Compact borrowed mvcc garbage after split/merge, KIKIMR-14979

ref:58c92a54c8a558f4781de7663cdd048fe0fae7ee

9 files changed, 60 insertions, 1 deletions

diff --git a/ydb/core/tablet_flat/flat_comp.h b/ydb/core/tablet_flat/flat_comp.h
index a768863386e..e010255a59b 100644
--- a/ydb/core/tablet_flat/flat_comp.h
+++ b/ydb/core/tablet_flat/flat_comp.h
@@ -312,6 +312,14 @@ namespace NTable {
             return false;
         }
 
+        /**
+         * Called to signal it's ok to compact borrowed garbage
+         * even if there's no private tablet data
+         */
+        virtual void AllowBorrowedGarbageCompaction() {
+            // nothing
+        }
+
         virtual ui64 GetLastFinishedForcedCompactionId() const = 0;
 
         virtual TInstant GetLastFinishedForcedCompactionTs() const = 0;
diff --git a/ydb/core/tablet_flat/flat_comp_gen.cpp b/ydb/core/tablet_flat/flat_comp_gen.cpp
index 16ec271bd16..63aeca89caa 100644
--- a/ydb/core/tablet_flat/flat_comp_gen.cpp
+++ b/ydb/core/tablet_flat/flat_comp_gen.cpp
@@ -1412,7 +1412,10 @@ bool TGenCompactionStrategy::MaybeAutoStartForceCompaction() {
     // belongs to our own tablet, i.e. not borrowed. This is so we don't
     // compact borrowed data mid-merge, which might cause epochs to become
     // out of sync across generations.
-    if (Generations && ForcedState == EForcedState::None && StatsPerTablet.contains(Backend->OwnerTabletId())) {
+    if (Generations &&
+        ForcedState == EForcedState::None &&
+        (CompactBorrowedGarbageAllowed || StatsPerTablet.contains(Backend->OwnerTabletId())))
+    {
         bool startForcedCompaction = NeedToStartForceCompaction();
         if (startForcedCompaction) {
             ForcedState = EForcedState::Pending;
@@ -1424,6 +1427,13 @@ bool TGenCompactionStrategy::MaybeAutoStartForceCompaction() {
     return false;
 }
 
+void TGenCompactionStrategy::AllowBorrowedGarbageCompaction() {
+    if (!CompactBorrowedGarbageAllowed) {
+        CompactBorrowedGarbageAllowed = true;
+        ReflectRemovedRowVersions();
+    }
+}
+
 }
 }
 }
diff --git a/ydb/core/tablet_flat/flat_comp_gen.h b/ydb/core/tablet_flat/flat_comp_gen.h
index 07b39415492..3ba12cd623b 100644
--- a/ydb/core/tablet_flat/flat_comp_gen.h
+++ b/ydb/core/tablet_flat/flat_comp_gen.h
@@ -41,6 +41,7 @@ namespace NCompGen {
         ui64 GetBackingSize(ui64 ownerTabletId) override;
         ui64 BeginMemCompaction(TTaskId taskId, TSnapEdge edge, ui64 forcedCompactionId) override;
         bool ScheduleBorrowedCompaction() override;
+        void AllowBorrowedGarbageCompaction() override;
         ui64 GetLastFinishedForcedCompactionId() const override { return FinishedForcedGenCompactionId; }
         TInstant GetLastFinishedForcedCompactionTs() const override { return FinishedForcedGenCompactionTs; }
         TCompactionChanges CompactionFinished(
@@ -279,6 +280,8 @@ namespace NCompGen {
         TStats Stats;
         THashMap<ui64, TStats> StatsPerTablet;
         NPage::TGarbageStatsAgg GarbageStatsAgg;
+
+        bool CompactBorrowedGarbageAllowed = false;
     };
 
 }
diff --git a/ydb/core/tablet_flat/flat_executor.cpp b/ydb/core/tablet_flat/flat_executor.cpp
index 4b8fc706e76..76ef090dd1a 100644
--- a/ydb/core/tablet_flat/flat_executor.cpp
+++ b/ydb/core/tablet_flat/flat_executor.cpp
@@ -3708,6 +3708,12 @@ bool TExecutor::CompactTables() {
     }
 }
 
+void TExecutor::AllowBorrowedGarbageCompaction(ui32 tableId) {
+    if (CompactionLogic) {
+        return CompactionLogic->AllowBorrowedGarbageCompaction(tableId);
+    }
+}
+
 STFUNC(TExecutor::StateInit) {
     Y_UNUSED(ev);
     Y_UNUSED(ctx);
diff --git a/ydb/core/tablet_flat/flat_executor.h b/ydb/core/tablet_flat/flat_executor.h
index 5fd0e391d98..b9c6753bf03 100644
--- a/ydb/core/tablet_flat/flat_executor.h
+++ b/ydb/core/tablet_flat/flat_executor.h
@@ -628,6 +628,8 @@ public:
     ui64 CompactTable(ui32 tableId) override;
     bool CompactTables() override;
 
+    void AllowBorrowedGarbageCompaction(ui32 tableId) override;
+
     void FollowerAttached() override;
     void FollowerSyncComplete() override;
     void FollowerGcApplied(ui32 step, TDuration followerSyncDelay) override;
diff --git a/ydb/core/tablet_flat/flat_executor_compaction_logic.cpp b/ydb/core/tablet_flat/flat_executor_compaction_logic.cpp
index b28c5214496..514aea3725a 100644
--- a/ydb/core/tablet_flat/flat_executor_compaction_logic.cpp
+++ b/ydb/core/tablet_flat/flat_executor_compaction_logic.cpp
@@ -201,6 +201,13 @@ TFinishedCompactionInfo TCompactionLogic::GetFinishedCompactionInfo(ui32 table)
         tableInfo->Strategy->GetLastFinishedForcedCompactionTs());
 }
 
+void TCompactionLogic::AllowBorrowedGarbageCompaction(ui32 table) {
+    auto *tableInfo = State->Tables.FindPtr(table);
+    Y_VERIFY(tableInfo && tableInfo->Strategy, "Cannot AllowBorrowedGarbageCompaction for unexpected table %" PRIu32, table);
+    tableInfo->AllowBorrowedGarbageCompaction = true;
+    tableInfo->Strategy->AllowBorrowedGarbageCompaction();
+}
+
 TReflectSchemeChangesResult TCompactionLogic::ReflectSchemeChanges()
 {
     TReflectSchemeChangesResult result;
@@ -248,6 +255,10 @@ TReflectSchemeChangesResult TCompactionLogic::ReflectSchemeChanges()
             } else {
                 table.Strategy->Start({ });
             }
+
+            if (table.AllowBorrowedGarbageCompaction) {
+                table.Strategy->AllowBorrowedGarbageCompaction();
+            }
         } else {
             Y_VERIFY(table.Strategy);
             table.Strategy->ReflectSchema();
diff --git a/ydb/core/tablet_flat/flat_executor_compaction_logic.h b/ydb/core/tablet_flat/flat_executor_compaction_logic.h
index 7993fdb83c0..be7a3cda8d4 100644
--- a/ydb/core/tablet_flat/flat_executor_compaction_logic.h
+++ b/ydb/core/tablet_flat/flat_executor_compaction_logic.h
@@ -100,6 +100,8 @@ struct TCompactionLogicState {
 
         bool ChangesRequested = false;
 
+        bool AllowBorrowedGarbageCompaction = true;
+
         TTableInfo() = default;
 
         ~TTableInfo();
@@ -217,6 +219,8 @@ public:
 
     TFinishedCompactionInfo GetFinishedCompactionInfo(ui32 table);
 
+    void AllowBorrowedGarbageCompaction(ui32 table);
+
     TReflectSchemeChangesResult ReflectSchemeChanges();
     void ReflectRemovedRowVersions(ui32 table);
     void UpdateInMemStatsStep(ui32 table, ui32 steps, ui64 size);
diff --git a/ydb/core/tablet_flat/tablet_flat_executor.h b/ydb/core/tablet_flat/tablet_flat_executor.h
index 8787bdb066f..efbae96a110 100644
--- a/ydb/core/tablet_flat/tablet_flat_executor.h
+++ b/ydb/core/tablet_flat/tablet_flat_executor.h
@@ -528,6 +528,12 @@ namespace NFlatExecutorSetup {
         virtual ui64 CompactTable(ui32 tableId) = 0;
         virtual bool CompactTables() = 0;
 
+        // Signal executor that it's ok to compact borrowed data in the given
+        // table even if there's no local modifications. Useful after finishing
+        // snapshot transfer on datashard split/merge so any mvcc data and/or
+        // erases can be compacted normally.
+        virtual void AllowBorrowedGarbageCompaction(ui32 tableId) = 0;
+
         virtual void RenderHtmlPage(NMon::TEvRemoteHttpInfo::TPtr&) const = 0;
         virtual void RenderHtmlCounters(NMon::TEvRemoteHttpInfo::TPtr&) const = 0;
         virtual void RenderHtmlDb(NMon::TEvRemoteHttpInfo::TPtr &ev, const TActorContext &ctx) const = 0;
diff --git a/ydb/core/tx/datashard/datashard.cpp b/ydb/core/tx/datashard/datashard.cpp
index e7e5deb690c..f75f4b3585e 100644
--- a/ydb/core/tx/datashard/datashard.cpp
+++ b/ydb/core/tx/datashard/datashard.cpp
@@ -380,6 +380,15 @@ void TDataShard::SwitchToWork(const TActorContext &ctx) {
     LOG_INFO_S(ctx, NKikimrServices::TX_DATASHARD, "Switched to work state "
          << DatashardStateName(State) << " tabletId " << TabletID());
 
+    if (State == TShardState::Ready && DstSplitDescription) {
+        // This shard was created as a result of split/merge (and not e.g. copy table)
+        // Signal executor that it should compact borrowed garbage even if this
+        // shard has no private data.
+        for (const auto& pr : TableInfos) {
+            Executor()->AllowBorrowedGarbageCompaction(pr.second->LocalTid);
+        }
+    }
+
     // Cleanup any removed snapshots from the previous generation
     Execute(new TTxCleanupRemovedSnapshots(this), ctx);