Merge ea10ee3 into 63653ea

Author: CyberROFL

ydb/core/protos/counters_datashard.proto

@@ -490,4 +490,5 @@ enum ETxTypes {
     TXTYPE_CLEANUP_VOLATILE = 80                          [(TxTypeOpts) = {Name: "TxCleanupVolatile"}];
     TXTYPE_PLAN_PREDICTED_TXS = 81                        [(TxTypeOpts) = {Name: "TxPlanPredictedTxs"}];
     TXTYPE_WRITE = 82                                     [(TxTypeOpts) = {Name: "TxWrite"}];
+    TXTYPE_REMOVE_SCHEMA_SNAPSHOTS = 83                   [(TxTypeOpts) = {Name: "TxRemoveSchemaSnapshots"}];
 }

ydb/core/tx/datashard/datashard.cpp

@@ -855,6 +855,45 @@ void TDataShard::PersistChangeRecord(NIceDb::TNiceDb& db, const TChangeRecord& r
             NIceDb::TUpdate<Schema::ChangeRecordDetails::Kind>(record.GetKind()),
             NIceDb::TUpdate<Schema::ChangeRecordDetails::Body>(record.GetBody()),
             NIceDb::TUpdate<Schema::ChangeRecordDetails::Source>(record.GetSource()));
+
+        auto res = ChangesQueue.emplace(record.GetOrder(), record);
+        Y_VERIFY_S(res.second, "Duplicate change record: " << record.GetOrder());
+
+        if (res.first->second.SchemaVersion) {
+            res.first->second.SchemaSnapshotAcquired = SchemaSnapshotManager.AcquireReference(
+                TSchemaSnapshotKey(res.first->second.TableId, res.first->second.SchemaVersion));
+        }
+
+        const auto key = TCommittingChangeRecordsKey::FromRecord(record);
+        if (!CommittingChangeRecords.contains(key)) {
+            db.GetDatabase().OnCommit([this, key] {
+                auto it = CommittingChangeRecords.find(key);
+                Y_ABORT_UNLESS(it != CommittingChangeRecords.end());
+                CommittingChangeRecords.erase(it);
+            });
+            db.GetDatabase().OnRollback([this, key] {
+                auto it = CommittingChangeRecords.find(key);
+                Y_ABORT_UNLESS(it != CommittingChangeRecords.end());
+
+                for (const auto order : it->second) {
+                    auto cIt = ChangesQueue.find(order);
+                    Y_VERIFY_S(cIt != ChangesQueue.end(), "Cannot find change record: " << order);
+
+                    if (cIt->second.SchemaSnapshotAcquired) {
+                        const auto snapshotKey = TSchemaSnapshotKey(cIt->second.TableId, cIt->second.SchemaVersion);
+                        if (const auto last = SchemaSnapshotManager.ReleaseReference(snapshotKey)) {
+                            ScheduleRemoveSchemaSnapshot(snapshotKey);
+                        }
+                    }
+
+                    ChangesQueue.erase(cIt);
+                }
+
+                CommittingChangeRecords.erase(it);
+            });
+        }
+
+        CommittingChangeRecords[key].push_back(record.GetOrder());
     } else {
         auto& state = LockChangeRecords[lockId];
         Y_ABORT_UNLESS(state.Changes.empty() || state.Changes.back().LockOffset < record.GetLockOffset(),
@@ -934,6 +973,14 @@ void TDataShard::CommitLockChangeRecords(NIceDb::TNiceDb& db, ui64 lockId, ui64
         committed.Step = rowVersion.Step;
         committed.TxId = rowVersion.TxId;
         collected.push_back(committed);
+
+        auto res = ChangesQueue.emplace(committed.Order, committed);
+        Y_VERIFY_S(res.second, "Duplicate change record: " << committed.Order);
+
+        if (res.first->second.SchemaVersion) {
+            res.first->second.SchemaSnapshotAcquired = SchemaSnapshotManager.AcquireReference(
+                TSchemaSnapshotKey(res.first->second.TableId, res.first->second.SchemaVersion));
+        }
     }
 
     Y_VERIFY_S(!CommittedLockChangeRecords.contains(lockId), "Cannot commit lock " << lockId << " more than once");
@@ -960,7 +1007,26 @@ void TDataShard::CommitLockChangeRecords(NIceDb::TNiceDb& db, ui64 lockId, ui64
         LockChangeRecords.erase(it);
     });
     db.GetDatabase().OnRollback([this, lockId]() {
-        CommittedLockChangeRecords.erase(lockId);
+        auto it = CommittedLockChangeRecords.find(lockId);
+        Y_VERIFY_S(it != CommittedLockChangeRecords.end(), "Unexpected failure to find lockId# " << lockId);
+
+        for (size_t i = 0; i < it->second.Count; ++i) {
+            const ui64 order = it->second.Order + i;
+
+            auto cIt = ChangesQueue.find(order);
+            Y_VERIFY_S(cIt != ChangesQueue.end(), "Cannot find change record: " << order);
+
+            if (cIt->second.SchemaSnapshotAcquired) {
+                const auto snapshotKey = TSchemaSnapshotKey(cIt->second.TableId, cIt->second.SchemaVersion);
+                if (const auto last = SchemaSnapshotManager.ReleaseReference(snapshotKey)) {
+                    ScheduleRemoveSchemaSnapshot(snapshotKey);
+                }
+            }
+
+            ChangesQueue.erase(cIt);
+        }
+
+        CommittedLockChangeRecords.erase(it);
     });
 }
 
@@ -1022,23 +1088,9 @@ void TDataShard::RemoveChangeRecord(NIceDb::TNiceDb& db, ui64 order) {
     ChangesQueueBytes -= record.BodySize;
 
     if (record.SchemaSnapshotAcquired) {
-        Y_ABORT_UNLESS(record.TableId);
-        auto tableIt = TableInfos.find(record.TableId.LocalPathId);
-
-        if (tableIt != TableInfos.end()) {
-            const auto snapshotKey = TSchemaSnapshotKey(record.TableId, record.SchemaVersion);
-            const bool last = SchemaSnapshotManager.ReleaseReference(snapshotKey);
-
-            if (last) {
-                const auto* snapshot = SchemaSnapshotManager.FindSnapshot(snapshotKey);
-                Y_ABORT_UNLESS(snapshot);
-
-                if (snapshot->Schema->GetTableSchemaVersion() < tableIt->second->GetTableSchemaVersion()) {
-                    SchemaSnapshotManager.RemoveShapshot(db, snapshotKey);
-                }
-            }
-        } else {
-            Y_DEBUG_ABORT_UNLESS(State == TShardState::PreOffline);
+        const auto snapshotKey = TSchemaSnapshotKey(record.TableId, record.SchemaVersion);
+        if (const bool last = SchemaSnapshotManager.ReleaseReference(snapshotKey)) {
+            ScheduleRemoveSchemaSnapshot(snapshotKey);
         }
     }
 
@@ -1081,22 +1133,15 @@ void TDataShard::EnqueueChangeRecords(TVector<IDataShardChangeCollector::TChange
     for (const auto& record : records) {
         forward.emplace_back(record.Order, record.PathId, record.BodySize);
 
-        auto res = ChangesQueue.emplace(
-            std::piecewise_construct,
-            std::forward_as_tuple(record.Order),
-            std::forward_as_tuple(record, now, cookie)
-        );
-        if (res.second) {
-            ChangesList.PushBack(&res.first->second);
+        auto it = ChangesQueue.find(record.Order);
+        Y_ABORT_UNLESS(it != ChangesQueue.end());
 
-            Y_ABORT_UNLESS(ChangesQueueBytes <= (Max<ui64>() - record.BodySize));
-            ChangesQueueBytes += record.BodySize;
+        it->second.EnqueuedAt = now;
+        it->second.ReservationCookie = cookie;
+        ChangesList.PushBack(&it->second);
 
-            if (record.SchemaVersion) {
-                res.first->second.SchemaSnapshotAcquired = SchemaSnapshotManager.AcquireReference(
-                    TSchemaSnapshotKey(record.TableId, record.SchemaVersion));
-            }
-        }
+        Y_ABORT_UNLESS(ChangesQueueBytes <= (Max<ui64>() - record.BodySize));
+        ChangesQueueBytes += record.BodySize;
     }
 
     if (auto it = ChangeQueueReservations.find(cookie); it != ChangeQueueReservations.end()) {
@@ -1265,6 +1310,14 @@ bool TDataShard::LoadChangeRecords(NIceDb::TNiceDb& db, TVector<IDataShardChange
             .SchemaVersion = schemaVersion,
         });
 
+        auto res = ChangesQueue.emplace(records.back().Order, records.back());
+        Y_VERIFY_S(res.second, "Duplicate change record: " << records.back().Order);
+
+        if (res.first->second.SchemaVersion) {
+            res.first->second.SchemaSnapshotAcquired = SchemaSnapshotManager.AcquireReference(
+                TSchemaSnapshotKey(res.first->second.TableId, res.first->second.SchemaVersion));
+        }
+
         if (!rowset.Next()) {
             return false;
         }
@@ -1363,6 +1416,14 @@ bool TDataShard::LoadChangeRecordCommits(NIceDb::TNiceDb& db, TVector<IDataShard
             });
             entry.Count++;
             needSort = true;
+
+            auto res = ChangesQueue.emplace(records.back().Order, records.back());
+            Y_VERIFY_S(res.second, "Duplicate change record: " << records.back().Order);
+
+            if (res.first->second.SchemaVersion) {
+                res.first->second.SchemaSnapshotAcquired = SchemaSnapshotManager.AcquireReference(
+                    TSchemaSnapshotKey(res.first->second.TableId, res.first->second.SchemaVersion));
+            }
         }
 
         LockChangeRecords.erase(lockId);
@@ -1421,6 +1482,51 @@ void TDataShard::ScheduleRemoveAbandonedLockChanges() {
     }
 }
 
+void TDataShard::ScheduleRemoveSchemaSnapshot(const TSchemaSnapshotKey& key) {
+    Y_ABORT_UNLESS(!SchemaSnapshotManager.HasReference(key));
+
+    const auto* snapshot = SchemaSnapshotManager.FindSnapshot(key);
+    Y_ABORT_UNLESS(snapshot);
+
+    auto it = TableInfos.find(key.PathId);
+    if (it == TableInfos.end()) {
+        Y_DEBUG_ABORT_UNLESS(State == TShardState::PreOffline);
+        return;
+    }
+
+    if (snapshot->Schema->GetTableSchemaVersion() < it->second->GetTableSchemaVersion()) {
+        bool wasEmpty = PendingSchemaSnapshotsToRemove.empty();
+        PendingSchemaSnapshotsToRemove.push_back(key);
+        if (wasEmpty) {
+            Send(SelfId(), new TEvPrivate::TEvRemoveSchemaSnapshots);
+        }
+    }
+}
+
+void TDataShard::ScheduleRemoveAbandonedSchemaSnapshots() {
+    bool wasEmpty = PendingSchemaSnapshotsToRemove.empty();
+
+    for (const auto& [key, snapshot] : SchemaSnapshotManager.GetSnapshots()) {
+        auto it = TableInfos.find(key.PathId);
+        if (it == TableInfos.end()) {
+            Y_DEBUG_ABORT_UNLESS(State == TShardState::PreOffline);
+            break;
+        }
+        if (SchemaSnapshotManager.HasReference(key)) {
+            continue;
+        }
+        if (snapshot.Schema->GetTableSchemaVersion() >= it->second->GetTableSchemaVersion()) {
+            continue;
+        }
+
+        PendingSchemaSnapshotsToRemove.push_back(key);
+    }
+
+    if (wasEmpty && !PendingSchemaSnapshotsToRemove.empty()) {
+        Send(SelfId(), new TEvPrivate::TEvRemoveSchemaSnapshots);
+    }
+}
+
 void TDataShard::PersistSchemeTxResult(NIceDb::TNiceDb &db, const TSchemaOperation &op) {
     db.Table<Schema::SchemaOperations>().Key(op.TxId).Update(
         NIceDb::TUpdate<Schema::SchemaOperations::Success>(op.Success),
@@ -1649,8 +1755,18 @@ void TDataShard::AddSchemaSnapshot(const TPathId& pathId, ui64 tableSchemaVersio
     Y_ABORT_UNLESS(TableInfos.contains(pathId.LocalPathId));
     auto tableInfo = TableInfos[pathId.LocalPathId];
 
-    const auto key = TSchemaSnapshotKey(pathId.OwnerId, pathId.LocalPathId, tableSchemaVersion);
+    const auto key = TSchemaSnapshotKey(pathId, tableSchemaVersion);
     SchemaSnapshotManager.AddSnapshot(txc.DB, key, TSchemaSnapshot(tableInfo, step, txId));
+
+    const auto& snapshots = SchemaSnapshotManager.GetSnapshots();
+    for (auto it = snapshots.lower_bound(TSchemaSnapshotKey(pathId, 1)); it != snapshots.end(); ++it) {
+        if (it->first == key) {
+            break;
+        }
+        if (!SchemaSnapshotManager.HasReference(it->first)) {
+            ScheduleRemoveSchemaSnapshot(it->first);
+        }
+    }
 }
 
 void TDataShard::PersistLastLoanTableTid(NIceDb::TNiceDb& db, ui32 localTid) {

ydb/core/tx/datashard/datashard__init.cpp

@@ -425,6 +425,12 @@ bool TDataShard::TTxInit::ReadEverything(TTransactionContext &txc) {
             return false;
     }
 
+    if (Self->State != TShardState::Offline && txc.DB.GetScheme().GetTableInfo(Schema::SchemaSnapshots::TableId)) {
+        if (!Self->SchemaSnapshotManager.Load(db)) {
+            return false;
+        }
+    }
+
     if (Self->State != TShardState::Offline && txc.DB.GetScheme().GetTableInfo(Schema::ChangeRecords::TableId)) {
         if (!Self->LoadChangeRecords(db, ChangeRecords)) {
             return false;
@@ -512,12 +518,6 @@ bool TDataShard::TTxInit::ReadEverything(TTransactionContext &txc) {
         }
     }
 
-    if (Self->State != TShardState::Offline && txc.DB.GetScheme().GetTableInfo(Schema::SchemaSnapshots::TableId)) {
-        if (!Self->SchemaSnapshotManager.Load(db)) {
-            return false;
-        }
-    }
-
     if (Self->State != TShardState::Offline && txc.DB.GetScheme().GetTableInfo(Schema::Locks::TableId)) {
         TDataShardLocksDb locksDb(*Self, txc);
         if (!Self->SysLocks.Load(locksDb)) {
@@ -547,6 +547,7 @@ bool TDataShard::TTxInit::ReadEverything(TTransactionContext &txc) {
     Self->SubscribeNewLocks();
 
     Self->ScheduleRemoveAbandonedLockChanges();
+    Self->ScheduleRemoveAbandonedSchemaSnapshots();
 
     return true;
 }

ydb/core/tx/datashard/datashard_change_sending.cpp

@@ -340,15 +340,13 @@ class TDataShard::TTxRemoveChangeRecords: public TTransactionBase<TDataShard> {
             Self->RemoveChangeRecordsInFly = false;
         }
 
-        if (!Self->ChangesQueue) { // double check queue
-            if (ChangeExchangeSplit) {
-                Self->KillChangeSender(ctx);
-                Self->ChangeExchangeSplitter.DoSplit(ctx);
-            }
+        if (ChangeExchangeSplit) {
+            Self->KillChangeSender(ctx);
+            Self->ChangeExchangeSplitter.DoSplit(ctx);
+        }
 
-            for (const auto dstTabletId : ActivationList) {
-                Self->ChangeSenderActivator.DoSend(dstTabletId, ctx);
-            }
+        for (const auto dstTabletId : ActivationList) {
+            Self->ChangeSenderActivator.DoSend(dstTabletId, ctx);
         }
 
         Self->CheckStateChange(ctx);