fix cs latency spikes

zverevgeny · zverevgeny · commit 899de9dd8a1c · 2024-07-04T17:44:24.000+03:00
diff --git a/ydb/core/tx/columnshard/background_controller.h b/ydb/core/tx/columnshard/background_controller.h
@@ -17,6 +17,7 @@ class TBackgroundController {
 
     bool ActiveCleanupPortions = false;
     bool ActiveCleanupTables = false;
+    bool ActiveCleanupInsertTable = false;
     YDB_READONLY(TMonotonic, LastIndexationInstant, TMonotonic::Zero());
 public:
     THashSet<NOlap::TPortionAddress> GetConflictTTLPortions() const;
@@ -66,6 +67,18 @@ class TBackgroundController {
     bool IsCleanupTablesActive() const {
         return ActiveCleanupTables;
     }
+
+    void StartCleanupInsertTable() {
+        Y_ABORT_UNLESS(!ActiveCleanupInsertTable);
+        ActiveCleanupInsertTable = true;
+    }
+    void FinishCleanupInsertTable() {
+        Y_ABORT_UNLESS(ActiveCleanupInsertTable);
+        ActiveCleanupInsertTable = false;
+    }
+    bool IsCleanupInsertTableActive() const {
+        return ActiveCleanupInsertTable;
+    }
 };
 
 }
diff --git a/ydb/core/tx/columnshard/blobs_action/abstract/blob_set.h b/ydb/core/tx/columnshard/blobs_action/abstract/blob_set.h
@@ -89,7 +89,18 @@ class TTabletsByBlob {
     bool IsEmpty() const {
         return Data.empty();
     }
-
+    using TGenStep = std::tuple<ui32, ui32>;
+    std::deque<TUnifiedBlobId> GroupByGenStep() const {
+        std::deque<TUnifiedBlobId> result;
+        for (const auto& i : Data) {
+            result.emplace_back(i.first);
+        }
+        const auto pred = [](const TUnifiedBlobId& l, const TUnifiedBlobId& r) {
+            return TGenStep(l.GetLogoBlobId().Generation(), l.GetLogoBlobId().Step()) < TGenStep(r.GetLogoBlobId().Generation(), r.GetLogoBlobId().Step());
+        };
+        std::sort(result.begin(), result.end(), pred);
+        return result;
+    }
     template <class TFilter>
     TTabletsByBlob ExtractBlobs(const TFilter& filter, const std::optional<ui32> countLimit = {}) {
         TTabletsByBlob result;
diff --git a/ydb/core/tx/columnshard/blobs_action/bs/blob_manager.cpp b/ydb/core/tx/columnshard/blobs_action/bs/blob_manager.cpp
@@ -211,7 +211,7 @@ std::vector<TGenStep> TBlobManager::FindNewGCBarriers() {
     return result;
 }
 
-std::shared_ptr<NBlobOperations::NBlobStorage::TGCTask> TBlobManager::BuildGCTask(const TString& storageId, 
+std::shared_ptr<NBlobOperations::NBlobStorage::TGCTask> TBlobManager::BuildGCTask(const TString& storageId,
     const std::shared_ptr<TBlobManager>& manager, const std::shared_ptr<NDataSharing::TStorageSharedBlobsManager>& sharedBlobsInfo,
     const std::shared_ptr<NBlobOperations::TRemoveGCCounters>& counters) noexcept {
     AFL_VERIFY(!CollectGenStepInFlight);
@@ -244,7 +244,7 @@ std::shared_ptr<NBlobOperations::NBlobStorage::TGCTask> TBlobManager::BuildGCTas
         }
     }
 
-    static const ui32 blobsGCCountLimit = 50000;
+    static const ui32 blobsGCCountLimit = 500000;
 
     const auto predShared = [&](const TUnifiedBlobId& id, const THashSet<TTabletId>& /*tabletIds*/) {
         return id.GetLogoBlobId().TabletID() != (ui64)SelfTabletId;
@@ -262,9 +262,8 @@ std::shared_ptr<NBlobOperations::NBlobStorage::TGCTask> TBlobManager::BuildGCTas
 
         TTabletsByBlob extractedOld = BlobsToDelete.ExtractBlobs(predRemoveOld, blobsGCCountLimit - extractedToRemoveFromDB.GetSize());
         extractedToRemoveFromDB.Add(extractedOld);
-        TTabletId tabletId;
-        TUnifiedBlobId unifiedBlobId;
-        while (extractedOld.ExtractFront(tabletId, unifiedBlobId)) {
+        for (TTabletsByBlob::TIterator itExtractedOld(extractedOld); itExtractedOld.IsValid(); ++itExtractedOld) {
+            const TUnifiedBlobId unifiedBlobId = itExtractedOld.GetBlobId();
             auto logoBlobId = unifiedBlobId.GetLogoBlobId();
             AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("to_delete_gc", logoBlobId);
             NBlobOperations::NBlobStorage::TGCTask::TGCLists& gl = perGroupGCListsInFlight[unifiedBlobId.GetDsGroup()];
@@ -278,6 +277,10 @@ std::shared_ptr<NBlobOperations::NBlobStorage::TGCTask> TBlobManager::BuildGCTas
 
 
     std::deque<TUnifiedBlobId> keepsToErase;
+    std::deque<TUnifiedBlobId> deleteIndex;
+    if (extractedToRemoveFromDB.GetSize() + keepsToErase.size() < blobsGCCountLimit) {
+        deleteIndex = BlobsToDelete.GroupByGenStep();
+    }
     for (auto&& newCollectGenStep : newCollectGenSteps) {
         AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("event", "PreparePerGroupGCRequests")("gen", std::get<0>(newCollectGenStep))("step", std::get<1>(newCollectGenStep));
         BlobsManagerCounters.OnNewCollectStep(std::get<0>(newCollectGenStep), std::get<1>(newCollectGenStep));
@@ -297,20 +300,38 @@ std::shared_ptr<NBlobOperations::NBlobStorage::TGCTask> TBlobManager::BuildGCTas
                 perGroupGCListsInFlight[blobGroup].KeepList.insert(*keepBlobIt);
                 keepsToErase.emplace_back(TUnifiedBlobId(blobGroup, *keepBlobIt));
                 AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("to_keep_gc", *keepBlobIt);
+                if (extractedToRemoveFromDB.GetSize() + keepsToErase.size() > blobsGCCountLimit) {
+                    AFL_WARN(NKikimrServices::TX_COLUMNSHARD)("event", "a lot of blobs to gc")("to_remove", extractedToRemoveFromDB.GetSize())("keeps_to_erase", keepsToErase.size())("limit", blobsGCCountLimit);
+                    break;
+                }
+            }
+            if (extractedToRemoveFromDB.GetSize() + keepsToErase.size() > blobsGCCountLimit) {
+                AFL_WARN(NKikimrServices::TX_COLUMNSHARD)("event", "a lot of blobs to gc")("to_remove", extractedToRemoveFromDB.GetSize())("keeps_to_erase", keepsToErase.size())("limit", blobsGCCountLimit);
+                break;
             }
             BlobsToKeep.erase(BlobsToKeep.begin(), keepBlobIt);
             BlobsManagerCounters.OnBlobsKeep(BlobsToKeep);
 
-            const auto predSelf = [&](const TUnifiedBlobId& id, const THashSet<TTabletId>& /*tabletIds*/) {
-                auto logoBlobId = id.GetLogoBlobId();
-                TGenStep genStep{logoBlobId.Generation(), logoBlobId.Step()};
-                return genStep <= newCollectGenStep && id.GetLogoBlobId().TabletID() == (ui64)SelfTabletId;
-            };
-            TTabletsByBlob extractedSelf = BlobsToDelete.ExtractBlobs(predSelf);
+            TTabletsByBlob extractedSelf;
+            {
+                while (deleteIndex.size()) {
+                    const auto& blobId = deleteIndex.front().GetLogoBlobId();
+                    if (newCollectGenStep < TGenStep(blobId.Generation(), blobId.Step())) {
+                        break;
+                    }
+                    BlobsToDelete.ExtractBlobTo(deleteIndex.front(), extractedSelf);
+                    if (extractedToRemoveFromDB.GetSize() + extractedSelf.GetSize() + keepsToErase.size() > blobsGCCountLimit) {
+                        AFL_WARN(NKikimrServices::TX_COLUMNSHARD)("event", "a lot of blobs to gc")("to_remove", extractedToRemoveFromDB.GetSize())("keeps_to_erase", keepsToErase.size())("limit", blobsGCCountLimit);
+                        break;
+                    }
+                    deleteIndex.pop_front();
+                }
+
+            }
+
             extractedToRemoveFromDB.Add(extractedSelf);
-            TTabletId tabletId;
-            TUnifiedBlobId unifiedBlobId;
-            while (extractedSelf.ExtractFront(tabletId, unifiedBlobId)) {
+            for (TTabletsByBlob::TIterator itExtractedSelf(extractedSelf); itExtractedSelf.IsValid(); ++itExtractedSelf) {
+                const TUnifiedBlobId unifiedBlobId = itExtractedSelf.GetBlobId();
                 auto logoBlobId = unifiedBlobId.GetLogoBlobId();
                 AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("to_delete_gc", logoBlobId);
                 NBlobOperations::NBlobStorage::TGCTask::TGCLists& gl = perGroupGCListsInFlight[unifiedBlobId.GetDsGroup()];
@@ -359,6 +380,8 @@ std::shared_ptr<NBlobOperations::NBlobStorage::TGCTask> TBlobManager::BuildGCTas
     return result;
 }
 
+
+
 TBlobBatch TBlobManager::StartBlobBatch(ui32 channel) {
     ++CountersUpdate.BatchesStarted;
     Y_ABORT_UNLESS(channel == BLOB_CHANNEL, "Support for mutiple blob channels is not implemented yet");
diff --git a/ydb/core/tx/columnshard/blobs_action/transaction/tx_gc_insert_table.cpp b/ydb/core/tx/columnshard/blobs_action/transaction/tx_gc_insert_table.cpp
@@ -30,7 +30,8 @@ void TTxInsertTableCleanup::Complete(const TActorContext& /*ctx*/) {
 
     Y_ABORT_UNLESS(BlobsAction);
     BlobsAction->OnCompleteTxAfterRemoving(true);
-    Self->EnqueueBackgroundActivities();
+    Self->BackgroundController.FinishCleanupInsertTable();
+    Self->SetupCleanupInsertTable();
 }
 
 }
diff --git a/ydb/core/tx/columnshard/columnshard_impl.cpp b/ydb/core/tx/columnshard/columnshard_impl.cpp
@@ -845,13 +845,18 @@ void TColumnShard::Handle(TEvPrivate::TEvGarbageCollectionFinished::TPtr& ev, co
 }
 
 void TColumnShard::SetupCleanupInsertTable() {
+    if (BackgroundController.IsCleanupInsertTableActive()) {
+        ACFL_DEBUG("background", "cleanup_insert_table")("skip_reason", "in_progress");
+        return;
+    }
     auto writeIdsToCleanup = InsertTable->OldWritesToAbort(AppData()->TimeProvider->Now());
 
     if (!InsertTable->GetAborted().size() && !writeIdsToCleanup.size()) {
         return;
     }
     AFL_INFO(NKikimrServices::TX_COLUMNSHARD)("event", "cleanup_started")("aborted", InsertTable->GetAborted().size())("to_cleanup", writeIdsToCleanup.size());
 
+    BackgroundController.StartCleanupInsertTable();
     Execute(new TTxInsertTableCleanup(this, std::move(writeIdsToCleanup)), TActorContext::AsActorContext());
 }
 
diff --git a/ydb/core/tx/columnshard/engines/column_engine.cpp b/ydb/core/tx/columnshard/engines/column_engine.cpp
@@ -9,15 +9,16 @@ const std::shared_ptr<arrow::Schema>& IColumnEngine::GetReplaceKey() const {
 }
 
 ui64 IColumnEngine::GetMetadataLimit() {
+    static const auto MemoryTotal = NSystemInfo::TotalMemorySize();
     if (!HasAppData()) {
-        AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("total", NSystemInfo::TotalMemorySize());
-        return NSystemInfo::TotalMemorySize() * 0.3;
+        AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("total", MemoryTotal);
+        return MemoryTotal * 0.3;
     } else if (AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().HasAbsoluteValue()) {
         AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("value", AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetAbsoluteValue());
         return AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetAbsoluteValue();
     } else {
-        AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("total", NSystemInfo::TotalMemorySize())("kff", AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetTotalRatio());
-        return NSystemInfo::TotalMemorySize() * AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetTotalRatio();
+        AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("total", MemoryTotal)("kff", AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetTotalRatio());
+        return MemoryTotal * AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetTotalRatio();
     }
 }
 

Original file line number	Diff line number	Diff line change
`@@ -30,7 +30,8 @@ void TTxInsertTableCleanup::Complete(const TActorContext& /ctx/) {`
`30`	`30`
`31`	`31`	`Y_ABORT_UNLESS(BlobsAction);`
`32`	`32`	`BlobsAction->OnCompleteTxAfterRemoving(true);`
`33`		`- Self->EnqueueBackgroundActivities();`
	`33`	`+ Self->BackgroundController.FinishCleanupInsertTable();`
	`34`	`+ Self->SetupCleanupInsertTable();`
`34`	`35`	`}`
`35`	`36`
`36`	`37`	`}`
Original file line number	Diff line number	Diff line change
`@@ -845,13 +845,18 @@ void TColumnShard::Handle(TEvPrivate::TEvGarbageCollectionFinished::TPtr& ev, co`
`845`	`845`	`}`
`846`	`846`
`847`	`847`	`void TColumnShard::SetupCleanupInsertTable() {`
	`848`	`+ if (BackgroundController.IsCleanupInsertTableActive()) {`
	`849`	`+ ACFL_DEBUG("background", "cleanup_insert_table")("skip_reason", "in_progress");`
	`850`	`+ return;`
	`851`	`+ }`
`848`	`852`	`auto writeIdsToCleanup = InsertTable->OldWritesToAbort(AppData()->TimeProvider->Now());`
`849`	`853`
`850`	`854`	`if (!InsertTable->GetAborted().size() && !writeIdsToCleanup.size()) {`
`851`	`855`	`return;`
`852`	`856`	`}`
`853`	`857`	`AFL_INFO(NKikimrServices::TX_COLUMNSHARD)("event", "cleanup_started")("aborted", InsertTable->GetAborted().size())("to_cleanup", writeIdsToCleanup.size());`
`854`	`858`
	`859`	`+ BackgroundController.StartCleanupInsertTable();`
`855`	`860`	`Execute(new TTxInsertTableCleanup(this, std::move(writeIdsToCleanup)), TActorContext::AsActorContext());`
`856`	`861`	`}`
`857`	`862`
Original file line number	Diff line number	Diff line change
`@@ -9,15 +9,16 @@ const std::shared_ptr<arrow::Schema>& IColumnEngine::GetReplaceKey() const {`
`9`	`9`	`}`
`10`	`10`
`11`	`11`	`ui64 IColumnEngine::GetMetadataLimit() {`
	`12`	`+ static const auto MemoryTotal = NSystemInfo::TotalMemorySize();`
`12`	`13`	`if (!HasAppData()) {`
`13`		`- AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("total", NSystemInfo::TotalMemorySize());`
`14`		`- return NSystemInfo::TotalMemorySize() * 0.3;`
	`14`	`+ AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("total", MemoryTotal);`
	`15`	`+ return MemoryTotal * 0.3;`
`15`	`16`	`} else if (AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().HasAbsoluteValue()) {`
`16`	`17`	`AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("value", AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetAbsoluteValue());`
`17`	`18`	`return AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetAbsoluteValue();`
`18`	`19`	`} else {`
`19`		`- AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("total", NSystemInfo::TotalMemorySize())("kff", AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetTotalRatio());`
`20`		`- return NSystemInfo::TotalMemorySize() * AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetTotalRatio();`
	`20`	`+ AFL_DEBUG(NKikimrServices::TX_COLUMNSHARD)("total", MemoryTotal)("kff", AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetTotalRatio());`
	`21`	`+ return MemoryTotal * AppDataVerified().ColumnShardConfig.GetIndexMetadataMemoryLimit().GetTotalRatio();`
`21`	`22`	`}`
`22`	`23`	`}`
`23`	`24`