improve node deletion in hive (#7218)

Author: vporyadke

ydb/core/mind/hive/hive_impl.cpp

@@ -3,6 +3,7 @@
 #include <ydb/core/cms/console/console.h>
 #include <ydb/core/cms/console/configs_dispatcher.h>
 #include <ydb/core/protos/counters_hive.pb.h>
+#include <ydb/core/protos/node_broker.pb.h>
 #include <ydb/core/util/tuples.h>
 #include <ydb/library/yverify_stream/yverify_stream.h>
 #include <ydb/library/actors/interconnect/interconnect.h>
@@ -96,10 +97,12 @@ void THive::RestartPipeTx(ui64 tabletId) {
 
 bool THive::TryToDeleteNode(TNodeInfo* node) {
     if (node->CanBeDeleted()) {
+        BLOG_I("TryToDeleteNode(" << node->Id << "): deleting");
         DeleteNode(node->Id);
         return true;
     }
     if (!node->DeletionScheduled) {
+        BLOG_D("TryToDeleteNode(" << node->Id << "): waiting " << GetNodeDeletePeriod());
         Schedule(GetNodeDeletePeriod(), new TEvPrivate::TEvDeleteNode(node->Id));
         node->DeletionScheduled = true;
     }
@@ -989,8 +992,9 @@ void THive::OnActivateExecutor(const TActorContext&) {
     BuildLocalConfig();
     ClusterConfig = AppData()->HiveConfig;
     SpreadNeighbours = ClusterConfig.GetSpreadNeighbours();
+    NodeBrokerEpoch = TDuration::MicroSeconds(NKikimrNodeBroker::TConfig().GetEpochDuration());
     Send(NConsole::MakeConfigsDispatcherID(SelfId().NodeId()),
-        new NConsole::TEvConfigsDispatcher::TEvSetConfigSubscriptionRequest(NKikimrConsole::TConfigItem::HiveConfigItem));
+        new NConsole::TEvConfigsDispatcher::TEvSetConfigSubscriptionRequest({NKikimrConsole::TConfigItem::HiveConfigItem, NKikimrConsole::TConfigItem::NodeBrokerConfigItem}));
     Execute(CreateInitScheme());
     if (!ResponsivenessPinger) {
         ResponsivenessPinger = new TTabletResponsivenessPinger(TabletCounters->Simple()[NHive::COUNTER_RESPONSE_TIME_USEC], TDuration::Seconds(1));
@@ -2218,7 +2222,9 @@ void THive::Handle(TEvHive::TEvInitiateTabletExternalBoot::TPtr& ev) {
 void THive::Handle(NConsole::TEvConsole::TEvConfigNotificationRequest::TPtr& ev) {
     const NKikimrConsole::TConfigNotificationRequest& record = ev->Get()->Record;
     ClusterConfig = record.GetConfig().GetHiveConfig();
-    BLOG_D("Received TEvConsole::TEvConfigNotificationRequest with update of cluster config: " << ClusterConfig.ShortDebugString());
+    NodeBrokerEpoch = TDuration::MicroSeconds(record.GetConfig().GetNodeBrokerConfig().GetEpochDuration());
+    BLOG_D("Received TEvConsole::TEvConfigNotificationRequest with update of cluster config: " << ClusterConfig.ShortDebugString() 
+           << "; " << record.GetConfig().GetNodeBrokerConfig().ShortDebugString());
     BuildCurrentConfig();
     Send(ev->Sender, new NConsole::TEvConsole::TEvConfigNotificationResponse(record), 0, ev->Cookie);
 }

ydb/core/mind/hive/hive_impl.h

@@ -445,6 +445,7 @@ class THive : public TActor<THive>, public TTabletExecutedFlat, public THiveShar
 
     NKikimrConfig::THiveConfig ClusterConfig;
     NKikimrConfig::THiveConfig DatabaseConfig;
+    TDuration NodeBrokerEpoch;
     std::unordered_map<TTabletTypes::EType, NKikimrConfig::THiveTabletLimit> TabletLimit; // built from CurrentConfig
     std::unordered_map<TTabletTypes::EType, NKikimrHive::TDataCentersPreference> DefaultDataCentersPreference;
     std::unordered_map<TDataCenterId, std::unordered_set<TNodeId>> RegisteredDataCenterNodes;
@@ -740,7 +741,11 @@ TTabletInfo* FindTabletEvenInDeleting(TTabletId tabletId, TFollowerId followerId
     }
 
     TDuration GetNodeDeletePeriod() const {
-        return TDuration::Seconds(CurrentConfig.GetNodeDeletePeriod());
+        if (CurrentConfig.HasNodeDeletePeriod()) {
+            return TDuration::Seconds(CurrentConfig.GetNodeDeletePeriod());
+        } else {
+            return NodeBrokerEpoch;
+        }
     }
 
     ui64 GetDrainInflight() const {

ydb/core/mind/hive/tx__load_everything.cpp

@@ -338,9 +338,9 @@ class TTxLoadEverything : public TTransactionBase<THive> {
                     // it's safe to call here, because there is no any tablets in the node yet
                     node.BecomeDisconnected();
                 }
-                if (node.CanBeDeleted()) {
+                if (Self->TryToDeleteNode(&node)) {
+                    // node is deleted from hashmap
                     db.Table<Schema::Node>().Key(nodeId).Delete();
-                    Self->Nodes.erase(nodeId);
                 } else if (node.IsUnknown() && node.LocationAcquired) {
                     Self->AddRegisteredDataCentersNode(node.Location.GetDataCenterId(), node.Id);
                 }

ydb/core/protos/config.proto

@@ -1439,7 +1439,7 @@ message THiveConfig {
     optional double MinPeriodBetweenReassign = 33 [default = 300.0]; // seconds
     optional double TabletRestartWatchPeriod = 34 [default = 3600.0]; // seconds
     optional double NodeRestartWatchPeriod = 35 [default = 3600.0]; // seconds
-    optional uint64 NodeDeletePeriod = 36 [default = 14400]; // seconds
+    optional uint64 NodeDeletePeriod = 36 [default = 3600]; // seconds
     repeated THiveTabletLimit DefaultTabletLimit = 37;
     repeated THiveTabletPreference DefaultTabletPreference = 38;
     optional uint64 SystemTabletCategoryId = 39 [default = 1];