25-1-1: add sensor for down nodes (#11592) (#18477)

Author: vporyadke

ydb/core/mind/hive/hive_impl.cpp

@@ -1708,6 +1708,22 @@ void THive::UpdateCounterTabletChannelHistorySize() {
     }
 }
 
+void THive::UpdateCounterNodesDown(i64 nodesDownDiff) {
+    if (TabletCounters != nullptr) {
+        auto& counter = TabletCounters->Simple()[NHive::COUNTER_NODES_DOWN];
+        auto newValue = counter.Get() + nodesDownDiff;
+        counter.Set(newValue);
+    }
+}
+
+void THive::UpdateCounterNodesFrozen(i64 nodesFrozenDiff) {
+    if (TabletCounters != nullptr) {
+        auto& counter = TabletCounters->Simple()[NHive::COUNTER_NODES_FROZEN];
+        auto newValue = counter.Get() + nodesFrozenDiff;
+        counter.Set(newValue);
+    }
+}
+
 void THive::RecordTabletMove(const TTabletMoveInfo& moveInfo) {
     TabletMoveHistory.PushBack(moveInfo);
     TabletCounters->Cumulative()[NHive::COUNTER_TABLETS_MOVED].Increment(1);

ydb/core/mind/hive/hive_impl.h

@@ -679,6 +679,8 @@ TTabletInfo* FindTabletEvenInDeleting(TTabletId tabletId, TFollowerId followerId
     void UpdateCounterTabletsStarting(i64 tabletsStartingDiff);
     void UpdateCounterPingQueueSize();
     void UpdateCounterTabletChannelHistorySize();
+    void UpdateCounterNodesDown(i64 nodesDownDiff);
+    void UpdateCounterNodesFrozen(i64 nodesFrozenDiff);
     void RecordTabletMove(const TTabletMoveInfo& info);
     bool DomainHasNodes(const TSubDomainKey &domainKey) const;
     void ProcessBootQueue();

ydb/core/mind/hive/node_info.cpp

@@ -360,6 +360,7 @@ void TNodeInfo::SendReconnect(const TActorId& local) {
 }
 
 void TNodeInfo::SetDown(bool down) {
+    Hive.UpdateCounterNodesDown(static_cast<i64>(down) - static_cast<i64>(Down));
     Down = down;
     if (Down) {
         Hive.ObjectDistributions.RemoveNode(*this);
@@ -370,6 +371,7 @@ void TNodeInfo::SetDown(bool down) {
 }
 
 void TNodeInfo::SetFreeze(bool freeze) {
+    Hive.UpdateCounterNodesFrozen(static_cast<i64>(freeze) - static_cast<i64>(Freeze));
     Freeze = freeze;
     if (Freeze) {
         for (const auto& [state, tablets] : Tablets) {

ydb/core/mind/hive/tx__load_everything.cpp

@@ -326,6 +326,9 @@ class TTxLoadEverything : public TTransactionBase<THive> {
                     // That was not persisted to avoid issues with downgrades
                     node.Down = true;
                 }
+                if (node.Down) {
+                    Self->UpdateCounterNodesDown(+1);
+                }
                 if (nodeRowset.HaveValue<Schema::Node::Location>()) {
                     auto location = nodeRowset.GetValue<Schema::Node::Location>();
                     if (location.HasDataCenter()) {

ydb/core/protos/counters_hive.proto

@@ -34,6 +34,8 @@ enum ESimpleCounters {
     COUNTER_NODES_RECOMMENDED = 24 [(CounterOpts) = {Name: "NodesRecommended"}];
     COUNTER_NODES_RECOMMENDED_DRY_RUN = 25 [(CounterOpts) = {Name: "NodesRecommendedDryRun"}];
     COUNTER_AVG_CPU_UTILIZATION = 26 [(CounterOpts) = {Name: "AvgCPUUtilization"}];
+    COUNTER_NODES_DOWN = 27 [(CounterOpts) = {Name: "NodesDown"}];
+    COUNTER_NODES_FROZEN = 28 [(CounterOpts) = {Name: "NodesFrozen"}];
 }
 
 enum ECumulativeCounters {