schemeshard: preserialize Table.SplitBoundary for describe result (ydb-platform#6648)

Preserialize table's split boundaries the same way table partitions are. The size of both depend on the same variable: number of shards in the table, but TablePartitions was preserialized (and cached) while Table.SplitBoundaries wasn't. Preserializing all potentially huge parts of DescribeSchemeResult message before it gets to the interconnect saves interconnect actors additional serialization costs. And when partitioning of the huge tables goes through the period of a rapid change that additional serialization causes interconnect to overload. Single shortcoming though: preserialized SplitBoundary is not used (cannot be used) when boundaries of the index tables are requested through describe request on table index. KIKIMR-21686

Author: ijon

ydb/core/tx/schemeshard/schemeshard__operation_alter_continuous_backup.cpp

@@ -83,7 +83,7 @@ TVector<ISubOperation::TPtr> CreateAlterContinuousBackup(TOperationId opId, cons
  const NScheme::TTypeRegistry* typeRegistry = AppData(context.Ctx)->TypeRegistry;
 
  NKikimrSchemeOp::TTableDescription schema;
- context.SS->DescribeTable(table, typeRegistry, true, false, &schema);
+ context.SS->DescribeTable(table, typeRegistry, true, &schema);
  schema.MutablePartitionConfig()->CopyFrom(table->TableDescription.GetPartitionConfig());
 
  TString errStr;

ydb/core/tx/schemeshard/schemeshard__operation_copy_table.cpp

@@ -14,7 +14,7 @@ void PrepareScheme(NKikimrSchemeOp::TTableDescription* schema, const TString& na
  const NScheme::TTypeRegistry* typeRegistry = AppData(context.Ctx)->TypeRegistry;
 
  NKikimrSchemeOp::TTableDescription completedSchema;
- context.SS->DescribeTable(srcTableInfo, typeRegistry, true, false, &completedSchema);
+ context.SS->DescribeTable(srcTableInfo, typeRegistry, true, &completedSchema);
  completedSchema.SetName(name);
 
  //inherit all from Src except PartitionConfig, PartitionConfig could be altered

ydb/core/tx/schemeshard/schemeshard_impl.cpp

@@ -598,8 +598,9 @@ void TSchemeShard::ClearDescribePathCaches(const TPathElement::TPtr node, bool f
  } else if (node->PathType == NKikimrSchemeOp::EPathType::EPathTypeTable) {
  Y_ABORT_UNLESS(Tables.contains(node->PathId));
  TTableInfo::TPtr tabletInfo = Tables.at(node->PathId);
- tabletInfo->PreSerializedPathDescription.clear();
- tabletInfo->PreSerializedPathDescriptionWithoutRangeKey.clear();
+ tabletInfo->PreserializedTablePartitions.clear();
+ tabletInfo->PreserializedTablePartitionsNoKeys.clear();
+ tabletInfo->PreserializedTableSplitBoundaries.clear();
  }
 }
 

ydb/core/tx/schemeshard/schemeshard_impl.h

@@ -1015,7 +1015,7 @@ class TSchemeShard
  void FillAsyncIndexInfo(const TPathId& tableId, NKikimrTxDataShard::TFlatSchemeTransaction& tx);
 
  void DescribeTable(const TTableInfo::TPtr tableInfo, const NScheme::TTypeRegistry* typeRegistry,
- bool fillConfig, bool fillBoundaries, NKikimrSchemeOp::TTableDescription* entry) const;
+ bool fillConfig, NKikimrSchemeOp::TTableDescription* entry) const;
  void DescribeTableIndex(const TPathId& pathId, const TString& name,
  bool fillConfig, bool fillBoundaries, NKikimrSchemeOp::TIndexDescription& entry
  ) const;
@@ -1031,7 +1031,6 @@ class TSchemeShard
  void DescribeReplication(const TPathId& pathId, const TString& name, NKikimrSchemeOp::TReplicationDescription& desc);
  void DescribeReplication(const TPathId& pathId, const TString& name, TReplicationInfo::TPtr info, NKikimrSchemeOp::TReplicationDescription& desc);
  void DescribeBlobDepot(const TPathId& pathId, const TString& name, NKikimrSchemeOp::TBlobDepotDescription& desc);
- static void FillTableBoundaries(const TTableInfo::TPtr tableInfo, google::protobuf::RepeatedPtrField<NKikimrSchemeOp::TSplitBoundary>& boundaries);
 
  void Handle(NKikimr::NOlap::NBackground::TEvExecuteGeneralLocalTransaction::TPtr& ev, const TActorContext& ctx);
  void Handle(NKikimr::NOlap::NBackground::TEvRemoveSession::TPtr& ev, const TActorContext& ctx);

ydb/core/tx/schemeshard/schemeshard_info_types.cpp

@@ -362,10 +362,10 @@ TTableInfo::TAlterDataPtr TTableInfo::CreateAlterData(
  const TTableInfo::TColumn& sourceColumn = source->Columns[colId];
 
  if (col.HasDefaultFromSequence()) {
- if (sourceColumn.PType.GetTypeId() != NScheme::NTypeIds::Int64 
+ if (sourceColumn.PType.GetTypeId() != NScheme::NTypeIds::Int64
  && NPg::PgTypeIdFromTypeDesc(sourceColumn.PType.GetTypeDesc()) != INT8OID) {
- TString sequenceType = sourceColumn.PType.GetTypeId() == NScheme::NTypeIds::Pg 
- ? NPg::PgTypeNameFromTypeDesc(NPg::TypeDescFromPgTypeId(INT8OID)) 
+ TString sequenceType = sourceColumn.PType.GetTypeId() == NScheme::NTypeIds::Pg
+ ? NPg::PgTypeNameFromTypeDesc(NPg::TypeDescFromPgTypeId(INT8OID))
  : NScheme::TypeName(NScheme::NTypeIds::Int64);
  errStr = Sprintf(
  "Sequence value type '%s' must be equal to the column type '%s'", sequenceType.c_str(),
@@ -423,7 +423,7 @@ TTableInfo::TAlterDataPtr TTableInfo::CreateAlterData(
  return nullptr;
  default:
  break;
- } 
+ }
  }
  } else {
  auto* typeDesc = NPg::TypeDescFromPgTypeName(typeName);
@@ -1586,8 +1586,9 @@ void TTableInfo::SetPartitioning(TVector<TTableShardInfo>&& newPartitioning) {
  Stats.PartitionStats.swap(newPartitionStats);
  Stats.Aggregated = newAggregatedStats;
  Partitions.swap(newPartitioning);
- PreSerializedPathDescription.clear();
- PreSerializedPathDescriptionWithoutRangeKey.clear();
+ PreserializedTablePartitions.clear();
+ PreserializedTablePartitionsNoKeys.clear();
+ PreserializedTableSplitBoundaries.clear();
 
  CondEraseSchedule.clear();
  InFlightCondErase.clear();

ydb/core/tx/schemeshard/schemeshard_info_types.h

@@ -436,8 +436,11 @@ struct TTableInfo : public TSimpleRefCount<TTableInfo> {
  TMap<TTxId, TBackupRestoreResult> BackupHistory;
  TMap<TTxId, TBackupRestoreResult> RestoreHistory;
 
- TString PreSerializedPathDescription;
- TString PreSerializedPathDescriptionWithoutRangeKey;
+ // Preserialized TDescribeSchemeResult with PathDescription.TablePartitions field filled
+ TString PreserializedTablePartitions;
+ TString PreserializedTablePartitionsNoKeys;
+ // Preserialized TDescribeSchemeResult with PathDescription.Table.SplitBoundary field filled
+ TString PreserializedTableSplitBoundaries;
 
  THashMap<TShardIdx, NKikimrSchemeOp::TPartitionConfig> PerShardPartitionConfig;
 

ydb/core/tx/schemeshard/schemeshard_path_describer.cpp

@@ -223,6 +223,68 @@ void TPathDescriber::DescribeDir(const TPath& path) {
  DescribeChildren(path);
 }
 
+void FillTableBoundaries(
+ google::protobuf::RepeatedPtrField<NKikimrSchemeOp::TSplitBoundary>* result,
+ const TTableInfo::TPtr tableInfo
+) {
+ TString errStr;
+ // Number of split boundaries equals to number of partitions - 1
+ result->Reserve(tableInfo->GetPartitions().size() - 1);
+ for (ui32 pi = 0; pi < tableInfo->GetPartitions().size() - 1; ++pi) {
+ const auto& p = tableInfo->GetPartitions()[pi];
+ TSerializedCellVec endKey(p.EndOfRange);
+ auto boundary = result->Add()->MutableKeyPrefix();
+ for (ui32 ki = 0; ki < endKey.GetCells().size(); ++ki){
+ const auto& c = endKey.GetCells()[ki];
+ auto type = tableInfo->Columns[tableInfo->KeyColumnIds[ki]].PType;
+ bool ok = NMiniKQL::CellToValue(type, c, *boundary->AddTuple(), errStr);
+ Y_ABORT_UNLESS(ok, "Failed to build key tuple at position %" PRIu32 " error: %s", ki, errStr.data());
+ }
+ }
+}
+
+void FillTablePartitions(
+ google::protobuf::RepeatedPtrField<NKikimrSchemeOp::TTablePartition>* result,
+ const TTableInfo::TPtr tableInfo,
+ const THashMap<TShardIdx, TShardInfo>& shardInfos,
+ bool includeKeys
+) {
+ result->Reserve(tableInfo->GetPartitions().size());
+ for (auto& p : tableInfo->GetPartitions()) {
+ const auto& tabletId = ui64(shardInfos.at(p.ShardIdx).TabletID);
+ const auto& key = p.EndOfRange;
+
+ auto part = result->Add();
+ part->SetDatashardId(tabletId);
+ if (includeKeys) {
+ // Currently we only support uniform partitioning where each range is [start, end)
+ // +inf as the end of the last range is represented by empty TCell vector
+ part->SetIsPoint(false);
+ part->SetIsInclusive(false);
+ part->SetEndOfRangeKeyPrefix(key);
+ }
+ }
+}
+
+const TString& GetSerializedTablePartitions(
+ const TTableInfo::TPtr tableInfo,
+ const THashMap<TShardIdx, TShardInfo>& shardInfos,
+ bool returnRangeKey
+) {
+ TString& cache = (returnRangeKey
+ ? tableInfo->PreserializedTablePartitions
+ : tableInfo->PreserializedTablePartitionsNoKeys
+ );
+
+ if (cache.empty()) {
+ NKikimrScheme::TEvDescribeSchemeResult result;
+ FillTablePartitions(result.MutablePathDescription()->MutableTablePartitions(), tableInfo, shardInfos, returnRangeKey);
+ Y_PROTOBUF_SUPPRESS_NODISCARD result.SerializeToString(&cache);
+ }
+
+ return cache;
+}
+
 void TPathDescriber::DescribeTable(const TActorContext& ctx, TPathId pathId, TPathElement::TPtr pathEl) {
  const NScheme::TTypeRegistry* typeRegistry = AppData(ctx)->TypeRegistry;
  const TTableInfo::TPtr tableInfo = *Self->Tables.FindPtr(pathId);
@@ -244,50 +306,30 @@ void TPathDescriber::DescribeTable(const TActorContext& ctx, TPathId pathId, TPa
  returnRangeKey = Params.GetOptions().GetReturnRangeKey();
  }
 
- Self->DescribeTable(tableInfo, typeRegistry, returnConfig, returnBoundaries, entry);
+ Self->DescribeTable(tableInfo, typeRegistry, returnConfig, entry);
  entry->SetName(pathEl->Name);
 
- if (returnPartitioning) {
- // partitions
- if (tableInfo->PreSerializedPathDescription.empty()) {
+ if (returnBoundaries) {
+ // split boundaries (split keys without shard's tablet-ids)
+ if (tableInfo->PreserializedTableSplitBoundaries.empty()) {
  NKikimrScheme::TEvDescribeSchemeResult preSerializedResult;
- NKikimrScheme::TEvDescribeSchemeResult preSerializedResultWithoutRangeKey;
-
- NKikimrSchemeOp::TPathDescription& pathDescription = *preSerializedResult.MutablePathDescription();
- NKikimrSchemeOp::TPathDescription& pathDescriptionWithoutRangeKey = *preSerializedResultWithoutRangeKey.MutablePathDescription();
-
- pathDescription.MutableTablePartitions()->Reserve(tableInfo->GetPartitions().size());
- pathDescriptionWithoutRangeKey.MutableTablePartitions()->Reserve(tableInfo->GetPartitions().size());
- for (auto& p : tableInfo->GetPartitions()) {
- auto part = pathDescription.AddTablePartitions();
- auto partWithoutRangeKey = pathDescriptionWithoutRangeKey.AddTablePartitions();
- auto datashardIdx = p.ShardIdx;
- auto datashardTabletId = Self->ShardInfos[datashardIdx].TabletID;
- // Currently we only support uniform partitioning where each range is [start, end)
- // +inf as the end of the last range is represented by empty TCell vector
- part->SetDatashardId(ui64(datashardTabletId));
- partWithoutRangeKey->SetDatashardId(ui64(datashardTabletId));
-
- part->SetIsPoint(false);
- partWithoutRangeKey->SetIsPoint(false);
-
- part->SetIsInclusive(false);
- partWithoutRangeKey->SetIsInclusive(false);
-
- part->SetEndOfRangeKeyPrefix(p.EndOfRange);
- }
- Y_PROTOBUF_SUPPRESS_NODISCARD preSerializedResult.SerializeToString(&tableInfo->PreSerializedPathDescription);
- Y_PROTOBUF_SUPPRESS_NODISCARD preSerializedResultWithoutRangeKey.SerializeToString(&tableInfo->PreSerializedPathDescriptionWithoutRangeKey);
- }
- if (returnRangeKey) {
- Result->PreSerializedData += tableInfo->PreSerializedPathDescription;
- } else {
- Result->PreSerializedData += tableInfo->PreSerializedPathDescriptionWithoutRangeKey;
- }
- if (!pathEl->IsCreateFinished()) {
- tableInfo->PreSerializedPathDescription.clear(); // KIKIMR-4337
- tableInfo->PreSerializedPathDescriptionWithoutRangeKey.clear();
+ auto& tableDesc = *preSerializedResult.MutablePathDescription()->MutableTable();
+ FillTableBoundaries(tableDesc.MutableSplitBoundary(), tableInfo);
+ Y_PROTOBUF_SUPPRESS_NODISCARD preSerializedResult.SerializeToString(&tableInfo->PreserializedTableSplitBoundaries);
  }
+ Result->PreSerializedData += tableInfo->PreserializedTableSplitBoundaries;
+ }
+
+ if (returnPartitioning) {
+ // partitions (shard tablet-ids with range keys)
+ Result->PreSerializedData += GetSerializedTablePartitions(tableInfo, Self->ShardInfos, returnRangeKey);
+ }
+
+ // KIKIMR-4337: table info is in flux until table is finally created
+ if (!pathEl->IsCreateFinished()) {
+ tableInfo->PreserializedTablePartitions.clear();
+ tableInfo->PreserializedTablePartitionsNoKeys.clear();
+ tableInfo->PreserializedTableSplitBoundaries.clear();
  }
 
  FillAggregatedStats(*Result->Record.MutablePathDescription(), tableInfo->GetStats());
@@ -1128,8 +1170,12 @@ THolder<TEvSchemeShard::TEvDescribeSchemeResultBuilder> DescribePath(
  return DescribePath(self, ctx, pathId, options);
 }
 
-void TSchemeShard::DescribeTable(const TTableInfo::TPtr tableInfo, const NScheme::TTypeRegistry* typeRegistry,
- bool fillConfig, bool fillBoundaries, NKikimrSchemeOp::TTableDescription* entry) const
+void TSchemeShard::DescribeTable(
+ const TTableInfo::TPtr tableInfo,
+ const NScheme::TTypeRegistry* typeRegistry,
+ bool fillConfig,
+ NKikimrSchemeOp::TTableDescription* entry
+ ) const
 {
  Y_UNUSED(typeRegistry);
  THashMap<ui32, TString> familyNames;
@@ -1198,10 +1244,6 @@ void TSchemeShard::DescribeTable(const TTableInfo::TPtr tableInfo, const NScheme
  FillPartitionConfig(tableInfo->PartitionConfig(), *entry->MutablePartitionConfig());
  }
 
- if (fillBoundaries) {
- FillTableBoundaries(tableInfo, *entry->MutableSplitBoundary());
- }
-
  if (tableInfo->HasTTLSettings()) {
  entry->MutableTTLSettings()->CopyFrom(tableInfo->TTLSettings());
  }
@@ -1244,32 +1286,32 @@ void TSchemeShard::DescribeTableIndex(const TPathId& pathId, const TString& name
  *entry.MutableDataColumnNames()->Add() = dataColumns;
  }
 
- auto* indexPath = PathsById.FindPtr(pathId);
+ auto indexPath = *PathsById.FindPtr(pathId);
  Y_ABORT_UNLESS(indexPath);
  const ui8 expectedIndexImplTableCount = indexInfo->Type == NKikimrSchemeOp::EIndexType::EIndexTypeGlobalVectorKmeansTree ? 2 : 1;
- Y_ABORT_UNLESS((*indexPath)->GetChildren().size() == expectedIndexImplTableCount);
+ Y_ABORT_UNLESS(indexPath->GetChildren().size() == expectedIndexImplTableCount);
 
  ui64 dataSize = 0;
- for (const auto& indexImplTablePathId : (*indexPath)->GetChildren()) {
- auto* tableInfo = Tables.FindPtr(indexImplTablePathId.second);
+ for (const auto& indexImplTablePathId : indexPath->GetChildren()) {
+ auto tableInfo = *Tables.FindPtr(indexImplTablePathId.second);
  Y_ABORT_UNLESS(tableInfo);
 
- const auto& tableStats = (*tableInfo)->GetStats().Aggregated;
+ const auto& tableStats = tableInfo->GetStats().Aggregated;
  dataSize += tableStats.DataSize + tableStats.IndexSize;
 
  auto* tableDescription = entry.AddIndexImplTableDescriptions();
  if (fillConfig) {
- FillPartitionConfig((*tableInfo)->PartitionConfig(), *tableDescription->MutablePartitionConfig());
+ FillPartitionConfig(tableInfo->PartitionConfig(), *tableDescription->MutablePartitionConfig());
  }
  if (fillBoundaries) {
- FillTableBoundaries(*tableInfo, *tableDescription->MutableSplitBoundary());
+ FillTableBoundaries(tableDescription->MutableSplitBoundary(), tableInfo);
  }
  }
  entry.SetDataSize(dataSize);
 
  if (indexInfo->Type == NKikimrSchemeOp::EIndexTypeGlobalVectorKmeansTree) {
  if (const auto* vectorIndexKmeansTreeDescription = std::get_if<NKikimrSchemeOp::TVectorIndexKmeansTreeDescription>(&indexInfo->SpecializedIndexDescription)) {
- const auto& indexInfoSettings = vectorIndexKmeansTreeDescription->GetSettings(); 
+ const auto& indexInfoSettings = vectorIndexKmeansTreeDescription->GetSettings();
  auto entrySettings = entry.MutableVectorIndexKmeansTreeDescription()->MutableSettings();
  if (indexInfoSettings.has_distance())
  entrySettings->set_distance(indexInfoSettings.distance());
@@ -1283,7 +1325,7 @@ void TSchemeShard::DescribeTableIndex(const TPathId& pathId, const TString& name
  Y_FAIL_S("SpecializedIndexDescription should be set");
  }
  }
- 
+
 }
 
 void TSchemeShard::DescribeCdcStream(const TPathId& pathId, const TString& name,
@@ -1429,22 +1471,5 @@ void TSchemeShard::DescribeBlobDepot(const TPathId& pathId, const TString& name,
  desc.SetTabletId(static_cast<ui64>(it->second->BlobDepotTabletId));
 }
 
-void TSchemeShard::FillTableBoundaries(const TTableInfo::TPtr tableInfo, google::protobuf::RepeatedPtrField<NKikimrSchemeOp::TSplitBoundary>& boundaries) {
- TString errStr;
- // Number of split boundaries equals to number of partitions - 1
- boundaries.Reserve(tableInfo->GetPartitions().size() - 1);
- for (ui32 pi = 0; pi < tableInfo->GetPartitions().size() - 1; ++pi) {
- const auto& p = tableInfo->GetPartitions()[pi];
- TSerializedCellVec endKey(p.EndOfRange);
- auto boundary = boundaries.Add()->MutableKeyPrefix();
- for (ui32 ki = 0; ki < endKey.GetCells().size(); ++ki){
- const auto& c = endKey.GetCells()[ki];
- auto type = tableInfo->Columns[tableInfo->KeyColumnIds[ki]].PType;
- bool ok = NMiniKQL::CellToValue(type, c, *boundary->AddTuple(), errStr);
- Y_ABORT_UNLESS(ok, "Failed to build key tuple at position %" PRIu32 " error: %s", ki, errStr.data());
- }
- }
-}
-
 } // NSchemeShard
 } // NKikimr