do not trigger emergency balancer when all nodes have high usage (#6532)

Author: vporyadke

ydb/core/mind/hive/hive_impl.cpp

@@ -2329,7 +2329,8 @@ void THive::Handle(TEvPrivate::TEvProcessTabletBalancer::TPtr&) {
         nodeUsageHistogram.IncrementFor(record.Usage * 100);
     }
 
-    if (stats.MaxUsage >= GetMaxNodeUsageToKick()) {
+    double minUsageToKick = GetMaxNodeUsageToKick() - GetNodeUsageRangeToKick();
+    if (stats.MaxUsage >= GetMaxNodeUsageToKick() && stats.MinUsage < minUsageToKick) {
         std::vector<TNodeId> overloadedNodes;
         for (const auto& [nodeId, nodeInfo] : Nodes) {
             if (nodeInfo.IsAlive() && !nodeInfo.Down && nodeInfo.IsOverloaded()) {

ydb/core/mind/hive/hive_impl.h

@@ -934,6 +934,10 @@ TTabletInfo* FindTabletEvenInDeleting(TTabletId tabletId, TFollowerId followerId
         return CurrentConfig.GetStorageBalancerInflight();
     }
 
+    double GetNodeUsageRangeToKick() const {
+        return CurrentConfig.GetNodeUsageRangeToKick();
+    }
+
     static void ActualizeRestartStatistics(google::protobuf::RepeatedField<google::protobuf::uint64>& restartTimestamps, ui64 barrier);
     static ui64 GetRestartsPerPeriod(const google::protobuf::RepeatedField<google::protobuf::uint64>& restartTimestamps, ui64 barrier);
     static bool IsSystemTablet(TTabletTypes::EType type);

ydb/core/mind/hive/hive_ut.cpp

@@ -3996,6 +3996,7 @@ Y_UNIT_TEST_SUITE(THiveTest) {
             // this value of MaxNodeUsageToKick is selected specifically to make test scenario work
             // in link with number of tablets and values of network usage metrics used below
             app.HiveConfig.SetMaxNodeUsageToKick(0.01);
+            app.HiveConfig.SetNodeUsageRangeToKick(0);
             app.HiveConfig.SetEmergencyBalancerInflight(1); // to ensure fair distribution
         });
 
@@ -4968,6 +4969,75 @@ Y_UNIT_TEST_SUITE(THiveTest) {
         UNIT_ASSERT_VALUES_EQUAL(newDistribution[1].size(), TABLETS_PER_NODE - 1);
     }
 
+    Y_UNIT_TEST(TestHiveBalancerHighUsage) {
+        static constexpr ui64 NUM_NODES = 2;
+        TTestBasicRuntime runtime(2, false);
+        Setup(runtime, true, 1, [](TAppPrepare& app) {
+            app.HiveConfig.SetTabletKickCooldownPeriod(0);
+            app.HiveConfig.SetResourceChangeReactionPeriod(0);
+        });
+        const int nodeBase = runtime.GetNodeId(0);
+        TActorId senderA = runtime.AllocateEdgeActor();
+        const ui64 hiveTablet = MakeDefaultHiveID();
+        const ui64 testerTablet = MakeTabletID(false, 1);
+
+        auto getDistribution = [hiveTablet, nodeBase, senderA, &runtime]() -> std::array<std::vector<ui64>, NUM_NODES> {
+            std::array<std::vector<ui64>, NUM_NODES> nodeTablets = {};
+            {
+                runtime.SendToPipe(hiveTablet, senderA, new TEvHive::TEvRequestHiveInfo());
+                TAutoPtr<IEventHandle> handle;
+                TEvHive::TEvResponseHiveInfo* response = runtime.GrabEdgeEventRethrow<TEvHive::TEvResponseHiveInfo>(handle);
+                for (const NKikimrHive::TTabletInfo& tablet : response->Record.GetTablets()) {
+                    UNIT_ASSERT_C(((int)tablet.GetNodeID() - nodeBase >= 0) && (tablet.GetNodeID() - nodeBase < NUM_NODES),
+                            "nodeId# " << tablet.GetNodeID() << " nodeBase# " << nodeBase);
+                    nodeTablets[tablet.GetNodeID() - nodeBase].push_back(tablet.GetTabletID());
+                }
+            }
+            return nodeTablets;
+        };
+
+        CreateTestBootstrapper(runtime, CreateTestTabletInfo(hiveTablet, TTabletTypes::Hive), &CreateDefaultHive);
+
+        // wait for creation of nodes
+        {
+            TDispatchOptions options;
+            options.FinalEvents.emplace_back(TEvLocal::EvStatus, NUM_NODES);
+            runtime.DispatchEvents(options);
+        }
+
+        TTabletTypes::EType tabletType = TTabletTypes::Dummy;
+        for (size_t i = 0; i < 2; ++i) {
+            THolder<TEvHive::TEvCreateTablet> ev(new TEvHive::TEvCreateTablet(testerTablet, 100500 + i, tabletType, BINDED_CHANNELS));
+            ev->Record.SetObjectId(i);
+            ui64 tabletId = SendCreateTestTablet(runtime, hiveTablet, testerTablet, std::move(ev), 0, true);
+            MakeSureTabletIsUp(runtime, tabletId, 0);
+        }
+
+        auto initialDistribution = getDistribution();
+
+        std::array<double, NUM_NODES> usages = {.89, .91};
+        for (ui32 i = 0; i < 2; ++i) {
+            for (ui32 node = 0; node < NUM_NODES; ++node) {
+                TActorId sender = runtime.AllocateEdgeActor(node);
+                THolder<TEvHive::TEvTabletMetrics> metrics = MakeHolder<TEvHive::TEvTabletMetrics>();
+                metrics->Record.SetTotalNodeUsage(usages[node]);
+
+                runtime.SendToPipe(hiveTablet, sender, metrics.Release(), node);
+            }
+        }
+
+        {
+            TDispatchOptions options;
+            options.FinalEvents.emplace_back(NHive::TEvPrivate::EvBalancerOut);
+            runtime.DispatchEvents(options, TDuration::Seconds(10));
+        }
+
+        // Check that balancer moved no tablets
+        auto newDistribution = getDistribution();
+
+        UNIT_ASSERT_EQUAL(initialDistribution, newDistribution);
+    }
+
     Y_UNIT_TEST(TestUpdateTabletsObjectUpdatesMetrics) {
         TTestBasicRuntime runtime(1, false);
         Setup(runtime, true);

ydb/core/mind/hive/monitoring.cpp

@@ -794,6 +794,7 @@ class TTxMonEvent_Settings : public TTransactionBase<THive>, public TLoggedMonTr
         UpdateConfig(db, "MinNetworkScatterToBalance", configUpdates);
         UpdateConfig(db, "MinCounterScatterToBalance", configUpdates);
         UpdateConfig(db, "MaxNodeUsageToKick", configUpdates, TSchemeIds::State::MaxNodeUsageToKick);
+        UpdateConfig(db, "NodeUsageRangeToKick", configUpdates);
         UpdateConfig(db, "ResourceChangeReactionPeriod", configUpdates, TSchemeIds::State::ResourceChangeReactionPeriod);
         UpdateConfig(db, "TabletKickCooldownPeriod", configUpdates, TSchemeIds::State::TabletKickCooldownPeriod);
         UpdateConfig(db, "SpreadNeighbours", configUpdates, TSchemeIds::State::SpreadNeighbours);
@@ -1140,6 +1141,7 @@ class TTxMonEvent_Settings : public TTransactionBase<THive>, public TLoggedMonTr
         ShowConfig(out, "MinCounterScatterToBalance");
         ShowConfig(out, "MinNodeUsageToBalance");
         ShowConfig(out, "MaxNodeUsageToKick");
+        ShowConfig(out, "NodeUsageRangeToKick");
         ShowConfig(out, "ResourceChangeReactionPeriod");
         ShowConfig(out, "TabletKickCooldownPeriod");
         ShowConfig(out, "NodeSelectStrategy");

ydb/core/protos/config.proto

@@ -1470,6 +1470,7 @@ message THiveConfig {
     optional double MinGroupUsageToBalance = 72 [default = 0.1];
     optional uint64 StorageBalancerInflight = 73 [default = 1];
     optional bool EnableDestroyOperations = 74 [default = false];
+    optional double NodeUsageRangeToKick = 75 [default = 0.2];
 }
 
 message TBlobCacheConfig {