::make_unique<ChildPolicyHelper>(
+        WeakRef(DEBUG_LOCATION, "ChildPolicyHelper"));
+    create_args.args = lb_policy_->channel_args_;
+    child_policy_ = MakeOrphanable<ChildPolicyHandler>(std::move(create_args),
+                                                       &grpc_lb_rls_trace);
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO,
+              "[rlslb %p] ChildPolicyWrapper=%p [%s], created new child policy "
+              "handler %p",
+              lb_policy_.get(), this, target_.c_str(), child_policy_.get());
+    }
+    grpc_pollset_set_add_pollset_set(child_policy_->interested_parties(),
+                                     lb_policy_->interested_parties());
+  }
+  // Send the child the updated config.
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO,
+            "[rlslb %p] ChildPolicyWrapper=%p [%s], updating child policy "
+            "handler %p",
+            lb_policy_.get(), this, target_.c_str(), child_policy_.get());
+  }
+  UpdateArgs update_args;
+  update_args.config = std::move(pending_config_);
+  update_args.addresses = lb_policy_->addresses_;
+  update_args.args = lb_policy_->channel_args_;
+  return child_policy_->UpdateLocked(std::move(update_args));
+}
+
+//
+// RlsLb::ChildPolicyWrapper::ChildPolicyHelper
+//
+
+void RlsLb::ChildPolicyWrapper::ChildPolicyHelper::UpdateState(
+    grpc_connectivity_state state, const absl::Status& status,
+    RefCountedPtr<SubchannelPicker> picker) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO,
+            "[rlslb %p] ChildPolicyWrapper=%p [%s] ChildPolicyHelper=%p: "
+            "UpdateState(state=%s, status=%s, picker=%p)",
+            wrapper_->lb_policy_.get(), wrapper_.get(),
+            wrapper_->target_.c_str(), this, ConnectivityStateName(state),
+            status.ToString().c_str(), picker.get());
+  }
+  {
+    MutexLock lock(&wrapper_->lb_policy_->mu_);
+    if (wrapper_->is_shutdown_) return;
+    // TODO(roth): It looks like this ignores subsequent TF updates that
+    // might change the status used to fail picks, which seems wrong.
+    if (wrapper_->connectivity_state_ == GRPC_CHANNEL_TRANSIENT_FAILURE &&
+        state != GRPC_CHANNEL_READY) {
+      return;
+    }
+    wrapper_->connectivity_state_ = state;
+    GPR_DEBUG_ASSERT(picker != nullptr);
+    if (picker != nullptr) {
+      wrapper_->picker_ = std::move(picker);
+    }
+  }
+  wrapper_->lb_policy_->UpdatePickerLocked();
+}
+
+//
+// RlsLb::Picker
+//
+
+// Builds the key to be used for a request based on path and initial_metadata.
+std::map<std::string, std::string> BuildKeyMap(
+    const RlsLbConfig::KeyBuilderMap& key_builder_map, absl::string_view path,
+    absl::string_view host,
+    const LoadBalancingPolicy::MetadataInterface* initial_metadata) {
+  size_t last_slash_pos = path.npos;  // May need this a few times, so cache it.
+  // Find key builder for this path.
+  auto it = key_builder_map.find(std::string(path));
+  if (it == key_builder_map.end()) {
+    // Didn't find exact match, try method wildcard.
+    last_slash_pos = path.rfind('/');
+    GPR_DEBUG_ASSERT(last_slash_pos != path.npos);
+    if (GPR_UNLIKELY(last_slash_pos == path.npos)) return {};
+    std::string service(path.substr(0, last_slash_pos + 1));
+    it = key_builder_map.find(service);
+    if (it == key_builder_map.end()) return {};
+  }
+  const RlsLbConfig::KeyBuilder* key_builder = &it->second;
+  // Construct key map using key builder.
+  std::map<std::string, std::string> key_map;
+  // Add header keys.
+  for (const auto& p : key_builder->header_keys) {
+    const std::string& key = p.first;
+    const std::vector<std::string>& header_names = p.second;
+    for (const std::string& header_name : header_names) {
+      std::string buffer;
+      absl::optional<absl::string_view> value =
+          initial_metadata->Lookup(header_name, &buffer);
+      if (value.has_value()) {
+        key_map[key] = std::string(*value);
+        break;
+      }
+    }
+  }
+  // Add constant keys.
+  key_map.insert(key_builder->constant_keys.begin(),
+                 key_builder->constant_keys.end());
+  // Add host key.
+  if (!key_builder->host_key.empty()) {
+    key_map[key_builder->host_key] = std::string(host);
+  }
+  // Add service key.
+  if (!key_builder->service_key.empty()) {
+    if (last_slash_pos == path.npos) {
+      last_slash_pos = path.rfind('/');
+      GPR_DEBUG_ASSERT(last_slash_pos != path.npos);
+      if (GPR_UNLIKELY(last_slash_pos == path.npos)) return {};
+    }
+    key_map[key_builder->service_key] =
+        std::string(path.substr(1, last_slash_pos - 1));
+  }
+  // Add method key.
+  if (!key_builder->method_key.empty()) {
+    if (last_slash_pos == path.npos) {
+      last_slash_pos = path.rfind('/');
+      GPR_DEBUG_ASSERT(last_slash_pos != path.npos);
+      if (GPR_UNLIKELY(last_slash_pos == path.npos)) return {};
+    }
+    key_map[key_builder->method_key] =
+        std::string(path.substr(last_slash_pos + 1));
+  }
+  return key_map;
+}
+
+RlsLb::Picker::Picker(RefCountedPtr<RlsLb> lb_policy)
+    : lb_policy_(std::move(lb_policy)), config_(lb_policy_->config_) {
+  if (lb_policy_->default_child_policy_ != nullptr) {
+    default_child_policy_ =
+        lb_policy_->default_child_policy_->Ref(DEBUG_LOCATION, "Picker");
+  }
+}
+
+LoadBalancingPolicy::PickResult RlsLb::Picker::Pick(PickArgs args) {
+  // Construct key for request.
+  RequestKey key = {
+      BuildKeyMap(config_->key_builder_map(), args.path,
+                  lb_policy_->channel_control_helper()->GetAuthority(),
+                  args.initial_metadata)};
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] picker=%p: request keys: %s",
+            lb_policy_.get(), this, key.ToString().c_str());
+  }
+  Timestamp now = Timestamp::Now();
+  MutexLock lock(&lb_policy_->mu_);
+  if (lb_policy_->is_shutdown_) {
+    return PickResult::Fail(
+        absl::UnavailableError("LB policy already shut down"));
+  }
+  // Check if there's a cache entry.
+  Cache::Entry* entry = lb_policy_->cache_.Find(key);
+  // If there is no cache entry, or if the cache entry is not in backoff
+  // and has a stale time in the past, and there is not already a
+  // pending RLS request for this key, then try to start a new RLS request.
+  if ((entry == nullptr ||
+       (entry->stale_time() < now && entry->backoff_time() < now)) &&
+      lb_policy_->request_map_.find(key) == lb_policy_->request_map_.end()) {
+    // Check if requests are being throttled.
+    if (lb_policy_->rls_channel_->ShouldThrottle()) {
+      // Request is throttled.
+      // If there is no non-expired data in the cache, then we use the
+      // default target if set, or else we fail the pick.
+      if (entry == nullptr || entry->data_expiration_time() < now) {
+        if (default_child_policy_ != nullptr) {
+          if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+            gpr_log(GPR_INFO,
+                    "[rlslb %p] picker=%p: RLS call throttled; "
+                    "using default target",
+                    lb_policy_.get(), this);
+          }
+          return default_child_policy_->Pick(args);
+        }
+        if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+          gpr_log(GPR_INFO,
+                  "[rlslb %p] picker=%p: RLS call throttled; failing pick",
+                  lb_policy_.get(), this);
+        }
+        return PickResult::Fail(
+            absl::UnavailableError("RLS request throttled"));
+      }
+    }
+    // Start the RLS call.
+    lb_policy_->rls_channel_->StartRlsCall(
+        key, (entry == nullptr || entry->data_expiration_time() < now) ? nullptr
+                                                                       : entry);
+  }
+  // If the cache entry exists, see if it has usable data.
+  if (entry != nullptr) {
+    // If the entry has non-expired data, use it.
+    if (entry->data_expiration_time() >= now) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+        gpr_log(GPR_INFO, "[rlslb %p] picker=%p: using cache entry %p",
+                lb_policy_.get(), this, entry);
+      }
+      return entry->Pick(args);
+    }
+    // If the entry is in backoff, then use the default target if set,
+    // or else fail the pick.
+    if (entry->backoff_time() >= now) {
+      if (default_child_policy_ != nullptr) {
+        if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+          gpr_log(
+              GPR_INFO,
+              "[rlslb %p] picker=%p: RLS call in backoff; using default target",
+              lb_policy_.get(), this);
+        }
+        return default_child_policy_->Pick(args);
+      }
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+        gpr_log(GPR_INFO,
+                "[rlslb %p] picker=%p: RLS call in backoff; failing pick",
+                lb_policy_.get(), this);
+      }
+      return PickResult::Fail(absl::UnavailableError(
+          absl::StrCat("RLS request failed: ", entry->status().ToString())));
+    }
+  }
+  // RLS call pending.  Queue the pick.
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] picker=%p: RLS request pending; queuing pick",
+            lb_policy_.get(), this);
+  }
+  return PickResult::Queue();
+}
+
+//
+// RlsLb::Cache::Entry::BackoffTimer
+//
+
+RlsLb::Cache::Entry::BackoffTimer::BackoffTimer(RefCountedPtr<Entry> entry,
+                                                Timestamp backoff_time)
+    : entry_(std::move(entry)) {
+  backoff_timer_task_handle_ =
+      entry_->lb_policy_->channel_control_helper()->GetEventEngine()->RunAfter(
+          backoff_time - Timestamp::Now(),
+          [self = Ref(DEBUG_LOCATION, "BackoffTimer")]() mutable {
+            ApplicationCallbackExecCtx callback_exec_ctx;
+            ExecCtx exec_ctx;
+            auto self_ptr = self.get();
+            self_ptr->entry_->lb_policy_->work_serializer()->Run(
+                [self = std::move(self)]() { self->OnBackoffTimerLocked(); },
+                DEBUG_LOCATION);
+          });
+}
+
+void RlsLb::Cache::Entry::BackoffTimer::Orphan() {
+  if (backoff_timer_task_handle_.has_value() &&
+      entry_->lb_policy_->channel_control_helper()->GetEventEngine()->Cancel(
+          *backoff_timer_task_handle_)) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO, "[rlslb %p] cache entry=%p %s, backoff timer canceled",
+              entry_->lb_policy_.get(), entry_.get(),
+              entry_->is_shutdown_ ? "(shut down)"
+                                   : entry_->lru_iterator_->ToString().c_str());
+    }
+  }
+  backoff_timer_task_handle_.reset();
+  Unref(DEBUG_LOCATION, "Orphan");
+}
+
+void RlsLb::Cache::Entry::BackoffTimer::OnBackoffTimerLocked() {
+  {
+    MutexLock lock(&entry_->lb_policy_->mu_);
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO, "[rlslb %p] cache entry=%p %s, backoff timer fired",
+              entry_->lb_policy_.get(), entry_.get(),
+              entry_->is_shutdown_ ? "(shut down)"
+                                   : entry_->lru_iterator_->ToString().c_str());
+    }
+    // Skip the update if Orphaned
+    if (!backoff_timer_task_handle_.has_value()) return;
+    backoff_timer_task_handle_.reset();
+  }
+  // The pick was in backoff state and there could be a pick queued if
+  // wait_for_ready is true. We'll update the picker for that case.
+  entry_->lb_policy_->UpdatePickerLocked();
+}
+
+//
+// RlsLb::Cache::Entry
+//
+
+std::unique_ptr<BackOff> MakeCacheEntryBackoff() {
+  return std::make_unique<BackOff>(
+      BackOff::Options()
+          .set_initial_backoff(kCacheBackoffInitial)
+          .set_multiplier(kCacheBackoffMultiplier)
+          .set_jitter(kCacheBackoffJitter)
+          .set_max_backoff(kCacheBackoffMax));
+}
+
+RlsLb::Cache::Entry::Entry(RefCountedPtr<RlsLb> lb_policy,
+                           const RequestKey& key)
+    : InternallyRefCounted<Entry>(
+          GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace) ? "CacheEntry" : nullptr),
+      lb_policy_(std::move(lb_policy)),
+      backoff_state_(MakeCacheEntryBackoff()),
+      min_expiration_time_(Timestamp::Now() + kMinExpirationTime),
+      lru_iterator_(lb_policy_->cache_.lru_list_.insert(
+          lb_policy_->cache_.lru_list_.end(), key)) {}
+
+void RlsLb::Cache::Entry::Orphan() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] cache entry=%p %s: cache entry evicted",
+            lb_policy_.get(), this, lru_iterator_->ToString().c_str());
+  }
+  is_shutdown_ = true;
+  lb_policy_->cache_.lru_list_.erase(lru_iterator_);
+  lru_iterator_ = lb_policy_->cache_.lru_list_.end();  // Just in case.
+  backoff_state_.reset();
+  if (backoff_timer_ != nullptr) {
+    backoff_timer_.reset();
+    lb_policy_->UpdatePickerAsync();
+  }
+  child_policy_wrappers_.clear();
+  Unref(DEBUG_LOCATION, "Orphan");
+}
+
+size_t RlsLb::Cache::Entry::Size() const {
+  // lru_iterator_ is not valid once we're shut down.
+  GPR_ASSERT(!is_shutdown_);
+  return lb_policy_->cache_.EntrySizeForKey(*lru_iterator_);
+}
+
+LoadBalancingPolicy::PickResult RlsLb::Cache::Entry::Pick(PickArgs args) {
+  size_t i = 0;
+  ChildPolicyWrapper* child_policy_wrapper = nullptr;
+  // Skip targets before the last one that are in state TRANSIENT_FAILURE.
+  for (; i < child_policy_wrappers_.size(); ++i) {
+    child_policy_wrapper = child_policy_wrappers_[i].get();
+    if (child_policy_wrapper->connectivity_state() ==
+            GRPC_CHANNEL_TRANSIENT_FAILURE &&
+        i < child_policy_wrappers_.size() - 1) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+        gpr_log(GPR_INFO,
+                "[rlslb %p] cache entry=%p %s: target %s (%" PRIuPTR
+                " of %" PRIuPTR ") in state TRANSIENT_FAILURE; skipping",
+                lb_policy_.get(), this, lru_iterator_->ToString().c_str(),
+                child_policy_wrapper->target().c_str(), i,
+                child_policy_wrappers_.size());
+      }
+      continue;
+    }
+    break;
+  }
+  // Child policy not in TRANSIENT_FAILURE or is the last target in
+  // the list, so delegate.
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO,
+            "[rlslb %p] cache entry=%p %s: target %s (%" PRIuPTR " of %" PRIuPTR
+            ") in state %s; delegating",
+            lb_policy_.get(), this, lru_iterator_->ToString().c_str(),
+            child_policy_wrapper->target().c_str(), i,
+            child_policy_wrappers_.size(),
+            ConnectivityStateName(child_policy_wrapper->connectivity_state()));
+  }
+  // Add header data.
+  // Note that even if the target we're using is in TRANSIENT_FAILURE,
+  // the pick might still succeed (e.g., if the child is ring_hash), so
+  // we need to pass the right header info down in all cases.
+  if (!header_data_.empty()) {
+    char* copied_header_data =
+        static_cast<char*>(args.call_state->Alloc(header_data_.length() + 1));
+    strcpy(copied_header_data, header_data_.c_str());
+    args.initial_metadata->Add(kRlsHeaderKey, copied_header_data);
+  }
+  return child_policy_wrapper->Pick(args);
+}
+
+void RlsLb::Cache::Entry::ResetBackoff() {
+  backoff_time_ = Timestamp::InfPast();
+  backoff_timer_.reset();
+}
+
+bool RlsLb::Cache::Entry::ShouldRemove() const {
+  Timestamp now = Timestamp::Now();
+  return data_expiration_time_ < now && backoff_expiration_time_ < now;
+}
+
+bool RlsLb::Cache::Entry::CanEvict() const {
+  Timestamp now = Timestamp::Now();
+  return min_expiration_time_ < now;
+}
+
+void RlsLb::Cache::Entry::MarkUsed() {
+  auto& lru_list = lb_policy_->cache_.lru_list_;
+  auto new_it = lru_list.insert(lru_list.end(), *lru_iterator_);
+  lru_list.erase(lru_iterator_);
+  lru_iterator_ = new_it;
+}
+
+std::vector<RlsLb::ChildPolicyWrapper*>
+RlsLb::Cache::Entry::OnRlsResponseLocked(
+    ResponseInfo response, std::unique_ptr<BackOff> backoff_state) {
+  // Move the entry to the end of the LRU list.
+  MarkUsed();
+  // If the request failed, store the failed status and update the
+  // backoff state.
+  if (!response.status.ok()) {
+    status_ = response.status;
+    if (backoff_state != nullptr) {
+      backoff_state_ = std::move(backoff_state);
+    } else {
+      backoff_state_ = MakeCacheEntryBackoff();
+    }
+    backoff_time_ = backoff_state_->NextAttemptTime();
+    Timestamp now = Timestamp::Now();
+    backoff_expiration_time_ = now + (backoff_time_ - now) * 2;
+    backoff_timer_ = MakeOrphanable<BackoffTimer>(
+        Ref(DEBUG_LOCATION, "BackoffTimer"), backoff_time_);
+    lb_policy_->UpdatePickerAsync();
+    return {};
+  }
+  // Request succeeded, so store the result.
+  header_data_ = std::move(response.header_data);
+  Timestamp now = Timestamp::Now();
+  data_expiration_time_ = now + lb_policy_->config_->max_age();
+  stale_time_ = now + lb_policy_->config_->stale_age();
+  status_ = absl::OkStatus();
+  backoff_state_.reset();
+  backoff_time_ = Timestamp::InfPast();
+  backoff_expiration_time_ = Timestamp::InfPast();
+  // Check if we need to update this list of targets.
+  bool targets_changed = [&]() ABSL_EXCLUSIVE_LOCKS_REQUIRED(&RlsLb::mu_) {
+    if (child_policy_wrappers_.size() != response.targets.size()) return true;
+    for (size_t i = 0; i < response.targets.size(); ++i) {
+      if (child_policy_wrappers_[i]->target() != response.targets[i]) {
+        return true;
+      }
+    }
+    return false;
+  }();
+  if (!targets_changed) {
+    // Targets didn't change, so we're not updating the list of child
+    // policies.  Return a new picker so that any queued requests can be
+    // re-processed.
+    lb_policy_->UpdatePickerAsync();
+    return {};
+  }
+  // Target list changed, so update it.
+  std::set<absl::string_view> old_targets;
+  for (RefCountedPtr<ChildPolicyWrapper>& child_policy_wrapper :
+       child_policy_wrappers_) {
+    old_targets.emplace(child_policy_wrapper->target());
+  }
+  bool update_picker = false;
+  std::vector<ChildPolicyWrapper*> child_policies_to_finish_update;
+  std::vector<RefCountedPtr<ChildPolicyWrapper>> new_child_policy_wrappers;
+  new_child_policy_wrappers.reserve(response.targets.size());
+  for (std::string& target : response.targets) {
+    auto it = lb_policy_->child_policy_map_.find(target);
+    if (it == lb_policy_->child_policy_map_.end()) {
+      auto new_child = MakeRefCounted<ChildPolicyWrapper>(
+          lb_policy_.Ref(DEBUG_LOCATION, "ChildPolicyWrapper"), target);
+      new_child->StartUpdate();
+      child_policies_to_finish_update.push_back(new_child.get());
+      new_child_policy_wrappers.emplace_back(std::move(new_child));
+    } else {
+      new_child_policy_wrappers.emplace_back(
+          it->second->Ref(DEBUG_LOCATION, "CacheEntry"));
+      // If the target already existed but was not previously used for
+      // this key, then we'll need to update the picker, since we
+      // didn't actually create a new child policy, which would have
+      // triggered an RLS picker update when it returned its first picker.
+      if (old_targets.find(target) == old_targets.end()) {
+        update_picker = true;
+      }
+    }
+  }
+  child_policy_wrappers_ = std::move(new_child_policy_wrappers);
+  if (update_picker) {
+    lb_policy_->UpdatePickerAsync();
+  }
+  return child_policies_to_finish_update;
+}
+
+//
+// RlsLb::Cache
+//
+
+RlsLb::Cache::Cache(RlsLb* lb_policy) : lb_policy_(lb_policy) {
+  StartCleanupTimer();
+}
+
+RlsLb::Cache::Entry* RlsLb::Cache::Find(const RequestKey& key) {
+  auto it = map_.find(key);
+  if (it == map_.end()) return nullptr;
+  it->second->MarkUsed();
+  return it->second.get();
+}
+
+RlsLb::Cache::Entry* RlsLb::Cache::FindOrInsert(const RequestKey& key) {
+  auto it = map_.find(key);
+  // If not found, create new entry.
+  if (it == map_.end()) {
+    size_t entry_size = EntrySizeForKey(key);
+    MaybeShrinkSize(size_limit_ - std::min(size_limit_, entry_size));
+    Entry* entry = new Entry(
+        lb_policy_->RefAsSubclass<RlsLb>(DEBUG_LOCATION, "CacheEntry"), key);
+    map_.emplace(key, OrphanablePtr<Entry>(entry));
+    size_ += entry_size;
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO, "[rlslb %p] key=%s: cache entry added, entry=%p",
+              lb_policy_, key.ToString().c_str(), entry);
+    }
+    return entry;
+  }
+  // Entry found, so use it.
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] key=%s: found cache entry %p", lb_policy_,
+            key.ToString().c_str(), it->second.get());
+  }
+  it->second->MarkUsed();
+  return it->second.get();
+}
+
+void RlsLb::Cache::Resize(size_t bytes) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] resizing cache to %" PRIuPTR " bytes",
+            lb_policy_, bytes);
+  }
+  size_limit_ = bytes;
+  MaybeShrinkSize(size_limit_);
+}
+
+void RlsLb::Cache::ResetAllBackoff() {
+  for (auto& p : map_) {
+    p.second->ResetBackoff();
+  }
+  lb_policy_->UpdatePickerAsync();
+}
+
+void RlsLb::Cache::Shutdown() {
+  map_.clear();
+  lru_list_.clear();
+  if (cleanup_timer_handle_.has_value() &&
+      lb_policy_->channel_control_helper()->GetEventEngine()->Cancel(
+          *cleanup_timer_handle_)) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO, "[rlslb %p] cache cleanup timer canceled", lb_policy_);
+    }
+  }
+  cleanup_timer_handle_.reset();
+}
+
+void RlsLb::Cache::StartCleanupTimer() {
+  cleanup_timer_handle_ =
+      lb_policy_->channel_control_helper()->GetEventEngine()->RunAfter(
+          kCacheCleanupTimerInterval,
+          [this, lb_policy = lb_policy_->Ref(DEBUG_LOCATION,
+                                             "CacheCleanupTimer")]() mutable {
+            ApplicationCallbackExecCtx callback_exec_ctx;
+            ExecCtx exec_ctx;
+            lb_policy_->work_serializer()->Run(
+                [this, lb_policy = std::move(lb_policy)]() {
+                  // The lb_policy ref is held until the callback completes
+                  OnCleanupTimer();
+                },
+                DEBUG_LOCATION);
+          });
+}
+
+void RlsLb::Cache::OnCleanupTimer() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] cache cleanup timer fired", lb_policy_);
+  }
+  MutexLock lock(&lb_policy_->mu_);
+  if (!cleanup_timer_handle_.has_value()) return;
+  if (lb_policy_->is_shutdown_) return;
+  for (auto it = map_.begin(); it != map_.end();) {
+    if (GPR_UNLIKELY(it->second->ShouldRemove() && it->second->CanEvict())) {
+      size_ -= it->second->Size();
+      it = map_.erase(it);
+    } else {
+      ++it;
+    }
+  }
+  StartCleanupTimer();
+}
+
+size_t RlsLb::Cache::EntrySizeForKey(const RequestKey& key) {
+  // Key is stored twice, once in LRU list and again in the cache map.
+  return (key.Size() * 2) + sizeof(Entry);
+}
+
+void RlsLb::Cache::MaybeShrinkSize(size_t bytes) {
+  while (size_ > bytes) {
+    auto lru_it = lru_list_.begin();
+    if (GPR_UNLIKELY(lru_it == lru_list_.end())) break;
+    auto map_it = map_.find(*lru_it);
+    GPR_ASSERT(map_it != map_.end());
+    if (!map_it->second->CanEvict()) break;
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO, "[rlslb %p] LRU eviction: removing entry %p %s",
+              lb_policy_, map_it->second.get(), lru_it->ToString().c_str());
+    }
+    size_ -= map_it->second->Size();
+    map_.erase(map_it);
+  }
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO,
+            "[rlslb %p] LRU pass complete: desired size=%" PRIuPTR
+            " size=%" PRIuPTR,
+            lb_policy_, bytes, size_);
+  }
+}
+
+//
+// RlsLb::RlsChannel::StateWatcher
+//
+
+void RlsLb::RlsChannel::StateWatcher::OnConnectivityStateChange(
+    grpc_connectivity_state new_state, const absl::Status& status) {
+  auto* lb_policy = rls_channel_->lb_policy_.get();
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO,
+            "[rlslb %p] RlsChannel=%p StateWatcher=%p: "
+            "state changed to %s (%s)",
+            lb_policy, rls_channel_.get(), this,
+            ConnectivityStateName(new_state), status.ToString().c_str());
+  }
+  if (rls_channel_->is_shutdown_) return;
+  MutexLock lock(&lb_policy->mu_);
+  if (new_state == GRPC_CHANNEL_READY && was_transient_failure_) {
+    was_transient_failure_ = false;
+    // Reset the backoff of all cache entries, so that we don't
+    // double-penalize if an RLS request fails while the channel is
+    // down, since the throttling for the channel being down is handled
+    // at the channel level instead of in the individual cache entries.
+    lb_policy->cache_.ResetAllBackoff();
+  } else if (new_state == GRPC_CHANNEL_TRANSIENT_FAILURE) {
+    was_transient_failure_ = true;
+  }
+}
+
+//
+// RlsLb::RlsChannel::Throttle
+//
+
+bool RlsLb::RlsChannel::Throttle::ShouldThrottle() {
+  Timestamp now = Timestamp::Now();
+  while (!requests_.empty() && now - requests_.front() > window_size_) {
+    requests_.pop_front();
+  }
+  while (!failures_.empty() && now - failures_.front() > window_size_) {
+    failures_.pop_front();
+  }
+  // Compute probability of throttling.
+  float num_requests = requests_.size();
+  float num_successes = num_requests - failures_.size();
+  // Note: it's possible that this ratio will be negative, in which case
+  // no throttling will be done.
+  float throttle_probability =
+      (num_requests - (num_successes * ratio_for_successes_)) /
+      (num_requests + padding_);
+  // Generate a random number for the request.
+  std::uniform_real_distribution<float> dist(0, 1.0);
+  // Check if we should throttle the request.
+  bool throttle = dist(rng_) < throttle_probability;
+  // If we're throttling, record the request and the failure.
+  if (throttle) {
+    requests_.push_back(now);
+    failures_.push_back(now);
+  }
+  return throttle;
+}
+
+void RlsLb::RlsChannel::Throttle::RegisterResponse(bool success) {
+  Timestamp now = Timestamp::Now();
+  requests_.push_back(now);
+  if (!success) failures_.push_back(now);
+}
+
+//
+// RlsLb::RlsChannel
+//
+
+RlsLb::RlsChannel::RlsChannel(RefCountedPtr<RlsLb> lb_policy)
+    : InternallyRefCounted<RlsChannel>(
+          GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace) ? "RlsChannel" : nullptr),
+      lb_policy_(std::move(lb_policy)) {
+  // Get channel creds from parent channel.
+  // Note that we are using the "unsafe" channel creds here, which do
+  // include any associated call creds.  This is safe in this case,
+  // because we are using the parent channel's authority on the RLS channel.
+  auto creds =
+      lb_policy_->channel_control_helper()->GetUnsafeChannelCredentials();
+  // Use the parent channel's authority.
+  auto authority = lb_policy_->channel_control_helper()->GetAuthority();
+  ChannelArgs args = ChannelArgs()
+                         .Set(GRPC_ARG_DEFAULT_AUTHORITY, authority)
+                         .Set(GRPC_ARG_CHANNELZ_IS_INTERNAL_CHANNEL, 1);
+  // Propagate fake security connector expected targets, if any.
+  // (This is ugly, but it seems better than propagating all channel args
+  // from the parent channel by default and then having a giant
+  // exclude list of args to strip out, like we do in grpclb.)
+  absl::optional<absl::string_view> fake_security_expected_targets =
+      lb_policy_->channel_args_.GetString(
+          GRPC_ARG_FAKE_SECURITY_EXPECTED_TARGETS);
+  if (fake_security_expected_targets.has_value()) {
+    args = args.Set(GRPC_ARG_FAKE_SECURITY_EXPECTED_TARGETS,
+                    *fake_security_expected_targets);
+  }
+  // Add service config args if needed.
+  const std::string& service_config =
+      lb_policy_->config_->rls_channel_service_config();
+  if (!service_config.empty()) {
+    args = args.Set(GRPC_ARG_SERVICE_CONFIG, service_config)
+               .Set(GRPC_ARG_SERVICE_CONFIG_DISABLE_RESOLUTION, 1);
+  }
+  channel_ = grpc_channel_create(lb_policy_->config_->lookup_service().c_str(),
+                                 creds.get(), args.ToC().get());
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] RlsChannel=%p: created channel %p for %s",
+            lb_policy_.get(), this, channel_,
+            lb_policy_->config_->lookup_service().c_str());
+  }
+  if (channel_ != nullptr) {
+    // Set up channelz linkage.
+    channelz::ChannelNode* child_channelz_node =
+        grpc_channel_get_channelz_node(channel_);
+    auto parent_channelz_node =
+        lb_policy_->channel_args_.GetObjectRef<channelz::ChannelNode>();
+    if (child_channelz_node != nullptr && parent_channelz_node != nullptr) {
+      parent_channelz_node->AddChildChannel(child_channelz_node->uuid());
+      parent_channelz_node_ = std::move(parent_channelz_node);
+    }
+    // Start connectivity watch.
+    ClientChannelFilter* client_channel =
+        ClientChannelFilter::GetFromChannel(Channel::FromC(channel_));
+    GPR_ASSERT(client_channel != nullptr);
+    watcher_ = new StateWatcher(Ref(DEBUG_LOCATION, "StateWatcher"));
+    client_channel->AddConnectivityWatcher(
+        GRPC_CHANNEL_IDLE,
+        OrphanablePtr<AsyncConnectivityStateWatcherInterface>(watcher_));
+  }
+}
+
+void RlsLb::RlsChannel::Orphan() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] RlsChannel=%p, channel=%p: shutdown",
+            lb_policy_.get(), this, channel_);
+  }
+  is_shutdown_ = true;
+  if (channel_ != nullptr) {
+    // Remove channelz linkage.
+    if (parent_channelz_node_ != nullptr) {
+      channelz::ChannelNode* child_channelz_node =
+          grpc_channel_get_channelz_node(channel_);
+      GPR_ASSERT(child_channelz_node != nullptr);
+      parent_channelz_node_->RemoveChildChannel(child_channelz_node->uuid());
+    }
+    // Stop connectivity watch.
+    if (watcher_ != nullptr) {
+      ClientChannelFilter* client_channel =
+          ClientChannelFilter::GetFromChannel(Channel::FromC(channel_));
+      GPR_ASSERT(client_channel != nullptr);
+      client_channel->RemoveConnectivityWatcher(watcher_);
+      watcher_ = nullptr;
+    }
+    grpc_channel_destroy_internal(channel_);
+  }
+  Unref(DEBUG_LOCATION, "Orphan");
+}
+
+void RlsLb::RlsChannel::StartRlsCall(const RequestKey& key,
+                                     Cache::Entry* stale_entry) {
+  std::unique_ptr<BackOff> backoff_state;
+  grpc_lookup_v1_RouteLookupRequest_Reason reason =
+      grpc_lookup_v1_RouteLookupRequest_REASON_MISS;
+  std::string stale_header_data;
+  if (stale_entry != nullptr) {
+    backoff_state = stale_entry->TakeBackoffState();
+    reason = grpc_lookup_v1_RouteLookupRequest_REASON_STALE;
+    stale_header_data = stale_entry->header_data();
+  }
+  lb_policy_->request_map_.emplace(
+      key, MakeOrphanable<RlsRequest>(
+               lb_policy_.Ref(DEBUG_LOCATION, "RlsRequest"), key,
+               lb_policy_->rls_channel_->Ref(DEBUG_LOCATION, "RlsRequest"),
+               std::move(backoff_state), reason, std::move(stale_header_data)));
+}
+
+void RlsLb::RlsChannel::ReportResponseLocked(bool response_succeeded) {
+  throttle_.RegisterResponse(response_succeeded);
+}
+
+void RlsLb::RlsChannel::ResetBackoff() {
+  GPR_DEBUG_ASSERT(channel_ != nullptr);
+  grpc_channel_reset_connect_backoff(channel_);
+}
+
+//
+// RlsLb::RlsRequest
+//
+
+RlsLb::RlsRequest::RlsRequest(RefCountedPtr<RlsLb> lb_policy, RequestKey key,
+                              RefCountedPtr<RlsChannel> rls_channel,
+                              std::unique_ptr<BackOff> backoff_state,
+                              grpc_lookup_v1_RouteLookupRequest_Reason reason,
+                              std::string stale_header_data)
+    : InternallyRefCounted<RlsRequest>(
+          GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace) ? "RlsRequest" : nullptr),
+      lb_policy_(std::move(lb_policy)),
+      key_(std::move(key)),
+      rls_channel_(std::move(rls_channel)),
+      backoff_state_(std::move(backoff_state)),
+      reason_(reason),
+      stale_header_data_(std::move(stale_header_data)) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO,
+            "[rlslb %p] rls_request=%p: RLS request created for key %s",
+            lb_policy_.get(), this, key_.ToString().c_str());
+  }
+  GRPC_CLOSURE_INIT(&call_complete_cb_, OnRlsCallComplete, this, nullptr);
+  ExecCtx::Run(
+      DEBUG_LOCATION,
+      GRPC_CLOSURE_INIT(&call_start_cb_, StartCall,
+                        Ref(DEBUG_LOCATION, "StartCall").release(), nullptr),
+      absl::OkStatus());
+}
+
+RlsLb::RlsRequest::~RlsRequest() { GPR_ASSERT(call_ == nullptr); }
+
+void RlsLb::RlsRequest::Orphan() {
+  if (call_ != nullptr) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO, "[rlslb %p] rls_request=%p %s: cancelling RLS call",
+              lb_policy_.get(), this, key_.ToString().c_str());
+    }
+    grpc_call_cancel_internal(call_);
+  }
+  Unref(DEBUG_LOCATION, "Orphan");
+}
+
+void RlsLb::RlsRequest::StartCall(void* arg, grpc_error_handle /*error*/) {
+  auto* request = static_cast<RlsRequest*>(arg);
+  request->lb_policy_->work_serializer()->Run(
+      [request]() {
+        request->StartCallLocked();
+        request->Unref(DEBUG_LOCATION, "StartCall");
+      },
+      DEBUG_LOCATION);
+}
+
+void RlsLb::RlsRequest::StartCallLocked() {
+  {
+    MutexLock lock(&lb_policy_->mu_);
+    if (lb_policy_->is_shutdown_) return;
+  }
+  Timestamp now = Timestamp::Now();
+  deadline_ = now + lb_policy_->config_->lookup_service_timeout();
+  grpc_metadata_array_init(&recv_initial_metadata_);
+  grpc_metadata_array_init(&recv_trailing_metadata_);
+  call_ = grpc_channel_create_pollset_set_call(
+      rls_channel_->channel(), nullptr, GRPC_PROPAGATE_DEFAULTS,
+      lb_policy_->interested_parties(),
+      grpc_slice_from_static_string(kRlsRequestPath), nullptr, deadline_,
+      nullptr);
+  grpc_op ops[6];
+  memset(ops, 0, sizeof(ops));
+  grpc_op* op = ops;
+  op->op = GRPC_OP_SEND_INITIAL_METADATA;
+  ++op;
+  op->op = GRPC_OP_SEND_MESSAGE;
+  send_message_ = MakeRequestProto();
+  op->data.send_message.send_message = send_message_;
+  ++op;
+  op->op = GRPC_OP_SEND_CLOSE_FROM_CLIENT;
+  ++op;
+  op->op = GRPC_OP_RECV_INITIAL_METADATA;
+  op->data.recv_initial_metadata.recv_initial_metadata =
+      &recv_initial_metadata_;
+  ++op;
+  op->op = GRPC_OP_RECV_MESSAGE;
+  op->data.recv_message.recv_message = &recv_message_;
+  ++op;
+  op->op = GRPC_OP_RECV_STATUS_ON_CLIENT;
+  op->data.recv_status_on_client.trailing_metadata = &recv_trailing_metadata_;
+  op->data.recv_status_on_client.status = &status_recv_;
+  op->data.recv_status_on_client.status_details = &status_details_recv_;
+  ++op;
+  Ref(DEBUG_LOCATION, "OnRlsCallComplete").release();
+  auto call_error = grpc_call_start_batch_and_execute(
+      call_, ops, static_cast<size_t>(op - ops), &call_complete_cb_);
+  GPR_ASSERT(call_error == GRPC_CALL_OK);
+}
+
+void RlsLb::RlsRequest::OnRlsCallComplete(void* arg, grpc_error_handle error) {
+  auto* request = static_cast<RlsRequest*>(arg);
+  request->lb_policy_->work_serializer()->Run(
+      [request, error]() {
+        request->OnRlsCallCompleteLocked(error);
+        request->Unref(DEBUG_LOCATION, "OnRlsCallComplete");
+      },
+      DEBUG_LOCATION);
+}
+
+void RlsLb::RlsRequest::OnRlsCallCompleteLocked(grpc_error_handle error) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    std::string status_message(StringViewFromSlice(status_details_recv_));
+    gpr_log(GPR_INFO,
+            "[rlslb %p] rls_request=%p %s, error=%s, status={%d, %s} RLS call "
+            "response received",
+            lb_policy_.get(), this, key_.ToString().c_str(),
+            StatusToString(error).c_str(), status_recv_,
+            status_message.c_str());
+  }
+  // Parse response.
+  ResponseInfo response;
+  if (!error.ok()) {
+    grpc_status_code code;
+    std::string message;
+    grpc_error_get_status(error, deadline_, &code, &message,
+                          /*http_error=*/nullptr, /*error_string=*/nullptr);
+    response.status =
+        absl::Status(static_cast<absl::StatusCode>(code), message);
+  } else if (status_recv_ != GRPC_STATUS_OK) {
+    response.status = absl::Status(static_cast<absl::StatusCode>(status_recv_),
+                                   StringViewFromSlice(status_details_recv_));
+  } else {
+    response = ParseResponseProto();
+  }
+  // Clean up call state.
+  grpc_byte_buffer_destroy(send_message_);
+  grpc_byte_buffer_destroy(recv_message_);
+  grpc_metadata_array_destroy(&recv_initial_metadata_);
+  grpc_metadata_array_destroy(&recv_trailing_metadata_);
+  CSliceUnref(status_details_recv_);
+  grpc_call_unref(call_);
+  call_ = nullptr;
+  // Return result to cache.
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] rls_request=%p %s: response info: %s",
+            lb_policy_.get(), this, key_.ToString().c_str(),
+            response.ToString().c_str());
+  }
+  std::vector<ChildPolicyWrapper*> child_policies_to_finish_update;
+  {
+    MutexLock lock(&lb_policy_->mu_);
+    if (lb_policy_->is_shutdown_) return;
+    rls_channel_->ReportResponseLocked(response.status.ok());
+    Cache::Entry* cache_entry = lb_policy_->cache_.FindOrInsert(key_);
+    child_policies_to_finish_update = cache_entry->OnRlsResponseLocked(
+        std::move(response), std::move(backoff_state_));
+    lb_policy_->request_map_.erase(key_);
+  }
+  // Now that we've released the lock, finish the update on any newly
+  // created child policies.
+  for (ChildPolicyWrapper* child : child_policies_to_finish_update) {
+    // TODO(roth): If the child reports an error with the update, we
+    // need to propagate that back to the resolver somehow.
+    (void)child->MaybeFinishUpdate();
+  }
+}
+
+grpc_byte_buffer* RlsLb::RlsRequest::MakeRequestProto() {
+  upb::Arena arena;
+  grpc_lookup_v1_RouteLookupRequest* req =
+      grpc_lookup_v1_RouteLookupRequest_new(arena.ptr());
+  grpc_lookup_v1_RouteLookupRequest_set_target_type(
+      req, upb_StringView_FromDataAndSize(kGrpc, sizeof(kGrpc) - 1));
+  for (const auto& kv : key_.key_map) {
+    grpc_lookup_v1_RouteLookupRequest_key_map_set(
+        req, upb_StringView_FromDataAndSize(kv.first.data(), kv.first.size()),
+        upb_StringView_FromDataAndSize(kv.second.data(), kv.second.size()),
+        arena.ptr());
+  }
+  grpc_lookup_v1_RouteLookupRequest_set_reason(req, reason_);
+  if (!stale_header_data_.empty()) {
+    grpc_lookup_v1_RouteLookupRequest_set_stale_header_data(
+        req, upb_StringView_FromDataAndSize(stale_header_data_.data(),
+                                            stale_header_data_.size()));
+  }
+  size_t len;
+  char* buf =
+      grpc_lookup_v1_RouteLookupRequest_serialize(req, arena.ptr(), &len);
+  grpc_slice send_slice = grpc_slice_from_copied_buffer(buf, len);
+  grpc_byte_buffer* byte_buffer = grpc_raw_byte_buffer_create(&send_slice, 1);
+  CSliceUnref(send_slice);
+  return byte_buffer;
+}
+
+RlsLb::ResponseInfo RlsLb::RlsRequest::ParseResponseProto() {
+  ResponseInfo response_info;
+  upb::Arena arena;
+  grpc_byte_buffer_reader bbr;
+  grpc_byte_buffer_reader_init(&bbr, recv_message_);
+  grpc_slice recv_slice = grpc_byte_buffer_reader_readall(&bbr);
+  grpc_byte_buffer_reader_destroy(&bbr);
+  grpc_lookup_v1_RouteLookupResponse* response =
+      grpc_lookup_v1_RouteLookupResponse_parse(
+          reinterpret_cast<const char*>(GRPC_SLICE_START_PTR(recv_slice)),
+          GRPC_SLICE_LENGTH(recv_slice), arena.ptr());
+  CSliceUnref(recv_slice);
+  if (response == nullptr) {
+    response_info.status = absl::InternalError("cannot parse RLS response");
+    return response_info;
+  }
+  size_t num_targets;
+  const upb_StringView* targets_strview =
+      grpc_lookup_v1_RouteLookupResponse_targets(response, &num_targets);
+  if (num_targets == 0) {
+    response_info.status =
+        absl::InvalidArgumentError("RLS response has no target entry");
+    return response_info;
+  }
+  response_info.targets.reserve(num_targets);
+  for (size_t i = 0; i < num_targets; ++i) {
+    response_info.targets.emplace_back(targets_strview[i].data,
+                                       targets_strview[i].size);
+  }
+  upb_StringView header_data_strview =
+      grpc_lookup_v1_RouteLookupResponse_header_data(response);
+  response_info.header_data =
+      std::string(header_data_strview.data, header_data_strview.size);
+  return response_info;
+}
+
+//
+// RlsLb
+//
+
+RlsLb::RlsLb(Args args) : LoadBalancingPolicy(std::move(args)), cache_(this) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] policy created", this);
+  }
+}
+
+bool EndpointsEqual(
+    const absl::StatusOr<std::shared_ptr<EndpointAddressesIterator>> endpoints1,
+    const absl::StatusOr<std::shared_ptr<EndpointAddressesIterator>>
+        endpoints2) {
+  if (endpoints1.status() != endpoints2.status()) return false;
+  if (endpoints1.ok()) {
+    std::vector<EndpointAddresses> e1_list;
+    (*endpoints1)->ForEach([&](const EndpointAddresses& endpoint) {
+      e1_list.push_back(endpoint);
+    });
+    size_t i = 0;
+    bool different = false;
+    (*endpoints2)->ForEach([&](const EndpointAddresses& endpoint) {
+      if (endpoint != e1_list[i++]) different = true;
+    });
+    if (different) return false;
+    if (i != e1_list.size()) return false;
+  }
+  return true;
+}
+
+absl::Status RlsLb::UpdateLocked(UpdateArgs args) {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] policy updated", this);
+  }
+  update_in_progress_ = true;
+  // Swap out config.
+  RefCountedPtr<RlsLbConfig> old_config = std::move(config_);
+  config_ = args.config.TakeAsSubclass<RlsLbConfig>();
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace) &&
+      (old_config == nullptr ||
+       old_config->child_policy_config() != config_->child_policy_config())) {
+    gpr_log(GPR_INFO, "[rlslb %p] updated child policy config: %s", this,
+            JsonDump(config_->child_policy_config()).c_str());
+  }
+  // Swap out addresses.
+  // If the new address list is an error and we have an existing address list,
+  // stick with the existing addresses.
+  absl::StatusOr<std::shared_ptr<EndpointAddressesIterator>> old_addresses;
+  if (args.addresses.ok()) {
+    old_addresses = std::move(addresses_);
+    addresses_ = std::move(args.addresses);
+  } else {
+    old_addresses = addresses_;
+  }
+  // Swap out channel args.
+  channel_args_ = std::move(args.args);
+  // Determine whether we need to update all child policies.
+  bool update_child_policies =
+      old_config == nullptr ||
+      old_config->child_policy_config() != config_->child_policy_config() ||
+      !EndpointsEqual(old_addresses, addresses_) || args.args != channel_args_;
+  // If default target changes, swap out child policy.
+  bool created_default_child = false;
+  if (old_config == nullptr ||
+      config_->default_target() != old_config->default_target()) {
+    if (config_->default_target().empty()) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+        gpr_log(GPR_INFO, "[rlslb %p] unsetting default target", this);
+      }
+      default_child_policy_.reset();
+    } else {
+      auto it = child_policy_map_.find(config_->default_target());
+      if (it == child_policy_map_.end()) {
+        if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+          gpr_log(GPR_INFO, "[rlslb %p] creating new default target", this);
+        }
+        default_child_policy_ = MakeRefCounted<ChildPolicyWrapper>(
+            RefAsSubclass<RlsLb>(DEBUG_LOCATION, "ChildPolicyWrapper"),
+            config_->default_target());
+        created_default_child = true;
+      } else {
+        if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+          gpr_log(GPR_INFO,
+                  "[rlslb %p] using existing child for default target", this);
+        }
+        default_child_policy_ =
+            it->second->Ref(DEBUG_LOCATION, "DefaultChildPolicy");
+      }
+    }
+  }
+  // Now grab the lock to swap out the state it guards.
+  {
+    MutexLock lock(&mu_);
+    // Swap out RLS channel if needed.
+    if (old_config == nullptr ||
+        config_->lookup_service() != old_config->lookup_service()) {
+      rls_channel_ = MakeOrphanable<RlsChannel>(
+          RefAsSubclass<RlsLb>(DEBUG_LOCATION, "RlsChannel"));
+    }
+    // Resize cache if needed.
+    if (old_config == nullptr ||
+        config_->cache_size_bytes() != old_config->cache_size_bytes()) {
+      cache_.Resize(static_cast<size_t>(config_->cache_size_bytes()));
+    }
+    // Start update of child policies if needed.
+    if (update_child_policies) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+        gpr_log(GPR_INFO, "[rlslb %p] starting child policy updates", this);
+      }
+      for (auto& p : child_policy_map_) {
+        p.second->StartUpdate();
+      }
+    } else if (created_default_child) {
+      if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+        gpr_log(GPR_INFO, "[rlslb %p] starting default child policy update",
+                this);
+      }
+      default_child_policy_->StartUpdate();
+    }
+  }
+  // Now that we've released the lock, finish update of child policies.
+  std::vector<std::string> errors;
+  if (update_child_policies) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO, "[rlslb %p] finishing child policy updates", this);
+    }
+    for (auto& p : child_policy_map_) {
+      absl::Status status = p.second->MaybeFinishUpdate();
+      if (!status.ok()) {
+        errors.emplace_back(
+            absl::StrCat("target ", p.first, ": ", status.ToString()));
+      }
+    }
+  } else if (created_default_child) {
+    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+      gpr_log(GPR_INFO, "[rlslb %p] finishing default child policy update",
+              this);
+    }
+    absl::Status status = default_child_policy_->MaybeFinishUpdate();
+    if (!status.ok()) {
+      errors.emplace_back(absl::StrCat("target ", config_->default_target(),
+                                       ": ", status.ToString()));
+    }
+  }
+  update_in_progress_ = false;
+  // In principle, we need to update the picker here only if the config
+  // fields used by the picker have changed.  However, it seems fragile
+  // to check individual fields, since the picker logic could change in
+  // the future to use additional config fields, and we might not
+  // remember to update the code here.  So for now, we just unconditionally
+  // update the picker here, even though it's probably redundant.
+  UpdatePickerLocked();
+  // Return status.
+  if (!errors.empty()) {
+    return absl::UnavailableError(absl::StrCat(
+        "errors from children: [", absl::StrJoin(errors, "; "), "]"));
+  }
+  return absl::OkStatus();
+}
+
+void RlsLb::ExitIdleLocked() {
+  MutexLock lock(&mu_);
+  for (auto& child_entry : child_policy_map_) {
+    child_entry.second->ExitIdleLocked();
+  }
+}
+
+void RlsLb::ResetBackoffLocked() {
+  {
+    MutexLock lock(&mu_);
+    rls_channel_->ResetBackoff();
+    cache_.ResetAllBackoff();
+  }
+  for (auto& child : child_policy_map_) {
+    child.second->ResetBackoffLocked();
+  }
+}
+
+void RlsLb::ShutdownLocked() {
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] policy shutdown", this);
+  }
+  MutexLock lock(&mu_);
+  is_shutdown_ = true;
+  config_.reset(DEBUG_LOCATION, "ShutdownLocked");
+  channel_args_ = ChannelArgs();
+  cache_.Shutdown();
+  request_map_.clear();
+  rls_channel_.reset();
+  default_child_policy_.reset();
+}
+
+void RlsLb::UpdatePickerAsync() {
+  // Run via the ExecCtx, since the caller may be holding the lock, and
+  // we don't want to be doing that when we hop into the WorkSerializer,
+  // in case the WorkSerializer callback happens to run inline.
+  ExecCtx::Run(
+      DEBUG_LOCATION,
+      GRPC_CLOSURE_CREATE(UpdatePickerCallback,
+                          Ref(DEBUG_LOCATION, "UpdatePickerCallback").release(),
+                          grpc_schedule_on_exec_ctx),
+      absl::OkStatus());
+}
+
+void RlsLb::UpdatePickerCallback(void* arg, grpc_error_handle /*error*/) {
+  auto* rls_lb = static_cast<RlsLb*>(arg);
+  rls_lb->work_serializer()->Run(
+      [rls_lb]() {
+        RefCountedPtr<RlsLb> lb_policy(rls_lb);
+        lb_policy->UpdatePickerLocked();
+        lb_policy.reset(DEBUG_LOCATION, "UpdatePickerCallback");
+      },
+      DEBUG_LOCATION);
+}
+
+void RlsLb::UpdatePickerLocked() {
+  // If we're in the process of propagating an update from our parent to
+  // our children, ignore any updates that come from the children.  We
+  // will instead return a new picker once the update has been seen by
+  // all children.  This avoids unnecessary picker churn while an update
+  // is being propagated to our children.
+  if (update_in_progress_) return;
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] updating picker", this);
+  }
+  grpc_connectivity_state state = GRPC_CHANNEL_IDLE;
+  if (!child_policy_map_.empty()) {
+    state = GRPC_CHANNEL_TRANSIENT_FAILURE;
+    int num_idle = 0;
+    int num_connecting = 0;
+    {
+      MutexLock lock(&mu_);
+      if (is_shutdown_) return;
+      for (auto& p : child_policy_map_) {
+        grpc_connectivity_state child_state = p.second->connectivity_state();
+        if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+          gpr_log(GPR_INFO, "[rlslb %p] target %s in state %s", this,
+                  p.second->target().c_str(),
+                  ConnectivityStateName(child_state));
+        }
+        if (child_state == GRPC_CHANNEL_READY) {
+          state = GRPC_CHANNEL_READY;
+          break;
+        } else if (child_state == GRPC_CHANNEL_CONNECTING) {
+          ++num_connecting;
+        } else if (child_state == GRPC_CHANNEL_IDLE) {
+          ++num_idle;
+        }
+      }
+      if (state != GRPC_CHANNEL_READY) {
+        if (num_connecting > 0) {
+          state = GRPC_CHANNEL_CONNECTING;
+        } else if (num_idle > 0) {
+          state = GRPC_CHANNEL_IDLE;
+        }
+      }
+    }
+  }
+  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_rls_trace)) {
+    gpr_log(GPR_INFO, "[rlslb %p] reporting state %s", this,
+            ConnectivityStateName(state));
+  }
+  absl::Status status;
+  if (state == GRPC_CHANNEL_TRANSIENT_FAILURE) {
+    status = absl::UnavailableError("no children available");
+  }
+  channel_control_helper()->UpdateState(
+      state, status,
+      MakeRefCounted<Picker>(RefAsSubclass<RlsLb>(DEBUG_LOCATION, "Picker")));
+}
+
+//
+// RlsLbFactory
+//
+
+struct GrpcKeyBuilder {
+  struct Name {
+    std::string service;
+    std::string method;
+
+    static const JsonLoaderInterface* JsonLoader(const JsonArgs&) {
+      static const auto* loader = JsonObjectLoader<Name>()
+                                      .Field("service", &Name::service)
+                                      .OptionalField("method", &Name::method)
+                                      .Finish();
+      return loader;
+    }
+  };
+
+  struct NameMatcher {
+    std::string key;
+    std::vector<std::string> names;
+    absl::optional<bool> required_match;
+
+    static const JsonLoaderInterface* JsonLoader(const JsonArgs&) {
+      static const auto* loader =
+          JsonObjectLoader<NameMatcher>()
+              .Field("key", &NameMatcher::key)
+              .Field("names", &NameMatcher::names)
+              .OptionalField("requiredMatch", &NameMatcher::required_match)
+              .Finish();
+      return loader;
+    }
+
+    void JsonPostLoad(const Json&, const JsonArgs&, ValidationErrors* errors) {
+      // key must be non-empty.
+      {
+        ValidationErrors::ScopedField field(errors, ".key");
+        if (!errors->FieldHasErrors() && key.empty()) {
+          errors->AddError("must be non-empty");
+        }
+      }
+      // List of header names must be non-empty.
+      {
+        ValidationErrors::ScopedField field(errors, ".names");
+        if (!errors->FieldHasErrors() && names.empty()) {
+          errors->AddError("must be non-empty");
+        }
+        // Individual header names must be non-empty.
+        for (size_t i = 0; i < names.size(); ++i) {
+          ValidationErrors::ScopedField field(errors,
+                                              absl::StrCat("[", i, "]"));
+          if (!errors->FieldHasErrors() && names[i].empty()) {
+            errors->AddError("must be non-empty");
+          }
+        }
+      }
+      // requiredMatch must not be present.
+      {
+        ValidationErrors::ScopedField field(errors, ".requiredMatch");
+        if (required_match.has_value()) {
+          errors->AddError("must not be present");
+        }
+      }
+    }
+  };
+
+  struct ExtraKeys {
+    absl::optional<std::string> host_key;
+    absl::optional<std::string> service_key;
+    absl::optional<std::string> method_key;
+
+    static const JsonLoaderInterface* JsonLoader(const JsonArgs&) {
+      static const auto* loader =
+          JsonObjectLoader<ExtraKeys>()
+              .OptionalField("host", &ExtraKeys::host_key)
+              .OptionalField("service", &ExtraKeys::service_key)
+              .OptionalField("method", &ExtraKeys::method_key)
+              .Finish();
+      return loader;
+    }
+
+    void JsonPostLoad(const Json&, const JsonArgs&, ValidationErrors* errors) {
+      auto check_field = [&](const std::string& field_name,
+                             absl::optional<std::string>* struct_field) {
+        ValidationErrors::ScopedField field(errors,
+                                            absl::StrCat(".", field_name));
+        if (struct_field->has_value() && (*struct_field)->empty()) {
+          errors->AddError("must be non-empty if set");
+        }
+      };
+      check_field("host", &host_key);
+      check_field("service", &service_key);
+      check_field("method", &method_key);
+    }
+  };
+
+  std::vector<Name> names;
+  std::vector<NameMatcher> headers;
+  ExtraKeys extra_keys;
+  std::map<std::string /*key*/, std::string /*value*/> constant_keys;
+
+  static const JsonLoaderInterface* JsonLoader(const JsonArgs&) {
+    static const auto* loader =
+        JsonObjectLoader<GrpcKeyBuilder>()
+            .Field("names", &GrpcKeyBuilder::names)
+            .OptionalField("headers", &GrpcKeyBuilder::headers)
+            .OptionalField("extraKeys", &GrpcKeyBuilder::extra_keys)
+            .OptionalField("constantKeys", &GrpcKeyBuilder::constant_keys)
+            .Finish();
+    return loader;
+  }
+
+  void JsonPostLoad(const Json&, const JsonArgs&, ValidationErrors* errors) {
+    // The names field must be non-empty.
+    {
+      ValidationErrors::ScopedField field(errors, ".names");
+      if (!errors->FieldHasErrors() && names.empty()) {
+        errors->AddError("must be non-empty");
+      }
+    }
+    // Make sure no key in constantKeys is empty.
+    if (constant_keys.find("") != constant_keys.end()) {
+      ValidationErrors::ScopedField field(errors, ".constantKeys[\"\"]");
+      errors->AddError("key must be non-empty");
+    }
+    // Check for duplicate keys.
+    std::set<absl::string_view> keys_seen;
+    auto duplicate_key_check_func = [&keys_seen, errors](
+                                        const std::string& key,
+                                        const std::string& field_name) {
+      if (key.empty()) return;  // Already generated an error about this.
+      ValidationErrors::ScopedField field(errors, field_name);
+      auto it = keys_seen.find(key);
+      if (it != keys_seen.end()) {
+        errors->AddError(absl::StrCat("duplicate key \"", key, "\""));
+      } else {
+        keys_seen.insert(key);
+      }
+    };
+    for (size_t i = 0; i < headers.size(); ++i) {
+      NameMatcher& header = headers[i];
+      duplicate_key_check_func(header.key,
+                               absl::StrCat(".headers[", i, "].key"));
+    }
+    for (const auto& p : constant_keys) {
+      duplicate_key_check_func(
+          p.first, absl::StrCat(".constantKeys[\"", p.first, "\"]"));
+    }
+    if (extra_keys.host_key.has_value()) {
+      duplicate_key_check_func(*extra_keys.host_key, ".extraKeys.host");
+    }
+    if (extra_keys.service_key.has_value()) {
+      duplicate_key_check_func(*extra_keys.service_key, ".extraKeys.service");
+    }
+    if (extra_keys.method_key.has_value()) {
+      duplicate_key_check_func(*extra_keys.method_key, ".extraKeys.method");
+    }
+  }
+};
+
+const JsonLoaderInterface* RlsLbConfig::RouteLookupConfig::JsonLoader(
+    const JsonArgs&) {
+  static const auto* loader =
+      JsonObjectLoader<RouteLookupConfig>()
+          // Note: Some fields require manual processing and are handled in
+          // JsonPostLoad() instead.
+          .Field("lookupService", &RouteLookupConfig::lookup_service)
+          .OptionalField("lookupServiceTimeout",
+                         &RouteLookupConfig::lookup_service_timeout)
+          .OptionalField("maxAge", &RouteLookupConfig::max_age)
+          .OptionalField("staleAge", &RouteLookupConfig::stale_age)
+          .Field("cacheSizeBytes", &RouteLookupConfig::cache_size_bytes)
+          .OptionalField("defaultTarget", &RouteLookupConfig::default_target)
+          .Finish();
+  return loader;
+}
+
+void RlsLbConfig::RouteLookupConfig::JsonPostLoad(const Json& json,
+                                                  const JsonArgs& args,
+                                                  ValidationErrors* errors) {
+  // Parse grpcKeybuilders.
+  auto grpc_keybuilders = LoadJsonObjectField<std::vector<GrpcKeyBuilder>>(
+      json.object(), args, "grpcKeybuilders", errors);
+  if (grpc_keybuilders.has_value()) {
+    ValidationErrors::ScopedField field(errors, ".grpcKeybuilders");
+    for (size_t i = 0; i < grpc_keybuilders->size(); ++i) {
+      ValidationErrors::ScopedField field(errors, absl::StrCat("[", i, "]"));
+      auto& grpc_keybuilder = (*grpc_keybuilders)[i];
+      // Construct KeyBuilder.
+      RlsLbConfig::KeyBuilder key_builder;
+      for (const auto& header : grpc_keybuilder.headers) {
+        key_builder.header_keys.emplace(header.key, header.names);
+      }
+      if (grpc_keybuilder.extra_keys.host_key.has_value()) {
+        key_builder.host_key = std::move(*grpc_keybuilder.extra_keys.host_key);
+      }
+      if (grpc_keybuilder.extra_keys.service_key.has_value()) {
+        key_builder.service_key =
+            std::move(*grpc_keybuilder.extra_keys.service_key);
+      }
+      if (grpc_keybuilder.extra_keys.method_key.has_value()) {
+        key_builder.method_key =
+            std::move(*grpc_keybuilder.extra_keys.method_key);
+      }
+      key_builder.constant_keys = std::move(grpc_keybuilder.constant_keys);
+      // Add entries to map.
+      for (const auto& name : grpc_keybuilder.names) {
+        std::string path = absl::StrCat("/", name.service, "/", name.method);
+        bool inserted = key_builder_map.emplace(path, key_builder).second;
+        if (!inserted) {
+          errors->AddError(absl::StrCat("duplicate entry for \"", path, "\""));
+        }
+      }
+    }
+  }
+  // Validate lookupService.
+  {
+    ValidationErrors::ScopedField field(errors, ".lookupService");
+    if (!errors->FieldHasErrors() &&
+        !CoreConfiguration::Get().resolver_registry().IsValidTarget(
+            lookup_service)) {
+      errors->AddError("must be valid gRPC target URI");
+    }
+  }
+  // Clamp maxAge to the max allowed value.
+  if (max_age > kMaxMaxAge) max_age = kMaxMaxAge;
+  // If staleAge is set, then maxAge must also be set.
+  if (json.object().find("staleAge") != json.object().end() &&
+      json.object().find("maxAge") == json.object().end()) {
+    ValidationErrors::ScopedField field(errors, ".maxAge");
+    errors->AddError("must be set if staleAge is set");
+  }
+  // Ignore staleAge if greater than or equal to maxAge.
+  if (stale_age >= max_age) stale_age = max_age;
+  // Validate cacheSizeBytes.
+  {
+    ValidationErrors::ScopedField field(errors, ".cacheSizeBytes");
+    if (!errors->FieldHasErrors() && cache_size_bytes <= 0) {
+      errors->AddError("must be greater than 0");
+    }
+  }
+  // Clamp cacheSizeBytes to the max allowed value.
+  if (cache_size_bytes > kMaxCacheSizeBytes) {
+    cache_size_bytes = kMaxCacheSizeBytes;
+  }
+  // Validate defaultTarget.
+  {
+    ValidationErrors::ScopedField field(errors, ".defaultTarget");
+    if (!errors->FieldHasErrors() &&
+        json.object().find("defaultTarget") != json.object().end() &&
+        default_target.empty()) {
+      errors->AddError("must be non-empty if set");
+    }
+  }
+}
+
+const JsonLoaderInterface* RlsLbConfig::JsonLoader(const JsonArgs&) {
+  static const auto* loader =
+      JsonObjectLoader<RlsLbConfig>()
+          // Note: Some fields require manual processing and are handled in
+          // JsonPostLoad() instead.
+          .Field("routeLookupConfig", &RlsLbConfig::route_lookup_config_)
+          .Field("childPolicyConfigTargetFieldName",
+                 &RlsLbConfig::child_policy_config_target_field_name_)
+          .Finish();
+  return loader;
+}
+
+void RlsLbConfig::JsonPostLoad(const Json& json, const JsonArgs&,
+                               ValidationErrors* errors) {
+  // Parse routeLookupChannelServiceConfig.
+  auto it = json.object().find("routeLookupChannelServiceConfig");
+  if (it != json.object().end()) {
+    ValidationErrors::ScopedField field(errors,
+                                        ".routeLookupChannelServiceConfig");
+    // Don't need to save the result here, just need the errors (if any).
+    ServiceConfigImpl::Create(ChannelArgs(), it->second, errors);
+  }
+  // Validate childPolicyConfigTargetFieldName.
+  {
+    ValidationErrors::ScopedField field(errors,
+                                        ".childPolicyConfigTargetFieldName");
+    if (!errors->FieldHasErrors() &&
+        child_policy_config_target_field_name_.empty()) {
+      errors->AddError("must be non-empty");
+    }
+  }
+  // Parse childPolicy.
+  {
+    ValidationErrors::ScopedField field(errors, ".childPolicy");
+    auto it = json.object().find("childPolicy");
+    if (it == json.object().end()) {
+      errors->AddError("field not present");
+    } else {
+      // Add target to all child policy configs in the list.
+      std::string target = route_lookup_config_.default_target.empty()
+                               ? kFakeTargetFieldValue
+                               : route_lookup_config_.default_target;
+      auto child_policy_config = InsertOrUpdateChildPolicyField(
+          child_policy_config_target_field_name_, target, it->second, errors);
+      if (child_policy_config.has_value()) {
+        child_policy_config_ = std::move(*child_policy_config);
+        // Parse the config.
+        auto parsed_config =
+            CoreConfiguration::Get()
+                .lb_policy_registry()
+                .ParseLoadBalancingConfig(child_policy_config_);
+        if (!parsed_config.ok()) {
+          errors->AddError(parsed_config.status().message());
+        } else {
+          // Find the chosen config and return it in JSON form.
+          // We remove all non-selected configs, and in the selected config,
+          // we leave the target field in place, set to the default value.
+          // This slightly optimizes what we need to do later when we update
+          // a child policy for a given target.
+          for (const Json& config : child_policy_config_.array()) {
+            if (config.object().begin()->first == (*parsed_config)->name()) {
+              child_policy_config_ = Json::FromArray({config});
+              break;
+            }
+          }
+          // If default target is set, set the default child config.
+          if (!route_lookup_config_.default_target.empty()) {
+            default_child_policy_parsed_config_ = std::move(*parsed_config);
+          }
+        }
+      }
+    }
+  }
+}
+
+class RlsLbFactory : public LoadBalancingPolicyFactory {
+ public:
+  absl::string_view name() const override { return kRls; }
+
+  OrphanablePtr<LoadBalancingPolicy> CreateLoadBalancingPolicy(
+      LoadBalancingPolicy::Args args) const override {
+    return MakeOrphanable<RlsLb>(std::move(args));
+  }
+
+  absl::StatusOr<RefCountedPtr<LoadBalancingPolicy::Config>>
+  ParseLoadBalancingConfig(const Json& json) const override {
+    return LoadFromJson<RefCountedPtr<RlsLbConfig>>(
+        json, JsonArgs(), "errors validing RLS LB policy config");
+  }
+};
+
+}  //  namespace
+
+void RegisterRlsLbPolicy(CoreConfiguration::Builder* builder) {
+  builder->lb_policy_registry()->RegisterLoadBalancingPolicyFactory(
+      std::make_unique<RlsLbFactory>());
+}
+
+}  // namespace grpc_core