Driver for running cleanup processes on a table. TableOptimizer provides functions for various cleanup processes that improve performance on a table. Only tables that have been modified using updates or deletes are candidates for cleanup. If the table descriptor corresponds to a sharded table, table optimizer processes each physical shard. More...

#include <TableOptimizer.h>

Collaboration diagram for TableOptimizer:

Public Member Functions
	TableOptimizer (const TableDescriptor td, Executor executor, const Catalog_Namespace::Catalog &cat)

void	recomputeMetadata () const
	Recomputes per-chunk metadata for each fragment in the table. Updates and deletes can cause chunk metadata to become wider than the values in the chunk. Recomputing the metadata narrows the range to fit the chunk, as well as setting or unsetting the nulls flag as appropriate. More...

void	recomputeMetadataUnlocked (const TableUpdateMetadata &table_update_metadata) const
	Recomputes column chunk metadata for the given set of fragments. The caller of this method is expected to have already acquired the executor lock. More...

void	vacuumDeletedRows () const
	Compacts fragments to remove deleted rows. When a row is deleted, a boolean deleted system column is set to true. Vacuuming removes all deleted rows from a fragment. Note that vacuuming is a checkpointing operation, so data on disk will increase even though the number of rows for the current epoch has decreased. More...

void	vacuumFragmentsAboveMinSelectivity (const TableUpdateMetadata &table_update_metadata) const

Private Member Functions
DeletedColumnStats	recomputeDeletedColumnMetadata (const TableDescriptor *td, const std::set< size_t > &fragment_indexes={}) const

void	recomputeColumnMetadata (const TableDescriptor td, const ColumnDescriptor cd, const std::unordered_map< int, size_t > &tuple_count_map, std::optional< Data_Namespace::MemoryLevel > memory_level, const std::set< size_t > &fragment_indexes) const

std::set< size_t >	getFragmentIndexes (const TableDescriptor *td, const std::set< int > &fragment_ids) const

void	vacuumFragments (const TableDescriptor *td, const std::set< int > &fragment_ids={}) const

DeletedColumnStats	getDeletedColumnStats (const TableDescriptor *td, const std::set< size_t > &fragment_indexes) const

Private Attributes
const TableDescriptor *	td_

Executor *	executor_

const Catalog_Namespace::Catalog &	cat_

Static Private Attributes
static constexpr size_t	ROW_SET_SIZE {1000000000}

Detailed Description

Driver for running cleanup processes on a table. TableOptimizer provides functions for various cleanup processes that improve performance on a table. Only tables that have been modified using updates or deletes are candidates for cleanup. If the table descriptor corresponds to a sharded table, table optimizer processes each physical shard.

Definition at line 38 of file TableOptimizer.h.

Constructor & Destructor Documentation

TableOptimizer::TableOptimizer	(	const TableDescriptor *	td,
		Executor *	executor,
		const Catalog_Namespace::Catalog &	cat
	)

Definition at line 29 of file TableOptimizer.cpp.

References CHECK.

     : td_(td), executor_(executor), cat_(cat) {
   CHECK(td);
 }

Member Function Documentation

DeletedColumnStats TableOptimizer::getDeletedColumnStats	(	const TableDescriptor *	td,
		const std::set< size_t > &	fragment_indexes
	)		const

private

Definition at line 222 of file TableOptimizer.cpp.

References anonymous_namespace{TableOptimizer.cpp}::build_ra_exe_unit(), cat_, CHECK_EQ, DeletedColumnStats::chunk_stats_per_fragment, ColumnDescriptor::columnId, CPU, executor_, anonymous_namespace{TableOptimizer.cpp}::get_compilation_options(), anonymous_namespace{TableOptimizer.cpp}::get_execution_options(), get_logical_type_info(), get_table_infos(), Catalog_Namespace::Catalog::getDatabaseId(), Catalog_Namespace::Catalog::getDeletedColumn(), TableDescriptor::hasDeletedCol, kCOUNT, LOG, anonymous_namespace{TableOptimizer.cpp}::set_metadata_from_results(), TableDescriptor::tableId, DeletedColumnStats::total_row_count, DeletedColumnStats::visible_row_count_per_fragment, and logger::WARNING.

Referenced by recomputeDeletedColumnMetadata(), and vacuumFragmentsAboveMinSelectivity().

                                                   {
   if (!td->hasDeletedCol) {
     return {};
   }
 
   auto cd = cat_.getDeletedColumn(td);
   const auto column_id = cd->columnId;
 
   const auto input_col_desc = std::make_shared<const InputColDescriptor>(
       column_id, td->tableId, cat_.getDatabaseId(), 0);
   const auto col_expr = makeExpr<Analyzer::ColumnVar>(
       cd->columnType, shared::ColumnKey{cat_.getDatabaseId(), td->tableId, column_id}, 0);
   const auto count_expr =
       makeExpr<Analyzer::AggExpr>(cd->columnType, kCOUNT, col_expr, false, nullptr);
 
   const auto ra_exe_unit = build_ra_exe_unit(input_col_desc, {count_expr.get()});
   const auto table_infos = get_table_infos(ra_exe_unit, executor_);
   CHECK_EQ(table_infos.size(), size_t(1));
 
   const auto co = get_compilation_options(ExecutorDeviceType::CPU);
   const auto eo = get_execution_options();
 
   DeletedColumnStats deleted_column_stats;
   Executor::PerFragmentCallBack compute_deleted_callback =
       [&deleted_column_stats, cd](
           ResultSetPtr results, const Fragmenter_Namespace::FragmentInfo& fragment_info) {
         // count number of tuples in $deleted as total number of tuples in table.
         if (cd->isDeletedCol) {
           deleted_column_stats.total_row_count += fragment_info.getPhysicalNumTuples();
         }
         if (fragment_info.getPhysicalNumTuples() == 0) {
           // TODO(adb): Should not happen, but just to be safe...
           LOG(WARNING) << "Skipping completely empty fragment for column "
                        << cd->columnName;
           return;
         }
 
         const auto row = results->getNextRow(false, false);
         CHECK_EQ(row.size(), size_t(1));
 
         const auto& ti = cd->columnType;
 
         auto chunk_metadata = std::make_shared<ChunkMetadata>();
         chunk_metadata->sqlType = get_logical_type_info(ti);
 
         const auto count_val = read_scalar_target_value<int64_t>(row[0]);
 
         // min element 0 max element 1
         std::vector<TargetValue> fakerow;
 
         auto num_tuples = static_cast<size_t>(count_val);
 
         // calculate min
         if (num_tuples == fragment_info.getPhysicalNumTuples()) {
           // nothing deleted
           // min = false;
           // max = false;
           fakerow.emplace_back(TargetValue{int64_t(0)});
           fakerow.emplace_back(TargetValue{int64_t(0)});
         } else {
           if (num_tuples == 0) {
             // everything marked as delete
             // min = true
             // max = true
             fakerow.emplace_back(TargetValue{int64_t(1)});
             fakerow.emplace_back(TargetValue{int64_t(1)});
           } else {
             // some deleted
             // min = false
             // max = true;
             fakerow.emplace_back(TargetValue{int64_t(0)});
             fakerow.emplace_back(TargetValue{int64_t(1)});
           }
         }
 
         // place manufacture min and max in fake row to use common infra
         if (!set_metadata_from_results(*chunk_metadata, fakerow, ti, false)) {
           LOG(WARNING) << "Unable to process new metadata values for column "
                        << cd->columnName;
           return;
         }
 
         deleted_column_stats.chunk_stats_per_fragment.emplace(
             std::make_pair(fragment_info.fragmentId, chunk_metadata->chunkStats));
         deleted_column_stats.visible_row_count_per_fragment.emplace(
             std::make_pair(fragment_info.fragmentId, num_tuples));
       };
 
   executor_->executeWorkUnitPerFragment(ra_exe_unit,
                                         table_infos[0],
                                         co,
                                         eo,
                                         cat_,
                                         compute_deleted_callback,
                                         fragment_indexes);
   return deleted_column_stats;
 }

Here is the call graph for this function:

Here is the caller graph for this function:

std::set< size_t > TableOptimizer::getFragmentIndexes	(	const TableDescriptor *	td,
		const std::set< int > &	fragment_ids
	)		const

private

Definition at line 421 of file TableOptimizer.cpp.

References CHECK, shared::contains(), and TableDescriptor::fragmenter.

Referenced by recomputeMetadataUnlocked(), and vacuumFragmentsAboveMinSelectivity().

                                            {
   CHECK(td->fragmenter);
   auto table_info = td->fragmenter->getFragmentsForQuery();
   std::set<size_t> fragment_indexes;
   for (size_t i = 0; i < table_info.fragments.size(); i++) {
     if (shared::contains(fragment_ids, table_info.fragments[i].fragmentId)) {
       fragment_indexes.emplace(i);
     }
   }
   return fragment_indexes;
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void TableOptimizer::recomputeColumnMetadata	(	const TableDescriptor *	td,
		const ColumnDescriptor *	cd,
		const std::unordered_map< int, size_t > &	tuple_count_map,
		std::optional< Data_Namespace::MemoryLevel >	memory_level,
		const std::set< size_t > &	fragment_indexes
	)		const

private

Definition at line 322 of file TableOptimizer.cpp.

References anonymous_namespace{TableOptimizer.cpp}::build_ra_exe_unit(), cat_, CHECK, CHECK_EQ, ColumnDescriptor::columnId, ColumnDescriptor::columnName, ColumnDescriptor::columnType, CPU, executor_, TableDescriptor::fragmenter, anonymous_namespace{TableOptimizer.cpp}::get_compilation_options(), anonymous_namespace{TableOptimizer.cpp}::get_execution_options(), get_logical_type_info(), get_table_infos(), Catalog_Namespace::Catalog::getDatabaseId(), logger::INFO, kCOUNT, kINT, kMAX, kMIN, LOG, anonymous_namespace{TableOptimizer.cpp}::set_metadata_from_results(), TableDescriptor::tableId, and logger::WARNING.

Referenced by recomputeMetadata(), and recomputeMetadataUnlocked().

                                                   {
   const auto ti = cd->columnType;
   if (ti.is_varlen()) {
     LOG(INFO) << "Skipping varlen column " << cd->columnName;
     return;
   }
 
   const auto column_id = cd->columnId;
   const auto input_col_desc = std::make_shared<const InputColDescriptor>(
       column_id, td->tableId, cat_.getDatabaseId(), 0);
   const auto col_expr = makeExpr<Analyzer::ColumnVar>(
       cd->columnType, shared::ColumnKey{cat_.getDatabaseId(), td->tableId, column_id}, 0);
   auto max_expr =
       makeExpr<Analyzer::AggExpr>(cd->columnType, kMAX, col_expr, false, nullptr);
   auto min_expr =
       makeExpr<Analyzer::AggExpr>(cd->columnType, kMIN, col_expr, false, nullptr);
   auto count_expr =
       makeExpr<Analyzer::AggExpr>(cd->columnType, kCOUNT, col_expr, false, nullptr);
 
   if (ti.is_string()) {
     const SQLTypeInfo fun_ti(kINT);
     const auto fun_expr = makeExpr<Analyzer::KeyForStringExpr>(col_expr);
     max_expr = makeExpr<Analyzer::AggExpr>(fun_ti, kMAX, fun_expr, false, nullptr);
     min_expr = makeExpr<Analyzer::AggExpr>(fun_ti, kMIN, fun_expr, false, nullptr);
   }
   const auto ra_exe_unit = build_ra_exe_unit(
       input_col_desc, {min_expr.get(), max_expr.get(), count_expr.get()});
   const auto table_infos = get_table_infos(ra_exe_unit, executor_);
   CHECK_EQ(table_infos.size(), size_t(1));
 
   const auto co = get_compilation_options(ExecutorDeviceType::CPU);
   const auto eo = get_execution_options();
 
   std::unordered_map</*fragment_id*/ int, ChunkStats> stats_map;
 
   Executor::PerFragmentCallBack compute_metadata_callback =
       [&stats_map, &tuple_count_map, cd](
           ResultSetPtr results, const Fragmenter_Namespace::FragmentInfo& fragment_info) {
         if (fragment_info.getPhysicalNumTuples() == 0) {
           // TODO(adb): Should not happen, but just to be safe...
           LOG(WARNING) << "Skipping completely empty fragment for column "
                        << cd->columnName;
           return;
         }
 
         const auto row = results->getNextRow(false, false);
         CHECK_EQ(row.size(), size_t(3));
 
         const auto& ti = cd->columnType;
 
         auto chunk_metadata = std::make_shared<ChunkMetadata>();
         chunk_metadata->sqlType = get_logical_type_info(ti);
 
         const auto count_val = read_scalar_target_value<int64_t>(row[2]);
         if (count_val == 0) {
           // Assume chunk of all nulls, bail
           return;
         }
 
         bool has_nulls = true;  // default to wide
         auto tuple_count_itr = tuple_count_map.find(fragment_info.fragmentId);
         if (tuple_count_itr != tuple_count_map.end()) {
           has_nulls = !(static_cast<size_t>(count_val) == tuple_count_itr->second);
         } else {
           // no deleted column calc so use raw physical count
           has_nulls =
               !(static_cast<size_t>(count_val) == fragment_info.getPhysicalNumTuples());
         }
 
         if (!set_metadata_from_results(*chunk_metadata, row, ti, has_nulls)) {
           LOG(WARNING) << "Unable to process new metadata values for column "
                        << cd->columnName;
           return;
         }
 
         stats_map.emplace(
             std::make_pair(fragment_info.fragmentId, chunk_metadata->chunkStats));
       };
 
   executor_->executeWorkUnitPerFragment(ra_exe_unit,
                                         table_infos[0],
                                         co,
                                         eo,
                                         cat_,
                                         compute_metadata_callback,
                                         fragment_indexes);
 
   auto* fragmenter = td->fragmenter.get();
   CHECK(fragmenter);
   fragmenter->updateChunkStats(cd, stats_map, memory_level);
 }

Here is the call graph for this function:

Here is the caller graph for this function:

DeletedColumnStats TableOptimizer::recomputeDeletedColumnMetadata	(	const TableDescriptor *	td,
		const std::set< size_t > &	fragment_indexes = `{}`
	)		const

private

Definition at line 206 of file TableOptimizer.cpp.

References cat_, CHECK, TableDescriptor::fragmenter, Catalog_Namespace::Catalog::getDeletedColumn(), getDeletedColumnStats(), and TableDescriptor::hasDeletedCol.

Referenced by recomputeMetadata(), and recomputeMetadataUnlocked().

                                                   {
   if (!td->hasDeletedCol) {
     return {};
   }
 
   auto stats = getDeletedColumnStats(td, fragment_indexes);
   auto* fragmenter = td->fragmenter.get();
   CHECK(fragmenter);
   auto cd = cat_.getDeletedColumn(td);
   fragmenter->updateChunkStats(cd, stats.chunk_stats_per_fragment, {});
   fragmenter->setNumRows(stats.total_row_count);
   return stats;
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void TableOptimizer::recomputeMetadata ( ) const

Recomputes per-chunk metadata for each fragment in the table. Updates and deletes can cause chunk metadata to become wider than the values in the chunk. Recomputing the metadata narrows the range to fit the chunk, as well as setting or unsetting the nulls flag as appropriate.

Definition at line 134 of file TableOptimizer.cpp.

References cat_, CHECK_GE, Catalog_Namespace::DBMetadata::dbId, DEBUG_TIMER, executor_, Catalog_Namespace::Catalog::getAllColumnMetadataForTable(), Catalog_Namespace::Catalog::getCurrentDB(), Catalog_Namespace::Catalog::getDataMgr(), Catalog_Namespace::Catalog::getPhysicalTablesDescriptors(), lockmgr::TableLockMgrImpl< TableDataLockMgr >::getWriteLockForTable(), logger::INFO, LOG, TableDescriptor::nShards, recomputeColumnMetadata(), recomputeDeletedColumnMetadata(), ROW_SET_SIZE, TableDescriptor::tableId, TableDescriptor::tableName, and td_.

Referenced by Parser::OptimizeTableStmt::execute(), and migrations::MigrationMgr::migrateDateInDaysMetadata().

                                              {
   auto timer = DEBUG_TIMER(__func__);
   heavyai::unique_lock<heavyai::shared_mutex> lock(executor_->execute_mutex_);
 
   LOG(INFO) << "Recomputing metadata for " << td_->tableName;
 
   CHECK_GE(td_->tableId, 0);
 
   std::vector<const TableDescriptor*> table_descriptors;
   if (td_->nShards > 0) {
     const auto physical_tds = cat_.getPhysicalTablesDescriptors(td_);
     table_descriptors.insert(
         table_descriptors.begin(), physical_tds.begin(), physical_tds.end());
   } else {
     table_descriptors.push_back(td_);
   }
 
   auto& data_mgr = cat_.getDataMgr();
 
   // acquire write lock on table data
   auto data_lock = lockmgr::TableDataLockMgr::getWriteLockForTable(cat_, td_->tableName);
 
   for (const auto td : table_descriptors) {
     ScopeGuard row_set_holder = [this] { executor_->row_set_mem_owner_ = nullptr; };
     executor_->row_set_mem_owner_ =
         std::make_shared<RowSetMemoryOwner>(ROW_SET_SIZE, /*num_threads=*/1);
     const auto table_id = td->tableId;
     auto stats = recomputeDeletedColumnMetadata(td);
 
     // TODO(adb): Support geo
     auto col_descs = cat_.getAllColumnMetadataForTable(table_id, false, false, false);
     for (const auto& cd : col_descs) {
       recomputeColumnMetadata(td, cd, stats.visible_row_count_per_fragment, {}, {});
     }
     data_mgr.checkpoint(cat_.getCurrentDB().dbId, table_id);
     executor_->clearMetaInfoCache();
   }
 
   data_mgr.clearMemory(Data_Namespace::MemoryLevel::CPU_LEVEL);
   if (data_mgr.gpusPresent()) {
     data_mgr.clearMemory(Data_Namespace::MemoryLevel::GPU_LEVEL);
   }
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void TableOptimizer::recomputeMetadataUnlocked ( const TableUpdateMetadata & table_update_metadata ) const

Recomputes column chunk metadata for the given set of fragments. The caller of this method is expected to have already acquired the executor lock.

Definition at line 178 of file TableOptimizer.cpp.

References cat_, CHECK, TableUpdateMetadata::columns_for_metadata_update, Data_Namespace::CPU_LEVEL, DEBUG_TIMER, getFragmentIndexes(), Catalog_Namespace::Catalog::getMetadataForTable(), recomputeColumnMetadata(), and recomputeDeletedColumnMetadata().

                                                             {
   auto timer = DEBUG_TIMER(__func__);
   std::map<int, std::list<const ColumnDescriptor*>> columns_by_table_id;
   auto& columns_for_update = table_update_metadata.columns_for_metadata_update;
   for (const auto& entry : columns_for_update) {
     auto column_descriptor = entry.first;
     columns_by_table_id[column_descriptor->tableId].emplace_back(column_descriptor);
   }
 
   for (const auto& [table_id, columns] : columns_by_table_id) {
     auto td = cat_.getMetadataForTable(table_id);
     auto stats = recomputeDeletedColumnMetadata(td);
     for (const auto cd : columns) {
       CHECK(columns_for_update.find(cd) != columns_for_update.end());
       auto fragment_indexes = getFragmentIndexes(td, columns_for_update.find(cd)->second);
       recomputeColumnMetadata(td,
                               cd,
                               stats.visible_row_count_per_fragment,
                               Data_Namespace::MemoryLevel::CPU_LEVEL,
                               fragment_indexes);
     }
   }
 }

Here is the call graph for this function:

void TableOptimizer::vacuumDeletedRows ( ) const

Compacts fragments to remove deleted rows. When a row is deleted, a boolean deleted system column is set to true. Vacuuming removes all deleted rows from a fragment. Note that vacuuming is a checkpointing operation, so data on disk will increase even though the number of rows for the current epoch has decreased.

Definition at line 435 of file TableOptimizer.cpp.

References cat_, Catalog_Namespace::Catalog::checkpoint(), File_Namespace::GlobalFileMgr::compactDataFiles(), DEBUG_TIMER, Catalog_Namespace::Catalog::getDatabaseId(), Catalog_Namespace::Catalog::getDataMgr(), Data_Namespace::DataMgr::getGlobalFileMgr(), Catalog_Namespace::Catalog::getPhysicalTablesDescriptors(), Catalog_Namespace::Catalog::getTableEpochs(), lockmgr::TableLockMgrImpl< TableDataLockMgr >::getWriteLockForTable(), Catalog_Namespace::Catalog::removeFragmenterForTable(), Catalog_Namespace::Catalog::setTableEpochsLogExceptions(), TableDescriptor::tableId, td_, and vacuumFragments().

Referenced by Parser::OptimizeTableStmt::execute().

                                              {
   auto timer = DEBUG_TIMER(__func__);
   const auto table_id = td_->tableId;
   const auto db_id = cat_.getDatabaseId();
   const auto table_lock =
       lockmgr::TableDataLockMgr::getWriteLockForTable({db_id, table_id});
   const auto table_epochs = cat_.getTableEpochs(db_id, table_id);
   const auto shards = cat_.getPhysicalTablesDescriptors(td_);
   try {
     for (const auto shard : shards) {
       vacuumFragments(shard);
     }
     cat_.checkpoint(table_id);
   } catch (...) {
     cat_.setTableEpochsLogExceptions(db_id, table_epochs);
     throw;
   }
 
   for (auto shard : shards) {
     cat_.removeFragmenterForTable(shard->tableId);
     cat_.getDataMgr().getGlobalFileMgr()->compactDataFiles(cat_.getDatabaseId(),
                                                            shard->tableId);
   }
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void TableOptimizer::vacuumFragments	(	const TableDescriptor *	td,
		const std::set< int > &	fragment_ids = `{}`
	)		const

private

Definition at line 460 of file TableOptimizer.cpp.

References cat_, UpdelRoll::catalog, CHECK_EQ, CHUNK_KEY_COLUMN_IDX, CHUNK_KEY_FRAGMENT_IDX, ColumnDescriptor::columnId, shared::contains(), Data_Namespace::CPU_LEVEL, TableDescriptor::fragmenter, Chunk_NS::Chunk::getChunk(), Data_Namespace::DataMgr::getChunkMetadataVecForKeyPrefix(), Catalog_Namespace::Catalog::getDatabaseId(), Catalog_Namespace::Catalog::getDataMgr(), Catalog_Namespace::Catalog::getDeletedColumn(), Catalog_Namespace::Catalog::getLogicalTableId(), UpdelRoll::logicalTableId, UpdelRoll::memoryLevel, UpdelRoll::stageUpdate(), UpdelRoll::table_descriptor, and TableDescriptor::tableId.

Referenced by vacuumDeletedRows(), and vacuumFragmentsAboveMinSelectivity().

                                                                             {
   // "if not a table that supports delete return,  nothing more to do"
   const ColumnDescriptor* cd = cat_.getDeletedColumn(td);
   if (nullptr == cd) {
     return;
   }
   // vacuum chunks which show sign of deleted rows in metadata
   ChunkKey chunk_key_prefix = {cat_.getDatabaseId(), td->tableId, cd->columnId};
   ChunkMetadataVector chunk_metadata_vec;
   cat_.getDataMgr().getChunkMetadataVecForKeyPrefix(chunk_metadata_vec, chunk_key_prefix);
   for (auto& [chunk_key, chunk_metadata] : chunk_metadata_vec) {
     auto fragment_id = chunk_key[CHUNK_KEY_FRAGMENT_IDX];
     // If delete has occurred, only vacuum fragments that are in the fragment_ids set.
     // Empty fragment_ids set implies all fragments.
     if (chunk_metadata->chunkStats.max.tinyintval == 1 &&
         (fragment_ids.empty() || shared::contains(fragment_ids, fragment_id))) {
       UpdelRoll updel_roll;
       updel_roll.catalog = &cat_;
       updel_roll.logicalTableId = cat_.getLogicalTableId(td->tableId);
       updel_roll.memoryLevel = Data_Namespace::MemoryLevel::CPU_LEVEL;
       updel_roll.table_descriptor = td;
       CHECK_EQ(cd->columnId, chunk_key[CHUNK_KEY_COLUMN_IDX]);
       const auto chunk = Chunk_NS::Chunk::getChunk(cd,
                                                    &cat_.getDataMgr(),
                                                    chunk_key,
                                                    updel_roll.memoryLevel,
                                                    0,
                                                    chunk_metadata->numBytes,
                                                    chunk_metadata->numElements);
       td->fragmenter->compactRows(&cat_,
                                   td,
                                   fragment_id,
                                   td->fragmenter->getVacuumOffsets(chunk),
                                   updel_roll.memoryLevel,
                                   updel_roll);
       updel_roll.stageUpdate();
     }
   }
   td->fragmenter->resetSizesFromFragments();
 }

Here is the call graph for this function:

Here is the caller graph for this function:

void TableOptimizer::vacuumFragmentsAboveMinSelectivity ( const TableUpdateMetadata & table_update_metadata ) const

Vacuums fragments with a deleted rows percentage that exceeds the configured minimum vacuum selectivity threshold.

Definition at line 502 of file TableOptimizer.cpp.

References cat_, Catalog_Namespace::Catalog::checkpoint(), Catalog_Namespace::Catalog::checkpointWithAutoRollback(), DEBUG_TIMER, Data_Namespace::DISK_LEVEL, executor_, TableUpdateMetadata::fragments_with_deleted_rows, g_vacuum_min_selectivity, Catalog_Namespace::Catalog::getDatabaseId(), getDeletedColumnStats(), getFragmentIndexes(), Catalog_Namespace::Catalog::getMetadataForTable(), Catalog_Namespace::Catalog::getTableEpochs(), lockmgr::TableLockMgrImpl< TableDataLockMgr >::getWriteLockForTable(), TableDescriptor::persistenceLevel, shared::printContainer(), ROW_SET_SIZE, Catalog_Namespace::Catalog::setTableEpochsLogExceptions(), TableDescriptor::tableId, td_, vacuumFragments(), DeletedColumnStats::visible_row_count_per_fragment, and VLOG.

                                                             {
   if (td_->persistenceLevel != Data_Namespace::MemoryLevel::DISK_LEVEL) {
     return;
   }
   auto timer = DEBUG_TIMER(__func__);
   std::map<const TableDescriptor*, std::set<int32_t>> fragments_to_vacuum;
   for (const auto& [table_id, fragment_ids] :
        table_update_metadata.fragments_with_deleted_rows) {
     auto td = cat_.getMetadataForTable(table_id);
     // Skip automatic vacuuming for tables with uncapped epoch
     if (td->maxRollbackEpochs == -1) {
       continue;
     }
 
     DeletedColumnStats deleted_column_stats;
     {
       heavyai::unique_lock<heavyai::shared_mutex> executor_lock(
           executor_->execute_mutex_);
       ScopeGuard row_set_holder = [this] { executor_->row_set_mem_owner_ = nullptr; };
       executor_->row_set_mem_owner_ =
           std::make_shared<RowSetMemoryOwner>(ROW_SET_SIZE, /*num_threads=*/1);
       deleted_column_stats =
           getDeletedColumnStats(td, getFragmentIndexes(td, fragment_ids));
       executor_->clearMetaInfoCache();
     }
 
     std::set<int32_t> filtered_fragment_ids;
     for (const auto [fragment_id, visible_row_count] :
          deleted_column_stats.visible_row_count_per_fragment) {
       auto total_row_count =
           td->fragmenter->getFragmentInfo(fragment_id)->getPhysicalNumTuples();
       float deleted_row_count = total_row_count - visible_row_count;
       if ((deleted_row_count / total_row_count) >= g_vacuum_min_selectivity) {
         filtered_fragment_ids.emplace(fragment_id);
       }
     }
 
     if (!filtered_fragment_ids.empty()) {
       fragments_to_vacuum[td] = filtered_fragment_ids;
     }
   }
 
   if (!fragments_to_vacuum.empty()) {
     const auto db_id = cat_.getDatabaseId();
     const auto table_lock =
         lockmgr::TableDataLockMgr::getWriteLockForTable({db_id, td_->tableId});
     const auto table_epochs = cat_.getTableEpochs(db_id, td_->tableId);
     try {
       for (const auto& [td, fragment_ids] : fragments_to_vacuum) {
         vacuumFragments(td, fragment_ids);
         VLOG(1) << "Auto-vacuumed fragments: " << shared::printContainer(fragment_ids)
                 << ", table id: " << td->tableId;
       }
       cat_.checkpoint(td_->tableId);
     } catch (...) {
       cat_.setTableEpochsLogExceptions(db_id, table_epochs);
       throw;
     }
   } else {
     // Checkpoint, even when no data update occurs, in order to ensure that epochs are
     // uniformly incremented in distributed mode.
     cat_.checkpointWithAutoRollback(td_->tableId);
   }
 }

Here is the call graph for this function:

Member Data Documentation

const Catalog_Namespace::Catalog& TableOptimizer::cat_

private

Definition at line 98 of file TableOptimizer.h.

Referenced by getDeletedColumnStats(), recomputeColumnMetadata(), recomputeDeletedColumnMetadata(), recomputeMetadata(), recomputeMetadataUnlocked(), vacuumDeletedRows(), vacuumFragments(), and vacuumFragmentsAboveMinSelectivity().

Executor* TableOptimizer::executor_

private

Definition at line 97 of file TableOptimizer.h.

Referenced by getDeletedColumnStats(), recomputeColumnMetadata(), recomputeMetadata(), and vacuumFragmentsAboveMinSelectivity().

constexpr size_t TableOptimizer::ROW_SET_SIZE {1000000000}

staticprivate

Definition at line 101 of file TableOptimizer.h.

Referenced by recomputeMetadata(), and vacuumFragmentsAboveMinSelectivity().

const TableDescriptor* TableOptimizer::td_

private

Definition at line 96 of file TableOptimizer.h.

Referenced by recomputeMetadata(), vacuumDeletedRows(), and vacuumFragmentsAboveMinSelectivity().

The documentation for this class was generated from the following files:

/home/jenkins-slave/workspace/core-os-doxygen/QueryEngine/TableOptimizer.h
/home/jenkins-slave/workspace/core-os-doxygen/QueryEngine/TableOptimizer.cpp

Public Member Functions

Private Member Functions

Private Attributes

Static Private Attributes

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation

Member Data Documentation