grpc 1.63.0 → 1.72.0

data/src/core/load_balancing/pick_first/pick_first.cc

@@ -14,14 +14,17 @@
 // limitations under the License.
 //
 
-#include <grpc/support/port_platform.h>
-
 #include "src/core/load_balancing/pick_first/pick_first.h"
 
+#include <grpc/event_engine/event_engine.h>
+#include <grpc/impl/channel_arg_names.h>
+#include <grpc/impl/connectivity_state.h>
+#include <grpc/support/port_platform.h>
 #include <inttypes.h>
 #include <string.h>
 
 #include <memory>
+#include <optional>
 #include <set>
 #include <string>
 #include <type_traits>
@@ -29,47 +32,41 @@
 #include <vector>
 
 #include "absl/algorithm/container.h"
+#include "absl/log/check.h"
+#include "absl/log/log.h"
 #include "absl/random/random.h"
 #include "absl/status/status.h"
 #include "absl/status/statusor.h"
 #include "absl/strings/str_cat.h"
 #include "absl/strings/string_view.h"
-#include "absl/types/optional.h"
-
-#include <grpc/event_engine/event_engine.h>
-#include <grpc/impl/channel_arg_names.h>
-#include <grpc/impl/connectivity_state.h>
-#include <grpc/support/log.h>
-
-#include "src/core/load_balancing/health_check_client.h"
+#include "src/core/config/core_configuration.h"
 #include "src/core/lib/address_utils/sockaddr_utils.h"
 #include "src/core/lib/channel/channel_args.h"
-#include "src/core/lib/channel/metrics.h"
-#include "src/core/lib/config/core_configuration.h"
 #include "src/core/lib/debug/trace.h"
-#include "src/core/lib/gpr/useful.h"
-#include "src/core/lib/gprpp/crash.h"
-#include "src/core/lib/gprpp/debug_location.h"
-#include "src/core/lib/gprpp/orphanable.h"
-#include "src/core/lib/gprpp/ref_counted_ptr.h"
-#include "src/core/lib/gprpp/time.h"
-#include "src/core/lib/gprpp/work_serializer.h"
+#include "src/core/lib/experiments/experiments.h"
 #include "src/core/lib/iomgr/exec_ctx.h"
 #include "src/core/lib/iomgr/iomgr_fwd.h"
 #include "src/core/lib/iomgr/resolved_address.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/transport/connectivity_state.h"
+#include "src/core/load_balancing/health_check_client.h"
 #include "src/core/load_balancing/lb_policy.h"
 #include "src/core/load_balancing/lb_policy_factory.h"
 #include "src/core/load_balancing/subchannel_interface.h"
 #include "src/core/resolver/endpoint_addresses.h"
+#include "src/core/telemetry/metrics.h"
+#include "src/core/util/crash.h"
+#include "src/core/util/debug_location.h"
+#include "src/core/util/json/json.h"
+#include "src/core/util/json/json_args.h"
+#include "src/core/util/json/json_object_loader.h"
+#include "src/core/util/orphanable.h"
+#include "src/core/util/ref_counted_ptr.h"
+#include "src/core/util/time.h"
+#include "src/core/util/useful.h"
+#include "src/core/util/work_serializer.h"
 
 namespace grpc_core {
 
-TraceFlag grpc_lb_pick_first_trace(false, "pick_first");
-
 namespace {
 
 //
@@ -83,19 +80,25 @@ const auto kMetricDisconnections =
         "grpc.lb.pick_first.disconnections",
         "EXPERIMENTAL.  Number of times the selected subchannel becomes "
         "disconnected.",
-        "{disconnection}", {kMetricLabelTarget}, {}, false);
+        "{disconnection}", false)
+        .Labels(kMetricLabelTarget)
+        .Build();
 
 const auto kMetricConnectionAttemptsSucceeded =
     GlobalInstrumentsRegistry::RegisterUInt64Counter(
         "grpc.lb.pick_first.connection_attempts_succeeded",
-        "EXPERIMENTAL.  Number of successful connection attempts.",
-        "{attempt}", {kMetricLabelTarget}, {}, false);
+        "EXPERIMENTAL.  Number of successful connection attempts.", "{attempt}",
+        false)
+        .Labels(kMetricLabelTarget)
+        .Build();
 
 const auto kMetricConnectionAttemptsFailed =
     GlobalInstrumentsRegistry::RegisterUInt64Counter(
         "grpc.lb.pick_first.connection_attempts_failed",
-        "EXPERIMENTAL.  Number of failed connection attempts.",
-        "{attempt}", {kMetricLabelTarget}, {}, false);
+        "EXPERIMENTAL.  Number of failed connection attempts.", "{attempt}",
+        false)
+        .Labels(kMetricLabelTarget)
+        .Build();
 
 class PickFirstConfig final : public LoadBalancingPolicy::Config {
  public:
@@ -128,114 +131,147 @@ class PickFirst final : public LoadBalancingPolicy {
  private:
   ~PickFirst() override;
 
+  // A list of subchannels that we will attempt connections on.
   class SubchannelList final : public InternallyRefCounted<SubchannelList> {
    public:
+    // Data about the subchannel that is needed only while attempting to
+    // connect.
     class SubchannelData final {
      public:
+      // Stores the subchannel and its watcher.  This is the state that
+      // is retained once a subchannel is chosen.
+      class SubchannelState final
+          : public InternallyRefCounted<SubchannelState> {
+       public:
+        SubchannelState(SubchannelData* subchannel_data,
+                        RefCountedPtr<SubchannelInterface> subchannel);
+
+        void Orphan() override;
+
+        SubchannelInterface* subchannel() const { return subchannel_.get(); }
+
+        void RequestConnection() { subchannel_->RequestConnection(); }
+
+        void ResetBackoffLocked() { subchannel_->ResetBackoff(); }
+
+       private:
+        // Watcher for subchannel connectivity state.
+        class Watcher
+            : public SubchannelInterface::ConnectivityStateWatcherInterface {
+         public:
+          explicit Watcher(RefCountedPtr<SubchannelState> subchannel_state)
+              : subchannel_state_(std::move(subchannel_state)) {}
+
+          ~Watcher() override {
+            subchannel_state_.reset(DEBUG_LOCATION, "Watcher dtor");
+          }
+
+          void OnConnectivityStateChange(grpc_connectivity_state new_state,
+                                         absl::Status status) override {
+            subchannel_state_->OnConnectivityStateChange(new_state,
+                                                         std::move(status));
+          }
+
+          grpc_pollset_set* interested_parties() override {
+            return subchannel_state_->pick_first_->interested_parties();
+          }
+
+         private:
+          RefCountedPtr<SubchannelState> subchannel_state_;
+        };
+
+        // Selects this subchannel.  Called when the subchannel reports READY.
+        void Select();
+
+        // This method will be invoked once soon after instantiation to report
+        // the current connectivity state, and it will then be invoked again
+        // whenever the connectivity state changes.
+        void OnConnectivityStateChange(grpc_connectivity_state new_state,
+                                       absl::Status status);
+
+        // If non-null, then we are still part of a subchannel list
+        // trying to connect.
+        SubchannelData* subchannel_data_;
+
+        // TODO(roth): Once we remove pollset_set, we should no longer
+        // need to hold a ref to PickFirst.  Instead, we can make this a
+        // raw pointer and put it in an std::variant with subchannel_data_.
+        RefCountedPtr<PickFirst> pick_first_;
+
+        RefCountedPtr<SubchannelInterface> subchannel_;
+        SubchannelInterface::ConnectivityStateWatcherInterface* watcher_ =
+            nullptr;
+      };
+
       SubchannelData(SubchannelList* subchannel_list, size_t index,
                      RefCountedPtr<SubchannelInterface> subchannel);
 
-      SubchannelInterface* subchannel() const { return subchannel_.get(); }
-      absl::optional<grpc_connectivity_state> connectivity_state() const {
+      std::optional<grpc_connectivity_state> connectivity_state() const {
         return connectivity_state_;
       }
       const absl::Status& connectivity_status() const {
         return connectivity_status_;
       }
 
-      // Resets the connection backoff.
-      void ResetBackoffLocked() {
-        if (subchannel_ != nullptr) subchannel_->ResetBackoff();
-      }
+      void RequestConnection() { subchannel_state_->RequestConnection(); }
 
-      void RequestConnection() { subchannel_->RequestConnection(); }
+      // Resets the connection backoff.
+      void ResetBackoffLocked() { subchannel_state_->ResetBackoffLocked(); }
 
       // Requests a connection attempt to start on this subchannel,
       // with appropriate Connection Attempt Delay.
       // Used only during the Happy Eyeballs pass.
       void RequestConnectionWithTimer();
 
-      // Cancels any pending connectivity watch and unrefs the subchannel.
-      void ShutdownLocked();
-
       bool seen_transient_failure() const { return seen_transient_failure_; }
+      void set_seen_transient_failure() { seen_transient_failure_ = true; }
 
      private:
-      // Watcher for subchannel connectivity state.
-      class Watcher final
-          : public SubchannelInterface::ConnectivityStateWatcherInterface {
-       public:
-        Watcher(RefCountedPtr<SubchannelList> subchannel_list, size_t index)
-            : subchannel_list_(std::move(subchannel_list)), index_(index) {}
-
-        ~Watcher() override {
-          subchannel_list_.reset(DEBUG_LOCATION, "Watcher dtor");
-        }
-
-        void OnConnectivityStateChange(grpc_connectivity_state new_state,
-                                       absl::Status status) override {
-          subchannel_list_->subchannels_[index_].OnConnectivityStateChange(
-              new_state, std::move(status));
-        }
-
-        grpc_pollset_set* interested_parties() override {
-          return subchannel_list_->policy_->interested_parties();
-        }
-
-       private:
-        RefCountedPtr<SubchannelList> subchannel_list_;
-        const size_t index_;
-      };
-
       // This method will be invoked once soon after instantiation to report
       // the current connectivity state, and it will then be invoked again
       // whenever the connectivity state changes.
       void OnConnectivityStateChange(grpc_connectivity_state new_state,
                                      absl::Status status);
 
-      // Processes the connectivity change to READY for an unselected
-      // subchannel.
-      void ProcessUnselectedReadyLocked();
-
       // Backpointer to owning subchannel list.  Not owned.
       SubchannelList* subchannel_list_;
+      // Our index within subchannel_list_.
       const size_t index_;
-      // The subchannel.
-      RefCountedPtr<SubchannelInterface> subchannel_;
-      // Will be non-null when the subchannel's state is being watched.
-      SubchannelInterface::ConnectivityStateWatcherInterface* pending_watcher_ =
-          nullptr;
+      // Subchannel state.
+      OrphanablePtr<SubchannelState> subchannel_state_;
       // Data updated by the watcher.
-      absl::optional<grpc_connectivity_state> connectivity_state_;
+      std::optional<grpc_connectivity_state> connectivity_state_;
       absl::Status connectivity_status_;
       bool seen_transient_failure_ = false;
     };
 
     SubchannelList(RefCountedPtr<PickFirst> policy,
                    EndpointAddressesIterator* addresses,
-                   const ChannelArgs& args);
+                   const ChannelArgs& args, absl::string_view resolution_note);
 
     ~SubchannelList() override;
 
+    void Orphan() override;
+
     // The number of subchannels in the list.
     size_t size() const { return subchannels_.size(); }
 
     // Resets connection backoff of all subchannels.
     void ResetBackoffLocked();
 
-    void Orphan() override;
-
     bool IsHappyEyeballsPassComplete() const {
       // Checking attempting_index_ here is just an optimization -- if
       // we haven't actually tried all subchannels yet, then we don't
       // need to iterate.
       if (attempting_index_ < size()) return false;
-      for (const SubchannelData& sd : subchannels_) {
-        if (!sd.seen_transient_failure()) return false;
+      for (const auto& sd : subchannels_) {
+        if (!sd->seen_transient_failure()) return false;
       }
       return true;
     }
 
+    void ReportTransientFailure(absl::Status status);
+
    private:
     // Returns true if all subchannels have seen their initial
     // connectivity state notifications.
@@ -260,9 +296,10 @@ class PickFirst final : public LoadBalancingPolicy {
     RefCountedPtr<PickFirst> policy_;
 
     ChannelArgs args_;
+    std::string resolution_note_;
 
     // The list of subchannels.
-    std::vector<SubchannelData> subchannels_;
+    std::vector<std::unique_ptr<SubchannelData>> subchannels_;
 
     // Is this list shutting down? This may be true due to the shutdown of the
     // policy itself or because a newer update has arrived while this one hadn't
@@ -276,7 +313,7 @@ class PickFirst final : public LoadBalancingPolicy {
     // initial pass is over, this will be equal to size().
     size_t attempting_index_ = 0;
     // Happy Eyeballs timer handle.
-    absl::optional<grpc_event_engine::experimental::EventEngine::TaskHandle>
+    std::optional<grpc_event_engine::experimental::EventEngine::TaskHandle>
         timer_handle_;
 
     // After the initial Happy Eyeballs pass, the number of failures
@@ -290,8 +327,9 @@ class PickFirst final : public LoadBalancingPolicy {
   class HealthWatcher final
       : public SubchannelInterface::ConnectivityStateWatcherInterface {
    public:
-    explicit HealthWatcher(RefCountedPtr<PickFirst> policy)
-        : policy_(std::move(policy)) {}
+    HealthWatcher(RefCountedPtr<PickFirst> policy,
+                  absl::string_view resolution_note)
+        : policy_(std::move(policy)), resolution_note_(resolution_note) {}
 
     ~HealthWatcher() override {
       policy_.reset(DEBUG_LOCATION, "HealthWatcher dtor");
@@ -306,6 +344,7 @@ class PickFirst final : public LoadBalancingPolicy {
 
    private:
     RefCountedPtr<PickFirst> policy_;
+    std::string resolution_note_;
   };
 
   class Picker final : public SubchannelPicker {
@@ -330,6 +369,8 @@ class PickFirst final : public LoadBalancingPolicy {
 
   void UnsetSelectedSubchannel();
 
+  void GoIdle();
+
   // When ExitIdleLocked() is called, we create a subchannel_list_ and start
   // trying to connect, but we don't actually change state_ until the first
   // subchannel reports CONNECTING.  So in order to know if we're really
@@ -347,12 +388,15 @@ class PickFirst final : public LoadBalancingPolicy {
 
   // Lateset update args.
   UpdateArgs latest_update_args_;
-  // All our subchannels.
+  // The list of subchannels that we're currently trying to connect to.
+  // Will generally be null when selected_ is set, except when we get a
+  // resolver update and need to check initial connectivity states for
+  // the new list to decide whether we keep using the existing
+  // connection or go IDLE.
   OrphanablePtr<SubchannelList> subchannel_list_;
-  // Latest pending subchannel list.
-  OrphanablePtr<SubchannelList> latest_pending_subchannel_list_;
-  // Selected subchannel in subchannel_list_.
-  SubchannelList::SubchannelData* selected_ = nullptr;
+  // Selected subchannel.  Will generally be null when subchannel_list_
+  // is non-null, with the exception mentioned above.
+  OrphanablePtr<SubchannelList::SubchannelData::SubchannelState> selected_;
   // Health watcher for the selected subchannel.
   SubchannelInterface::ConnectivityStateWatcherInterface* health_watcher_ =
       nullptr;
@@ -380,44 +424,32 @@ PickFirst::PickFirst(Args args)
                     .GetInt(GRPC_ARG_HAPPY_EYEBALLS_CONNECTION_ATTEMPT_DELAY_MS)
                     .value_or(250),
                 100, 2000))) {
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO, "Pick First %p created.", this);
-  }
+  GRPC_TRACE_LOG(pick_first, INFO) << "Pick First " << this << " created.";
 }
 
 PickFirst::~PickFirst() {
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO, "Destroying Pick First %p", this);
-  }
-  GPR_ASSERT(subchannel_list_ == nullptr);
-  GPR_ASSERT(latest_pending_subchannel_list_ == nullptr);
+  GRPC_TRACE_LOG(pick_first, INFO) << "Destroying Pick First " << this;
+  CHECK_EQ(subchannel_list_.get(), nullptr);
 }
 
 void PickFirst::ShutdownLocked() {
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO, "Pick First %p Shutting down", this);
-  }
+  GRPC_TRACE_LOG(pick_first, INFO) << "Pick First " << this << " Shutting down";
   shutdown_ = true;
   UnsetSelectedSubchannel();
   subchannel_list_.reset();
-  latest_pending_subchannel_list_.reset();
 }
 
 void PickFirst::ExitIdleLocked() {
   if (shutdown_) return;
   if (IsIdle()) {
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-      gpr_log(GPR_INFO, "Pick First %p exiting idle", this);
-    }
+    GRPC_TRACE_LOG(pick_first, INFO)
+        << "Pick First " << this << " exiting idle";
     AttemptToConnectUsingLatestUpdateArgsLocked();
   }
 }
 
 void PickFirst::ResetBackoffLocked() {
   if (subchannel_list_ != nullptr) subchannel_list_->ResetBackoffLocked();
-  if (latest_pending_subchannel_list_ != nullptr) {
-    latest_pending_subchannel_list_->ResetBackoffLocked();
-  }
 }
 
 void PickFirst::AttemptToConnectUsingLatestUpdateArgsLocked() {
@@ -426,37 +458,24 @@ void PickFirst::AttemptToConnectUsingLatestUpdateArgsLocked() {
   if (latest_update_args_.addresses.ok()) {
     addresses = latest_update_args_.addresses->get();
   }
-  // Replace latest_pending_subchannel_list_.
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace) &&
-      latest_pending_subchannel_list_ != nullptr) {
-    gpr_log(GPR_INFO,
-            "[PF %p] Shutting down previous pending subchannel list %p", this,
-            latest_pending_subchannel_list_.get());
+  // Replace subchannel_list_.
+  if (GRPC_TRACE_FLAG_ENABLED(pick_first) && subchannel_list_ != nullptr) {
+    LOG(INFO) << "[PF " << this << "] Shutting down previous subchannel list "
+              << subchannel_list_.get();
   }
-  latest_pending_subchannel_list_ = MakeOrphanable<SubchannelList>(
-      RefAsSubclass<PickFirst>(), addresses, latest_update_args_.args);
+  subchannel_list_ = MakeOrphanable<SubchannelList>(
+      RefAsSubclass<PickFirst>(DEBUG_LOCATION, "SubchannelList"), addresses,
+      latest_update_args_.args, latest_update_args_.resolution_note);
   // Empty update or no valid subchannels.  Put the channel in
-  // TRANSIENT_FAILURE and request re-resolution.
-  if (latest_pending_subchannel_list_->size() == 0) {
+  // TRANSIENT_FAILURE and request re-resolution.  Also unset the
+  // current selected subchannel.
+  if (subchannel_list_->size() == 0) {
     channel_control_helper()->RequestReresolution();
-    absl::Status status =
-        latest_update_args_.addresses.ok()
-            ? absl::UnavailableError(absl::StrCat(
-                  "empty address list: ", latest_update_args_.resolution_note))
-            : latest_update_args_.addresses.status();
-    UpdateState(GRPC_CHANNEL_TRANSIENT_FAILURE, status,
-                MakeRefCounted<TransientFailurePicker>(status));
-  }
-  // If the new update is empty or we don't yet have a selected subchannel in
-  // the current list, replace the current subchannel list immediately.
-  if (latest_pending_subchannel_list_->size() == 0 || selected_ == nullptr) {
+    absl::Status status = latest_update_args_.addresses.ok()
+                              ? absl::UnavailableError("empty address list")
+                              : latest_update_args_.addresses.status();
+    subchannel_list_->ReportTransientFailure(std::move(status));
     UnsetSelectedSubchannel();
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace) &&
-        subchannel_list_ != nullptr) {
-      gpr_log(GPR_INFO, "[PF %p] Shutting down previous subchannel list %p",
-              this, subchannel_list_.get());
-    }
-    subchannel_list_ = std::move(latest_pending_subchannel_list_);
   }
 }
 
@@ -490,12 +509,13 @@ class AddressFamilyIterator final {
 };
 
 absl::Status PickFirst::UpdateLocked(UpdateArgs args) {
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
+  if (GRPC_TRACE_FLAG_ENABLED(pick_first)) {
     if (args.addresses.ok()) {
-      gpr_log(GPR_INFO, "Pick First %p received update", this);
+      LOG(INFO) << "Pick First " << this << " received update";
     } else {
-      gpr_log(GPR_INFO, "Pick First %p received update with address error: %s",
-              this, args.addresses.status().ToString().c_str());
+      LOG(INFO) << "Pick First " << this
+                << " received update with address error: "
+                << args.addresses.status();
     }
   }
   // Set return status based on the address list.
@@ -579,11 +599,25 @@ void PickFirst::UnsetSelectedSubchannel() {
   if (selected_ != nullptr && health_data_watcher_ != nullptr) {
     selected_->subchannel()->CancelDataWatcher(health_data_watcher_);
   }
-  selected_ = nullptr;
+  selected_.reset();
   health_watcher_ = nullptr;
   health_data_watcher_ = nullptr;
 }
 
+void PickFirst::GoIdle() {
+  // Unset the selected subchannel.
+  UnsetSelectedSubchannel();
+  // Drop the current subchannel list, if any.
+  subchannel_list_.reset();
+  // Request a re-resolution.
+  // TODO(qianchengz): We may want to request re-resolution in
+  // ExitIdleLocked() instead.
+  channel_control_helper()->RequestReresolution();
+  // Enter idle.
+  UpdateState(GRPC_CHANNEL_IDLE, absl::OkStatus(),
+              MakeRefCounted<QueuePicker>(Ref(DEBUG_LOCATION, "QueuePicker")));
+}
+
 //
 // PickFirst::HealthWatcher
 //
@@ -591,11 +625,10 @@ void PickFirst::UnsetSelectedSubchannel() {
 void PickFirst::HealthWatcher::OnConnectivityStateChange(
     grpc_connectivity_state new_state, absl::Status status) {
   if (policy_->health_watcher_ != this) return;
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO, "[PF %p] health watch state update: %s (%s)",
-            policy_.get(), ConnectivityStateName(new_state),
-            status.ToString().c_str());
-  }
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << policy_.get()
+      << "] health watch state update: " << ConnectivityStateName(new_state)
+      << " (" << status << ")";
   switch (new_state) {
     case GRPC_CHANNEL_READY:
       policy_->channel_control_helper()->UpdateState(
@@ -613,168 +646,188 @@ void PickFirst::HealthWatcher::OnConnectivityStateChange(
           new_state, absl::OkStatus(),
           MakeRefCounted<QueuePicker>(policy_->Ref()));
       break;
-    case GRPC_CHANNEL_TRANSIENT_FAILURE:
+    case GRPC_CHANNEL_TRANSIENT_FAILURE: {
+      std::string message = absl::StrCat("health watch: ", status.message());
+      if (!resolution_note_.empty()) {
+        absl::StrAppend(&message, " (", resolution_note_, ")");
+      }
       policy_->channel_control_helper()->UpdateState(
           GRPC_CHANNEL_TRANSIENT_FAILURE, status,
-          MakeRefCounted<TransientFailurePicker>(status));
+          MakeRefCounted<TransientFailurePicker>(
+              absl::UnavailableError(message)));
       break;
+    }
     case GRPC_CHANNEL_SHUTDOWN:
       Crash("health watcher reported state SHUTDOWN");
   }
 }
 
 //
-// PickFirst::SubchannelList::SubchannelData
+// PickFirst::SubchannelList::SubchannelData::SubchannelState
 //
 
-PickFirst::SubchannelList::SubchannelData::SubchannelData(
-    SubchannelList* subchannel_list, size_t index,
+PickFirst::SubchannelList::SubchannelData::SubchannelState::SubchannelState(
+    SubchannelData* subchannel_data,
     RefCountedPtr<SubchannelInterface> subchannel)
-    : subchannel_list_(subchannel_list),
-      index_(index),
+    : subchannel_data_(subchannel_data),
+      pick_first_(subchannel_data_->subchannel_list_->policy_),
       subchannel_(std::move(subchannel)) {
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO,
-            "[PF %p] subchannel list %p index %" PRIuPTR
-            " (subchannel %p): starting watch",
-            subchannel_list_->policy_.get(), subchannel_list_, index_,
-            subchannel_.get());
-  }
-  auto watcher = std::make_unique<Watcher>(
-      subchannel_list_->Ref(DEBUG_LOCATION, "Watcher"), index_);
-  pending_watcher_ = watcher.get();
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << pick_first_.get() << "] subchannel state " << this
+      << " (subchannel " << subchannel_.get() << "): starting watch";
+  auto watcher = std::make_unique<Watcher>(Ref(DEBUG_LOCATION, "Watcher"));
+  watcher_ = watcher.get();
   subchannel_->WatchConnectivityState(std::move(watcher));
 }
 
-void PickFirst::SubchannelList::SubchannelData::ShutdownLocked() {
-  if (subchannel_ != nullptr) {
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-      gpr_log(GPR_INFO,
-              "[PF %p] subchannel list %p index %" PRIuPTR " of %" PRIuPTR
-              " (subchannel %p): cancelling watch and unreffing subchannel",
-              subchannel_list_->policy_.get(), subchannel_list_, index_,
-              subchannel_list_->size(), subchannel_.get());
+void PickFirst::SubchannelList::SubchannelData::SubchannelState::Orphan() {
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << pick_first_.get() << "] subchannel state " << this
+      << " (subchannel " << subchannel_.get()
+      << "): cancelling watch and unreffing subchannel";
+  subchannel_data_ = nullptr;
+  subchannel_->CancelConnectivityStateWatch(watcher_);
+  watcher_ = nullptr;
+  subchannel_.reset();
+  pick_first_.reset();
+  Unref();
+}
+
+void PickFirst::SubchannelList::SubchannelData::SubchannelState::Select() {
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "Pick First " << pick_first_.get() << " selected subchannel "
+      << subchannel_.get();
+  CHECK_NE(subchannel_data_, nullptr);
+  pick_first_->UnsetSelectedSubchannel();  // Cancel health watch, if any.
+  pick_first_->selected_ = std::move(subchannel_data_->subchannel_state_);
+  // If health checking is enabled, start the health watch, but don't
+  // report a new picker -- we want to stay in CONNECTING while we wait
+  // for the health status notification.
+  // If health checking is NOT enabled, report READY.
+  if (pick_first_->enable_health_watch_) {
+    GRPC_TRACE_LOG(pick_first, INFO)
+        << "[PF " << pick_first_.get() << "] starting health watch";
+    auto watcher = std::make_unique<HealthWatcher>(
+        pick_first_.Ref(DEBUG_LOCATION, "HealthWatcher"),
+        subchannel_data_->subchannel_list_->resolution_note_);
+    pick_first_->health_watcher_ = watcher.get();
+    auto health_data_watcher = MakeHealthCheckWatcher(
+        pick_first_->work_serializer(),
+        subchannel_data_->subchannel_list_->args_, std::move(watcher));
+    pick_first_->health_data_watcher_ = health_data_watcher.get();
+    subchannel_->AddDataWatcher(std::move(health_data_watcher));
+  } else {
+    pick_first_->UpdateState(GRPC_CHANNEL_READY, absl::OkStatus(),
+                             MakeRefCounted<Picker>(subchannel_));
+  }
+  // Report successful connection.
+  // We consider it a successful connection attempt only if the
+  // previous state was CONNECTING.  In particular, we don't want to
+  // increment this counter if we got a new address list and found the
+  // existing connection already in state READY.
+  if (subchannel_data_->connectivity_state_ == GRPC_CHANNEL_CONNECTING) {
+    auto& stats_plugins =
+        pick_first_->channel_control_helper()->GetStatsPluginGroup();
+    stats_plugins.AddCounter(
+        kMetricConnectionAttemptsSucceeded, 1,
+        {pick_first_->channel_control_helper()->GetTarget()}, {});
+  }
+  // Drop our pointer to subchannel_data_, so that we know not to
+  // interact with it on subsequent connectivity state updates.
+  subchannel_data_ = nullptr;
+  // Clean up subchannel list.
+  pick_first_->subchannel_list_.reset();
+}
+
+void PickFirst::SubchannelList::SubchannelData::SubchannelState::
+    OnConnectivityStateChange(grpc_connectivity_state new_state,
+                              absl::Status status) {
+  if (watcher_ == nullptr) return;
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << pick_first_.get() << "] subchannel state " << this
+      << " (subchannel " << subchannel_.get()
+      << "): connectivity changed: new_state="
+      << ConnectivityStateName(new_state) << ", status=" << status
+      << ", watcher=" << watcher_ << ", subchannel_data_=" << subchannel_data_
+      << ", pick_first_->selected_=" << pick_first_->selected_.get();
+  // If we're still part of a subchannel list trying to connect, check
+  // if we're connected.
+  if (subchannel_data_ != nullptr) {
+    CHECK_EQ(pick_first_->subchannel_list_.get(),
+             subchannel_data_->subchannel_list_);
+    // If the subchannel is READY, use it.
+    // Otherwise, tell the subchannel list to keep trying.
+    if (new_state == GRPC_CHANNEL_READY) {
+      Select();
+    } else {
+      subchannel_data_->OnConnectivityStateChange(new_state, std::move(status));
     }
-    subchannel_->CancelConnectivityStateWatch(pending_watcher_);
-    pending_watcher_ = nullptr;
-    subchannel_.reset();
+    return;
   }
+  // We aren't trying to connect, so we must be the selected subchannel.
+  CHECK_EQ(pick_first_->selected_.get(), this);
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "Pick First " << pick_first_.get()
+      << " selected subchannel connectivity changed to "
+      << ConnectivityStateName(new_state);
+  // Any state change is considered to be a failure of the existing
+  // connection.  Report the failure.
+  auto& stats_plugins =
+      pick_first_->channel_control_helper()->GetStatsPluginGroup();
+  stats_plugins.AddCounter(kMetricDisconnections, 1,
+                           {pick_first_->channel_control_helper()->GetTarget()},
+                           {});
+  // Report IDLE.
+  pick_first_->GoIdle();
+}
+
+//
+// PickFirst::SubchannelList::SubchannelData
+//
+
+PickFirst::SubchannelList::SubchannelData::SubchannelData(
+    SubchannelList* subchannel_list, size_t index,
+    RefCountedPtr<SubchannelInterface> subchannel)
+    : subchannel_list_(subchannel_list), index_(index) {
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << subchannel_list_->policy_.get() << "] subchannel list "
+      << subchannel_list_ << " index " << index_
+      << ": creating subchannel data";
+  subchannel_state_ =
+      MakeOrphanable<SubchannelState>(this, std::move(subchannel));
 }
 
 void PickFirst::SubchannelList::SubchannelData::OnConnectivityStateChange(
     grpc_connectivity_state new_state, absl::Status status) {
   PickFirst* p = subchannel_list_->policy_.get();
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(
-        GPR_INFO,
-        "[PF %p] subchannel list %p index %" PRIuPTR " of %" PRIuPTR
-        " (subchannel %p): connectivity changed: old_state=%s, new_state=%s, "
-        "status=%s, shutting_down=%d, pending_watcher=%p, "
-        "seen_transient_failure=%d, p->selected_=%p, "
-        "p->subchannel_list_=%p, p->latest_pending_subchannel_list_=%p",
-        p, subchannel_list_, index_, subchannel_list_->size(),
-        subchannel_.get(),
-        (connectivity_state_.has_value()
-             ? ConnectivityStateName(*connectivity_state_)
-             : "N/A"),
-        ConnectivityStateName(new_state), status.ToString().c_str(),
-        subchannel_list_->shutting_down_, pending_watcher_,
-        seen_transient_failure_, p->selected_, p->subchannel_list_.get(),
-        p->latest_pending_subchannel_list_.get());
-  }
-  if (subchannel_list_->shutting_down_ || pending_watcher_ == nullptr) return;
-  auto& stats_plugins =
-      subchannel_list_->policy_->channel_control_helper()
-          ->GetStatsPluginGroup();
-  // The notification must be for a subchannel in either the current or
-  // latest pending subchannel lists.
-  GPR_ASSERT(subchannel_list_ == p->subchannel_list_.get() ||
-             subchannel_list_ == p->latest_pending_subchannel_list_.get());
-  GPR_ASSERT(new_state != GRPC_CHANNEL_SHUTDOWN);
-  absl::optional<grpc_connectivity_state> old_state = connectivity_state_;
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << p << "] subchannel list " << subchannel_list_ << " index "
+      << index_ << " of " << subchannel_list_->size() << " (subchannel_state "
+      << subchannel_state_.get() << "): connectivity changed: old_state="
+      << (connectivity_state_.has_value()
+              ? ConnectivityStateName(*connectivity_state_)
+              : "N/A")
+      << ", new_state=" << ConnectivityStateName(new_state)
+      << ", status=" << status
+      << ", seen_transient_failure=" << seen_transient_failure_
+      << ", p->selected_=" << p->selected_.get()
+      << ", p->subchannel_list_=" << p->subchannel_list_.get()
+      << ", p->subchannel_list_->shutting_down_="
+      << p->subchannel_list_->shutting_down_;
+  if (subchannel_list_->shutting_down_) return;
+  // The notification must be for a subchannel in the current list.
+  CHECK_EQ(subchannel_list_, p->subchannel_list_.get());
+  // SHUTDOWN should never happen.
+  CHECK_NE(new_state, GRPC_CHANNEL_SHUTDOWN);
+  // READY should be caught by SubchannelState, in which case it will
+  // not call us in the first place.
+  CHECK_NE(new_state, GRPC_CHANNEL_READY);
+  // Update state.
+  std::optional<grpc_connectivity_state> old_state = connectivity_state_;
   connectivity_state_ = new_state;
   connectivity_status_ = std::move(status);
-  // Handle updates for the currently selected subchannel.
-  if (p->selected_ == this) {
-    GPR_ASSERT(subchannel_list_ == p->subchannel_list_.get());
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-      gpr_log(GPR_INFO,
-              "Pick First %p selected subchannel connectivity changed to %s", p,
-              ConnectivityStateName(new_state));
-    }
-    // Any state change is considered to be a failure of the existing
-    // connection.
-    stats_plugins.AddCounter(
-        kMetricDisconnections, 1,
-        {subchannel_list_->policy_->channel_control_helper()->GetTarget()}, {});
-    // TODO(roth): We could check the connectivity states of all the
-    // subchannels here, just in case one of them happens to be READY,
-    // and we could switch to that rather than going IDLE.
-    // Request a re-resolution.
-    // TODO(qianchengz): We may want to request re-resolution in
-    // ExitIdleLocked().
-    p->channel_control_helper()->RequestReresolution();
-    // If there is a pending update, switch to the pending update.
-    if (p->latest_pending_subchannel_list_ != nullptr) {
-      if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-        gpr_log(GPR_INFO,
-                "Pick First %p promoting pending subchannel list %p to "
-                "replace %p",
-                p, p->latest_pending_subchannel_list_.get(),
-                p->subchannel_list_.get());
-      }
-      p->UnsetSelectedSubchannel();
-      p->subchannel_list_ = std::move(p->latest_pending_subchannel_list_);
-      // Set our state to that of the pending subchannel list.
-      if (p->subchannel_list_->IsHappyEyeballsPassComplete()) {
-        status = absl::UnavailableError(absl::StrCat(
-            "selected subchannel failed; switching to pending update; "
-            "last failure: ",
-            p->subchannel_list_->last_failure_.ToString()));
-        p->UpdateState(GRPC_CHANNEL_TRANSIENT_FAILURE, status,
-                       MakeRefCounted<TransientFailurePicker>(status));
-      } else if (p->state_ != GRPC_CHANNEL_TRANSIENT_FAILURE) {
-        p->UpdateState(GRPC_CHANNEL_CONNECTING, absl::Status(),
-                       MakeRefCounted<QueuePicker>(nullptr));
-      }
-      return;
-    }
-    // Enter idle.
-    p->UnsetSelectedSubchannel();
-    p->subchannel_list_.reset();
-    p->UpdateState(
-        GRPC_CHANNEL_IDLE, absl::Status(),
-        MakeRefCounted<QueuePicker>(p->Ref(DEBUG_LOCATION, "QueuePicker")));
-    return;
-  }
-  // If we get here, there are two possible cases:
-  // 1. We do not currently have a selected subchannel, and the update is
-  //    for a subchannel in p->subchannel_list_ that we're trying to
-  //    connect to.  The goal here is to find a subchannel that we can
-  //    select.
-  // 2. We do currently have a selected subchannel, and the update is
-  //    for a subchannel in p->latest_pending_subchannel_list_.  The
-  //    goal here is to find a subchannel from the update that we can
-  //    select in place of the current one.
-  // If the subchannel is READY, use it.
-  if (new_state == GRPC_CHANNEL_READY) {
-    // We consider it a successful connection attempt only if the
-    // previous state was CONNECTING.  In particular, we don't want to
-    // increment this counter if we got a new address list and found the
-    // existing connection already in state READY.
-    if (old_state == GRPC_CHANNEL_CONNECTING) {
-      stats_plugins.AddCounter(
-          kMetricConnectionAttemptsSucceeded, 1,
-          {subchannel_list_->policy_->channel_control_helper()->GetTarget()},
-          {});
-    }
-    ProcessUnselectedReadyLocked();
-    return;
-  }
   // Make sure we note when a subchannel has seen TRANSIENT_FAILURE.
-  bool prev_seen_transient_failure = seen_transient_failure_;
   if (new_state == GRPC_CHANNEL_TRANSIENT_FAILURE) {
-    seen_transient_failure_ = true;
     subchannel_list_->last_failure_ = connectivity_status_;
   }
   // If this is the initial connectivity state update for this subchannel,
@@ -787,15 +840,31 @@ void PickFirst::SubchannelList::SubchannelData::OnConnectivityStateChange(
   if (!subchannel_list_->AllSubchannelsSeenInitialState()) return;
   // If we're still here and this is the initial connectivity state
   // notification for this subchannel, that means it was the last one to
-  // see its initial notification.  Start trying to connect, starting
-  // with the first subchannel.
+  // see its initial notification.  So we now have enough state to
+  // figure out how to proceed.
   if (!old_state.has_value()) {
-    subchannel_list_->StartConnectingNextSubchannel();
+    // If we already have a selected subchannel and we got here, that
+    // means that none of the subchannels on the new list are in READY
+    // state, which means that the address we're currently connected to
+    // is not in the new list.  In that case, we drop the current
+    // connection and report IDLE.
+    if (p->selected_ != nullptr) {
+      GRPC_TRACE_LOG(pick_first, INFO)
+          << "[PF " << p << "] subchannel list " << subchannel_list_
+          << ": new update has no subchannels in state READY; dropping "
+             "existing connection and going IDLE";
+      p->GoIdle();
+    } else {
+      // Start trying to connect, starting with the first subchannel.
+      subchannel_list_->StartConnectingNextSubchannel();
+    }
     return;
   }
   // We've already started trying to connect.  Any subchannel that
   // reports TF is a connection attempt failure.
   if (new_state == GRPC_CHANNEL_TRANSIENT_FAILURE) {
+    auto& stats_plugins = subchannel_list_->policy_->channel_control_helper()
+                              ->GetStatsPluginGroup();
     stats_plugins.AddCounter(
         kMetricConnectionAttemptsFailed, 1,
         {subchannel_list_->policy_->channel_control_helper()->GetTarget()}, {});
@@ -803,6 +872,8 @@ void PickFirst::SubchannelList::SubchannelData::OnConnectivityStateChange(
   // Otherwise, process connectivity state change.
   switch (*connectivity_state_) {
     case GRPC_CHANNEL_TRANSIENT_FAILURE: {
+      bool prev_seen_transient_failure =
+          std::exchange(seen_transient_failure_, true);
       // If this is the first failure we've seen on this subchannel,
       // then we're still in the Happy Eyeballs pass.
       if (!prev_seen_transient_failure && seen_transient_failure_) {
@@ -840,8 +911,7 @@ void PickFirst::SubchannelList::SubchannelData::OnConnectivityStateChange(
                    ? ""
                    : "failed to connect to all addresses; last error: "),
               connectivity_status_.ToString()));
-          p->UpdateState(GRPC_CHANNEL_TRANSIENT_FAILURE, status,
-                         MakeRefCounted<TransientFailurePicker>(status));
+          subchannel_list_->ReportTransientFailure(std::move(status));
         }
       }
       break;
@@ -851,15 +921,15 @@ void PickFirst::SubchannelList::SubchannelData::OnConnectivityStateChange(
       // into a mode where we immediately try to connect to every
       // subchannel in parallel.
       if (subchannel_list_->IsHappyEyeballsPassComplete()) {
-        subchannel_->RequestConnection();
+        subchannel_state_->RequestConnection();
       }
       break;
     case GRPC_CHANNEL_CONNECTING:
-      // Only update connectivity state in case 1, and only if we're not
-      // already in TRANSIENT_FAILURE.
-      if (subchannel_list_ == p->subchannel_list_.get() &&
-          p->state_ != GRPC_CHANNEL_TRANSIENT_FAILURE) {
-        p->UpdateState(GRPC_CHANNEL_CONNECTING, absl::Status(),
+      // Only update connectivity state only if we're not already in
+      // TRANSIENT_FAILURE.
+      // TODO(roth): Squelch duplicate CONNECTING updates.
+      if (p->state_ != GRPC_CHANNEL_TRANSIENT_FAILURE) {
+        p->UpdateState(GRPC_CHANNEL_CONNECTING, absl::OkStatus(),
                        MakeRefCounted<QueuePicker>(nullptr));
       }
       break;
@@ -870,177 +940,99 @@ void PickFirst::SubchannelList::SubchannelData::OnConnectivityStateChange(
 }
 
 void PickFirst::SubchannelList::SubchannelData::RequestConnectionWithTimer() {
-  GPR_ASSERT(connectivity_state_.has_value());
+  CHECK(connectivity_state_.has_value());
   if (connectivity_state_ == GRPC_CHANNEL_IDLE) {
-    subchannel_->RequestConnection();
+    subchannel_state_->RequestConnection();
   } else {
-    GPR_ASSERT(connectivity_state_ == GRPC_CHANNEL_CONNECTING);
+    CHECK_EQ(connectivity_state_.value(), GRPC_CHANNEL_CONNECTING);
   }
   // If this is not the last subchannel in the list, start the timer.
   if (index_ != subchannel_list_->size() - 1) {
     PickFirst* p = subchannel_list_->policy_.get();
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-      gpr_log(GPR_INFO,
-              "Pick First %p subchannel list %p: starting Connection "
-              "Attempt Delay timer for %" PRId64 "ms for index %" PRIuPTR,
-              p, subchannel_list_, p->connection_attempt_delay_.millis(),
-              index_);
-    }
+    GRPC_TRACE_LOG(pick_first, INFO)
+        << "Pick First " << p << " subchannel list " << subchannel_list_
+        << ": starting Connection Attempt Delay timer for "
+        << p->connection_attempt_delay_.millis() << "ms for index " << index_;
     subchannel_list_->timer_handle_ =
         p->channel_control_helper()->GetEventEngine()->RunAfter(
             p->connection_attempt_delay_,
             [subchannel_list =
                  subchannel_list_->Ref(DEBUG_LOCATION, "timer")]() mutable {
-              ApplicationCallbackExecCtx application_exec_ctx;
               ExecCtx exec_ctx;
               auto* sl = subchannel_list.get();
               sl->policy_->work_serializer()->Run(
                   [subchannel_list = std::move(subchannel_list)]() {
-                    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-                      gpr_log(GPR_INFO,
-                              "Pick First %p subchannel list %p: Connection "
-                              "Attempt Delay timer fired (shutting_down=%d, "
-                              "selected=%p)",
-                              subchannel_list->policy_.get(),
-                              subchannel_list.get(),
-                              subchannel_list->shutting_down_,
-                              subchannel_list->policy_->selected_);
-                    }
+                    GRPC_TRACE_LOG(pick_first, INFO)
+                        << "Pick First " << subchannel_list->policy_.get()
+                        << " subchannel list " << subchannel_list.get()
+                        << ": Connection Attempt Delay timer fired "
+                           "(shutting_down="
+                        << subchannel_list->shutting_down_ << ", selected="
+                        << subchannel_list->policy_->selected_.get() << ")";
                     if (subchannel_list->shutting_down_) return;
                     if (subchannel_list->policy_->selected_ != nullptr) return;
                     ++subchannel_list->attempting_index_;
                     subchannel_list->StartConnectingNextSubchannel();
-                  },
-                  DEBUG_LOCATION);
+                  });
             });
   }
 }
 
-void PickFirst::SubchannelList::SubchannelData::ProcessUnselectedReadyLocked() {
-  PickFirst* p = subchannel_list_->policy_.get();
-  // Cancel Happy Eyeballs timer, if any.
-  if (subchannel_list_->timer_handle_.has_value()) {
-    p->channel_control_helper()->GetEventEngine()->Cancel(
-        *subchannel_list_->timer_handle_);
-  }
-  // If we get here, there are two possible cases:
-  // 1. We do not currently have a selected subchannel, and the update is
-  //    for a subchannel in p->subchannel_list_ that we're trying to
-  //    connect to.  The goal here is to find a subchannel that we can
-  //    select.
-  // 2. We do currently have a selected subchannel, and the update is
-  //    for a subchannel in p->latest_pending_subchannel_list_.  The
-  //    goal here is to find a subchannel from the update that we can
-  //    select in place of the current one.
-  GPR_ASSERT(subchannel_list_ == p->subchannel_list_.get() ||
-             subchannel_list_ == p->latest_pending_subchannel_list_.get());
-  // Case 2.  Promote p->latest_pending_subchannel_list_ to p->subchannel_list_.
-  if (subchannel_list_ == p->latest_pending_subchannel_list_.get()) {
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-      gpr_log(GPR_INFO,
-              "Pick First %p promoting pending subchannel list %p to "
-              "replace %p",
-              p, p->latest_pending_subchannel_list_.get(),
-              p->subchannel_list_.get());
-    }
-    p->UnsetSelectedSubchannel();
-    p->subchannel_list_ = std::move(p->latest_pending_subchannel_list_);
-  }
-  // Cases 1 and 2.
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO, "Pick First %p selected subchannel %p", p,
-            subchannel_.get());
-  }
-  p->selected_ = this;
-  // If health checking is enabled, start the health watch, but don't
-  // report a new picker -- we want to stay in CONNECTING while we wait
-  // for the health status notification.
-  // If health checking is NOT enabled, report READY.
-  if (p->enable_health_watch_) {
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-      gpr_log(GPR_INFO, "[PF %p] starting health watch", p);
-    }
-    auto watcher = std::make_unique<HealthWatcher>(
-        p->RefAsSubclass<PickFirst>(DEBUG_LOCATION, "HealthWatcher"));
-    p->health_watcher_ = watcher.get();
-    auto health_data_watcher = MakeHealthCheckWatcher(
-        p->work_serializer(), subchannel_list_->args_, std::move(watcher));
-    p->health_data_watcher_ = health_data_watcher.get();
-    subchannel_->AddDataWatcher(std::move(health_data_watcher));
-  } else {
-    p->UpdateState(GRPC_CHANNEL_READY, absl::Status(),
-                   MakeRefCounted<Picker>(subchannel()->Ref()));
-  }
-  // Unref all other subchannels in the list.
-  for (size_t i = 0; i < subchannel_list_->size(); ++i) {
-    if (i != index_) {
-      subchannel_list_->subchannels_[i].ShutdownLocked();
-    }
-  }
-}
-
 //
 // PickFirst::SubchannelList
 //
 
 PickFirst::SubchannelList::SubchannelList(RefCountedPtr<PickFirst> policy,
                                           EndpointAddressesIterator* addresses,
-                                          const ChannelArgs& args)
+                                          const ChannelArgs& args,
+                                          absl::string_view resolution_note)
     : InternallyRefCounted<SubchannelList>(
-          GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace) ? "SubchannelList"
-                                                            : nullptr),
+          GRPC_TRACE_FLAG_ENABLED(pick_first) ? "SubchannelList" : nullptr),
       policy_(std::move(policy)),
-      args_(args.Remove(GRPC_ARG_INTERNAL_PICK_FIRST_ENABLE_HEALTH_CHECKING)
-                .Remove(
-                    GRPC_ARG_INTERNAL_PICK_FIRST_OMIT_STATUS_MESSAGE_PREFIX)) {
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO, "[PF %p] Creating subchannel list %p - channel args: %s",
-            policy_.get(), this, args_.ToString().c_str());
-  }
+      args_(
+          args.Remove(GRPC_ARG_INTERNAL_PICK_FIRST_ENABLE_HEALTH_CHECKING)
+              .Remove(GRPC_ARG_INTERNAL_PICK_FIRST_OMIT_STATUS_MESSAGE_PREFIX)),
+      resolution_note_(resolution_note) {
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << policy_.get() << "] Creating subchannel list " << this
+      << " - channel args: " << args_.ToString();
   if (addresses == nullptr) return;
   // Create a subchannel for each address.
   addresses->ForEach([&](const EndpointAddresses& address) {
-    GPR_ASSERT(address.addresses().size() == 1);
+    CHECK_EQ(address.addresses().size(), 1u);
     RefCountedPtr<SubchannelInterface> subchannel =
         policy_->channel_control_helper()->CreateSubchannel(
             address.address(), address.args(), args_);
     if (subchannel == nullptr) {
       // Subchannel could not be created.
-      if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-        gpr_log(GPR_INFO,
-                "[PF %p] could not create subchannel for address %s, ignoring",
-                policy_.get(), address.ToString().c_str());
-      }
+      GRPC_TRACE_LOG(pick_first, INFO)
+          << "[PF " << policy_.get()
+          << "] could not create subchannel for address " << address.ToString()
+          << ", ignoring";
       return;
     }
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-      gpr_log(GPR_INFO,
-              "[PF %p] subchannel list %p index %" PRIuPTR
-              ": Created subchannel %p for address %s",
-              policy_.get(), this, subchannels_.size(), subchannel.get(),
-              address.ToString().c_str());
-    }
-    subchannels_.emplace_back(this, subchannels_.size(), std::move(subchannel));
+    GRPC_TRACE_LOG(pick_first, INFO)
+        << "[PF " << policy_.get() << "] subchannel list " << this << " index "
+        << subchannels_.size() << ": Created subchannel " << subchannel.get()
+        << " for address " << address.ToString();
+    subchannels_.emplace_back(std::make_unique<SubchannelData>(
+        this, subchannels_.size(), std::move(subchannel)));
   });
 }
 
 PickFirst::SubchannelList::~SubchannelList() {
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO, "[PF %p] Destroying subchannel_list %p", policy_.get(),
-            this);
-  }
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << policy_.get() << "] Destroying subchannel_list " << this;
 }
 
 void PickFirst::SubchannelList::Orphan() {
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO, "[PF %p] Shutting down subchannel_list %p", policy_.get(),
-            this);
-  }
-  GPR_ASSERT(!shutting_down_);
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "[PF " << policy_.get() << "] Shutting down subchannel_list " << this;
+  CHECK(!shutting_down_);
   shutting_down_ = true;
-  for (auto& sd : subchannels_) {
-    sd.ShutdownLocked();
-  }
+  // Shut down subchannels.
+  subchannels_.clear();
+  // Cancel Happy Eyeballs timer, if any.
   if (timer_handle_.has_value()) {
     policy_->channel_control_helper()->GetEventEngine()->Cancel(*timer_handle_);
   }
@@ -1049,23 +1041,33 @@ void PickFirst::SubchannelList::Orphan() {
 
 void PickFirst::SubchannelList::ResetBackoffLocked() {
   for (auto& sd : subchannels_) {
-    sd.ResetBackoffLocked();
+    sd->ResetBackoffLocked();
   }
 }
 
+void PickFirst::SubchannelList::ReportTransientFailure(absl::Status status) {
+  if (!resolution_note_.empty()) {
+    status = absl::Status(status.code(), absl::StrCat(status.message(), " (",
+                                                      resolution_note_, ")"));
+  }
+  policy_->UpdateState(GRPC_CHANNEL_TRANSIENT_FAILURE, status,
+                       MakeRefCounted<TransientFailurePicker>(status));
+}
+
 void PickFirst::SubchannelList::StartConnectingNextSubchannel() {
   // Find the next subchannel not in state TRANSIENT_FAILURE.
   // We skip subchannels in state TRANSIENT_FAILURE to avoid a
   // large recursion that could overflow the stack.
   for (; attempting_index_ < size(); ++attempting_index_) {
-    SubchannelData* sc = &subchannels_[attempting_index_];
-    GPR_ASSERT(sc->connectivity_state().has_value());
+    SubchannelData* sc = subchannels_[attempting_index_].get();
+    CHECK(sc->connectivity_state().has_value());
     if (sc->connectivity_state() != GRPC_CHANNEL_TRANSIENT_FAILURE) {
       // Found a subchannel not in TRANSIENT_FAILURE, so trigger a
       // connection attempt.
       sc->RequestConnectionWithTimer();
       return;
     }
+    sc->set_seen_transient_failure();
   }
   // If we didn't find a subchannel to request a connection on, check to
   // see if the Happy Eyeballs pass is complete.
@@ -1079,47 +1081,27 @@ void PickFirst::SubchannelList::MaybeFinishHappyEyeballsPass() {
   // We didn't find another subchannel not in state TRANSIENT_FAILURE,
   // so report TRANSIENT_FAILURE and switch to a mode in which we try to
   // connect to all addresses in parallel.
-  if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-    gpr_log(GPR_INFO,
-            "Pick First %p subchannel list %p failed to connect to "
-            "all subchannels",
-            policy_.get(), this);
-  }
-  // In case 2, swap to the new subchannel list.  This means reporting
-  // TRANSIENT_FAILURE and dropping the existing (working) connection,
-  // but we can't ignore what the control plane has told us.
-  if (policy_->latest_pending_subchannel_list_.get() == this) {
-    if (GRPC_TRACE_FLAG_ENABLED(grpc_lb_pick_first_trace)) {
-      gpr_log(GPR_INFO,
-              "Pick First %p promoting pending subchannel list %p to "
-              "replace %p",
-              policy_.get(), policy_->latest_pending_subchannel_list_.get(),
-              this);
-    }
-    policy_->UnsetSelectedSubchannel();
-    policy_->subchannel_list_ =
-        std::move(policy_->latest_pending_subchannel_list_);
-  }
-  // If this is the current subchannel list (either because we were
-  // in case 1 or because we were in case 2 and just promoted it to
-  // be the current list), re-resolve and report new state.
-  if (policy_->subchannel_list_.get() == this) {
-    policy_->channel_control_helper()->RequestReresolution();
-    absl::Status status = absl::UnavailableError(
-        absl::StrCat((policy_->omit_status_message_prefix_
-                          ? ""
-                          : "failed to connect to all addresses; last error: "),
-                     last_failure_.ToString()));
-    policy_->UpdateState(GRPC_CHANNEL_TRANSIENT_FAILURE, status,
-                         MakeRefCounted<TransientFailurePicker>(status));
-  }
+  GRPC_TRACE_LOG(pick_first, INFO)
+      << "Pick First " << policy_.get() << " subchannel list " << this
+      << " failed to connect to all subchannels";
+  // Re-resolve and report TRANSIENT_FAILURE.
+  policy_->channel_control_helper()->RequestReresolution();
+  absl::Status status = absl::UnavailableError(
+      absl::StrCat((policy_->omit_status_message_prefix_
+                        ? ""
+                        : "failed to connect to all addresses; last error: "),
+                   last_failure_.ToString()));
+  ReportTransientFailure(std::move(status));
+  // Drop the existing (working) connection, if any.  This may be
+  // sub-optimal, but we can't ignore what the control plane told us.
+  policy_->UnsetSelectedSubchannel();
   // We now transition into a mode where we try to connect to all
   // subchannels in parallel.  For any subchannel currently in IDLE,
   // trigger a connection attempt.  For any subchannel not currently in
   // IDLE, we will trigger a connection attempt when it does report IDLE.
-  for (SubchannelData& sd : subchannels_) {
-    if (sd.connectivity_state() == GRPC_CHANNEL_IDLE) {
-      sd.RequestConnection();
+  for (auto& sd : subchannels_) {
+    if (sd->connectivity_state() == GRPC_CHANNEL_IDLE) {
+      sd->RequestConnection();
     }
   }
 }