All Downloads are FREE. Search and download functionalities are using the official Maven repository.

com.facebook.presto.hive.metastore.SemiTransactionalHiveMetastore Maven / Gradle / Ivy

/*
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
package com.facebook.presto.hive.metastore;

import com.facebook.airlift.log.Logger;
import com.facebook.presto.common.predicate.Domain;
import com.facebook.presto.common.type.Type;
import com.facebook.presto.hive.HdfsContext;
import com.facebook.presto.hive.HdfsEnvironment;
import com.facebook.presto.hive.HiveBasicStatistics;
import com.facebook.presto.hive.HiveType;
import com.facebook.presto.hive.LocationHandle.WriteMode;
import com.facebook.presto.hive.PartitionNotFoundException;
import com.facebook.presto.hive.TableAlreadyExistsException;
import com.facebook.presto.spi.ConnectorSession;
import com.facebook.presto.spi.PrestoException;
import com.facebook.presto.spi.SchemaTableName;
import com.facebook.presto.spi.StandardErrorCode;
import com.facebook.presto.spi.TableNotFoundException;
import com.facebook.presto.spi.security.PrestoPrincipal;
import com.facebook.presto.spi.security.PrincipalType;
import com.facebook.presto.spi.security.RoleGrant;
import com.facebook.presto.spi.statistics.ColumnStatisticType;
import com.google.common.annotations.VisibleForTesting;
import com.google.common.base.Joiner;
import com.google.common.collect.ImmutableList;
import com.google.common.collect.ImmutableMap;
import com.google.common.collect.ImmutableSet;
import com.google.common.collect.Iterables;
import com.google.common.collect.Lists;
import com.google.common.util.concurrent.ListenableFuture;
import com.google.common.util.concurrent.ListeningExecutorService;
import org.apache.hadoop.fs.FileStatus;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;

import javax.annotation.concurrent.GuardedBy;

import java.io.FileNotFoundException;
import java.io.IOException;
import java.util.ArrayList;
import java.util.Collection;
import java.util.HashMap;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Optional;
import java.util.OptionalLong;
import java.util.Set;
import java.util.concurrent.atomic.AtomicBoolean;

import static com.facebook.airlift.concurrent.MoreFutures.getFutureValue;
import static com.facebook.presto.hive.HiveErrorCode.HIVE_CORRUPTED_COLUMN_STATISTICS;
import static com.facebook.presto.hive.HiveErrorCode.HIVE_FILESYSTEM_ERROR;
import static com.facebook.presto.hive.HiveErrorCode.HIVE_METASTORE_ERROR;
import static com.facebook.presto.hive.HiveErrorCode.HIVE_PATH_ALREADY_EXISTS;
import static com.facebook.presto.hive.HiveErrorCode.HIVE_TABLE_DROPPED_DURING_QUERY;
import static com.facebook.presto.hive.LocationHandle.WriteMode.DIRECT_TO_TARGET_NEW_DIRECTORY;
import static com.facebook.presto.hive.metastore.HivePrivilegeInfo.HivePrivilege.OWNERSHIP;
import static com.facebook.presto.hive.metastore.MetastoreUtil.PRESTO_QUERY_ID_NAME;
import static com.facebook.presto.hive.metastore.MetastoreUtil.convertPredicateToParts;
import static com.facebook.presto.hive.metastore.MetastoreUtil.createDirectory;
import static com.facebook.presto.hive.metastore.MetastoreUtil.getFileSystem;
import static com.facebook.presto.hive.metastore.MetastoreUtil.isPrestoView;
import static com.facebook.presto.hive.metastore.MetastoreUtil.pathExists;
import static com.facebook.presto.hive.metastore.MetastoreUtil.renameFile;
import static com.facebook.presto.hive.metastore.MetastoreUtil.toPartitionValues;
import static com.facebook.presto.hive.metastore.PrestoTableType.MANAGED_TABLE;
import static com.facebook.presto.hive.metastore.PrestoTableType.TEMPORARY_TABLE;
import static com.facebook.presto.hive.metastore.PrestoTableType.VIRTUAL_VIEW;
import static com.facebook.presto.hive.metastore.Statistics.ReduceOperator.SUBTRACT;
import static com.facebook.presto.hive.metastore.Statistics.merge;
import static com.facebook.presto.hive.metastore.Statistics.reduce;
import static com.facebook.presto.spi.StandardErrorCode.ALREADY_EXISTS;
import static com.facebook.presto.spi.StandardErrorCode.NOT_SUPPORTED;
import static com.facebook.presto.spi.StandardErrorCode.TRANSACTION_CONFLICT;
import static com.facebook.presto.spi.security.PrincipalType.USER;
import static com.google.common.base.MoreObjects.toStringHelper;
import static com.google.common.base.Preconditions.checkArgument;
import static com.google.common.base.Preconditions.checkState;
import static com.google.common.base.Verify.verify;
import static com.google.common.collect.ImmutableList.toImmutableList;
import static com.google.common.util.concurrent.Futures.whenAllSucceed;
import static com.google.common.util.concurrent.MoreExecutors.directExecutor;
import static java.lang.String.format;
import static java.util.Objects.requireNonNull;
import static java.util.stream.Collectors.toList;
import static org.apache.hadoop.hive.common.FileUtils.makePartName;

public class SemiTransactionalHiveMetastore
{
    private static final Logger log = Logger.get(SemiTransactionalHiveMetastore.class);
    private static final int PARTITION_COMMIT_BATCH_SIZE = 8;

    private final ExtendedHiveMetastore delegate;
    private final HdfsEnvironment hdfsEnvironment;
    private final ListeningExecutorService renameExecutor;
    private final boolean skipDeletionForAlter;
    private final boolean skipTargetCleanupOnRollback;

    @GuardedBy("this")
    private final Map> tableActions = new HashMap<>();
    @GuardedBy("this")
    private final Map, Action>> partitionActions = new HashMap<>();
    @GuardedBy("this")
    private final List declaredIntentionsToWrite = new ArrayList<>();
    @GuardedBy("this")
    private ExclusiveOperation bufferedExclusiveOperation;
    @GuardedBy("this")
    private State state = State.EMPTY;
    private boolean throwOnCleanupFailure;

    public SemiTransactionalHiveMetastore(
            HdfsEnvironment hdfsEnvironment,
            ExtendedHiveMetastore delegate,
            ListeningExecutorService renameExecutor,
            boolean skipDeletionForAlter,
            boolean skipTargetCleanupOnRollback)
    {
        this.hdfsEnvironment = requireNonNull(hdfsEnvironment, "hdfsEnvironment is null");
        this.delegate = requireNonNull(delegate, "delegate is null");
        this.renameExecutor = requireNonNull(renameExecutor, "renameExecutor is null");
        this.skipDeletionForAlter = skipDeletionForAlter;
        this.skipTargetCleanupOnRollback = skipTargetCleanupOnRollback;
    }

    public synchronized List getAllDatabases()
    {
        checkReadable();
        return delegate.getAllDatabases();
    }

    public synchronized Optional getDatabase(String databaseName)
    {
        checkReadable();
        return delegate.getDatabase(databaseName);
    }

    public synchronized Optional> getAllTables(String databaseName)
    {
        checkReadable();
        if (!tableActions.isEmpty()) {
            throw new UnsupportedOperationException("Listing all tables after adding/dropping/altering tables/views in a transaction is not supported");
        }
        return delegate.getAllTables(databaseName);
    }

    public synchronized Optional getTable(String databaseName, String tableName)
    {
        checkReadable();
        Action tableAction = tableActions.get(new SchemaTableName(databaseName, tableName));
        if (tableAction == null) {
            return delegate.getTable(databaseName, tableName);
        }
        switch (tableAction.getType()) {
            case ADD:
            case ALTER:
            case INSERT_EXISTING:
                return Optional.of(tableAction.getData().getAugmentedTableForInTransactionRead());
            case DROP:
                return Optional.empty();
            default:
                throw new IllegalStateException("Unknown action type");
        }
    }

    public synchronized Set getSupportedColumnStatistics(Type type)
    {
        return delegate.getSupportedColumnStatistics(type);
    }

    public synchronized PartitionStatistics getTableStatistics(String databaseName, String tableName)
    {
        checkReadable();
        Action tableAction = tableActions.get(new SchemaTableName(databaseName, tableName));
        if (tableAction == null) {
            return delegate.getTableStatistics(databaseName, tableName);
        }
        switch (tableAction.getType()) {
            case ADD:
            case ALTER:
            case INSERT_EXISTING:
                return tableAction.getData().getStatistics();
            case DROP:
                return PartitionStatistics.empty();
            default:
                throw new IllegalStateException("Unknown action type");
        }
    }

    public synchronized Map getPartitionStatistics(String databaseName, String tableName, Set partitionNames)
    {
        checkReadable();
        Optional
table = getTable(databaseName, tableName); if (!table.isPresent()) { return ImmutableMap.of(); } TableSource tableSource = getTableSource(databaseName, tableName); Map, Action> partitionActionsOfTable = partitionActions.computeIfAbsent(new SchemaTableName(databaseName, tableName), k -> new HashMap<>()); ImmutableSet.Builder partitionNamesToQuery = ImmutableSet.builder(); ImmutableMap.Builder resultBuilder = ImmutableMap.builder(); for (String partitionName : partitionNames) { List partitionValues = toPartitionValues(partitionName); Action partitionAction = partitionActionsOfTable.get(partitionValues); if (partitionAction == null) { switch (tableSource) { case PRE_EXISTING_TABLE: partitionNamesToQuery.add(partitionName); break; case CREATED_IN_THIS_TRANSACTION: resultBuilder.put(partitionName, PartitionStatistics.empty()); break; default: throw new UnsupportedOperationException("unknown table source"); } } else { resultBuilder.put(partitionName, partitionAction.getData().getStatistics()); } } Map delegateResult = delegate.getPartitionStatistics(databaseName, tableName, partitionNamesToQuery.build()); if (!delegateResult.isEmpty()) { resultBuilder.putAll(delegateResult); } else { partitionNamesToQuery.build().forEach(partitionName -> resultBuilder.put(partitionName, PartitionStatistics.empty())); } return resultBuilder.build(); } /** * This method can only be called when the table is known to exist */ @GuardedBy("this") private TableSource getTableSource(String databaseName, String tableName) { checkHoldsLock(); checkReadable(); Action tableAction = tableActions.get(new SchemaTableName(databaseName, tableName)); if (tableAction == null) { return TableSource.PRE_EXISTING_TABLE; } switch (tableAction.getType()) { case ADD: return TableSource.CREATED_IN_THIS_TRANSACTION; case ALTER: throw new IllegalStateException("Tables are never altered in the current implementation"); case DROP: throw new TableNotFoundException(new SchemaTableName(databaseName, tableName)); case INSERT_EXISTING: return TableSource.PRE_EXISTING_TABLE; default: throw new IllegalStateException("Unknown action type"); } } public synchronized HivePageSinkMetadata generatePageSinkMetadata(SchemaTableName schemaTableName) { checkReadable(); Optional
table = getTable(schemaTableName.getSchemaName(), schemaTableName.getTableName()); if (!table.isPresent()) { return new HivePageSinkMetadata(schemaTableName, Optional.empty(), ImmutableMap.of()); } Map, Action> partitionActionMap = partitionActions.get(schemaTableName); Map, Optional> modifiedPartitionMap; if (partitionActionMap == null) { modifiedPartitionMap = ImmutableMap.of(); } else { ImmutableMap.Builder, Optional> modifiedPartitionMapBuilder = ImmutableMap.builder(); for (Map.Entry, Action> entry : partitionActionMap.entrySet()) { modifiedPartitionMapBuilder.put(entry.getKey(), getPartitionFromPartitionAction(entry.getValue())); } modifiedPartitionMap = modifiedPartitionMapBuilder.build(); } return new HivePageSinkMetadata( schemaTableName, table, modifiedPartitionMap); } public synchronized Optional> getAllViews(String databaseName) { checkReadable(); if (!tableActions.isEmpty()) { throw new UnsupportedOperationException("Listing all tables after adding/dropping/altering tables/views in a transaction is not supported"); } return delegate.getAllViews(databaseName); } public synchronized void createDatabase(Database database) { setExclusive((delegate, hdfsEnvironment) -> delegate.createDatabase(database)); } public synchronized void dropDatabase(String schemaName) { setExclusive((delegate, hdfsEnvironment) -> delegate.dropDatabase(schemaName)); } public synchronized void renameDatabase(String source, String target) { setExclusive((delegate, hdfsEnvironment) -> delegate.renameDatabase(source, target)); } // TODO: Allow updating statistics for 2 tables in the same transaction public synchronized void setTableStatistics(Table table, PartitionStatistics tableStatistics) { setExclusive((delegate, hdfsEnvironment) -> delegate.updateTableStatistics(table.getDatabaseName(), table.getTableName(), statistics -> updatePartitionStatistics(statistics, tableStatistics))); } // TODO: Allow updating statistics for 2 tables in the same transaction public synchronized void setPartitionStatistics(Table table, Map, PartitionStatistics> partitionStatisticsMap) { setExclusive((delegate, hdfsEnvironment) -> partitionStatisticsMap.forEach((partitionValues, newPartitionStats) -> delegate.updatePartitionStatistics( table.getDatabaseName(), table.getTableName(), getPartitionName(table, partitionValues), oldPartitionStats -> updatePartitionStatistics(oldPartitionStats, newPartitionStats)))); } // For HiveBasicStatistics, we only overwrite the original statistics if the new one is not empty. // For HiveColumnStatistics, we always overwrite every statistics. // TODO: Collect file count, on-disk size and in-memory size during ANALYZE private PartitionStatistics updatePartitionStatistics(PartitionStatistics oldPartitionStats, PartitionStatistics newPartitionStats) { HiveBasicStatistics oldBasicStatistics = oldPartitionStats.getBasicStatistics(); HiveBasicStatistics newBasicStatistics = newPartitionStats.getBasicStatistics(); HiveBasicStatistics updatedBasicStatistics = new HiveBasicStatistics( firstPresent(newBasicStatistics.getFileCount(), oldBasicStatistics.getFileCount()), firstPresent(newBasicStatistics.getRowCount(), oldBasicStatistics.getRowCount()), firstPresent(newBasicStatistics.getInMemoryDataSizeInBytes(), oldBasicStatistics.getInMemoryDataSizeInBytes()), firstPresent(newBasicStatistics.getOnDiskDataSizeInBytes(), oldBasicStatistics.getOnDiskDataSizeInBytes())); return new PartitionStatistics(updatedBasicStatistics, newPartitionStats.getColumnStatistics()); } private static OptionalLong firstPresent(OptionalLong first, OptionalLong second) { return first.isPresent() ? first : second; } /** * {@code currentLocation} needs to be supplied if a writePath exists for the table. */ public synchronized void createTable( ConnectorSession session, Table table, PrincipalPrivileges principalPrivileges, Optional currentPath, boolean ignoreExisting, PartitionStatistics statistics) { setShared(); // When creating a table, it should never have partition actions. This is just a sanity check. checkNoPartitionAction(table.getDatabaseName(), table.getTableName()); SchemaTableName schemaTableName = new SchemaTableName(table.getDatabaseName(), table.getTableName()); Action oldTableAction = tableActions.get(schemaTableName); TableAndMore tableAndMore = new TableAndMore(table, Optional.of(principalPrivileges), currentPath, Optional.empty(), ignoreExisting, statistics, statistics); if (oldTableAction == null) { HdfsContext context = new HdfsContext(session, table.getDatabaseName(), table.getTableName()); tableActions.put(schemaTableName, new Action<>(ActionType.ADD, tableAndMore, context)); return; } switch (oldTableAction.getType()) { case DROP: throw new PrestoException(TRANSACTION_CONFLICT, "Dropping and then recreating the same table in a transaction is not supported"); case ADD: case ALTER: case INSERT_EXISTING: throw new TableAlreadyExistsException(schemaTableName); default: throw new IllegalStateException("Unknown action type"); } } public synchronized void dropTable(ConnectorSession session, String databaseName, String tableName) { setShared(); // Dropping table with partition actions requires cleaning up staging data, which is not implemented yet. checkNoPartitionAction(databaseName, tableName); SchemaTableName schemaTableName = new SchemaTableName(databaseName, tableName); Action oldTableAction = tableActions.get(schemaTableName); if (oldTableAction == null || oldTableAction.getType() == ActionType.ALTER) { HdfsContext context = new HdfsContext(session, databaseName, tableName); tableActions.put(schemaTableName, new Action<>(ActionType.DROP, null, context)); return; } switch (oldTableAction.getType()) { case DROP: throw new TableNotFoundException(schemaTableName); case ADD: case ALTER: case INSERT_EXISTING: throw new UnsupportedOperationException("dropping a table added/modified in the same transaction is not supported"); default: throw new IllegalStateException("Unknown action type"); } } public synchronized void replaceView(String databaseName, String tableName, Table table, PrincipalPrivileges principalPrivileges) { setExclusive((delegate, hdfsEnvironment) -> delegate.replaceTable(databaseName, tableName, table, principalPrivileges)); } public synchronized void renameTable(String databaseName, String tableName, String newDatabaseName, String newTableName) { setExclusive((delegate, hdfsEnvironment) -> delegate.renameTable(databaseName, tableName, newDatabaseName, newTableName)); } public synchronized void addColumn(String databaseName, String tableName, String columnName, HiveType columnType, String columnComment) { setExclusive((delegate, hdfsEnvironment) -> delegate.addColumn(databaseName, tableName, columnName, columnType, columnComment)); } public synchronized void renameColumn(String databaseName, String tableName, String oldColumnName, String newColumnName) { setExclusive((delegate, hdfsEnvironment) -> delegate.renameColumn(databaseName, tableName, oldColumnName, newColumnName)); } public synchronized void dropColumn(String databaseName, String tableName, String columnName) { setExclusive((delegate, hdfsEnvironment) -> delegate.dropColumn(databaseName, tableName, columnName)); } public synchronized void finishInsertIntoExistingTable( ConnectorSession session, String databaseName, String tableName, Path currentLocation, List fileNames, PartitionStatistics statisticsUpdate) { // Data can only be inserted into partitions and unpartitioned tables. They can never be inserted into a partitioned table. // Therefore, this method assumes that the table is unpartitioned. setShared(); SchemaTableName schemaTableName = new SchemaTableName(databaseName, tableName); Action oldTableAction = tableActions.get(schemaTableName); if (oldTableAction == null || oldTableAction.getData().getTable().getTableType().equals(TEMPORARY_TABLE)) { Table table = getTable(databaseName, tableName) .orElseThrow(() -> new TableNotFoundException(schemaTableName)); PartitionStatistics currentStatistics = getTableStatistics(databaseName, tableName); HdfsContext context = new HdfsContext(session, databaseName, tableName); tableActions.put( schemaTableName, new Action<>( ActionType.INSERT_EXISTING, new TableAndMore( table, Optional.empty(), Optional.of(currentLocation), Optional.of(fileNames), false, merge(currentStatistics, statisticsUpdate), statisticsUpdate), context)); return; } switch (oldTableAction.getType()) { case DROP: throw new TableNotFoundException(schemaTableName); case ADD: case ALTER: case INSERT_EXISTING: throw new UnsupportedOperationException("Inserting into an unpartitioned table that were added, altered, or inserted into in the same transaction is not supported"); default: throw new IllegalStateException("Unknown action type"); } } public synchronized void truncateUnpartitionedTable(ConnectorSession session, String databaseName, String tableName) { checkReadable(); Optional
table = getTable(databaseName, tableName); SchemaTableName schemaTableName = new SchemaTableName(databaseName, tableName); if (!table.isPresent()) { throw new TableNotFoundException(schemaTableName); } if (!table.get().getTableType().equals(MANAGED_TABLE)) { throw new PrestoException(NOT_SUPPORTED, "Cannot delete from non-managed Hive table"); } if (!table.get().getPartitionColumns().isEmpty()) { throw new IllegalArgumentException("Table is partitioned"); } Path path = new Path(table.get().getStorage().getLocation()); HdfsContext context = new HdfsContext(session, databaseName, tableName); setExclusive((delegate, hdfsEnvironment) -> { RecursiveDeleteResult recursiveDeleteResult = recursiveDeleteFiles(hdfsEnvironment, context, path, ImmutableList.of(""), false); if (!recursiveDeleteResult.getNotDeletedEligibleItems().isEmpty()) { throw new PrestoException(HIVE_FILESYSTEM_ERROR, format( "Error deleting from unpartitioned table %s. These items can not be deleted: %s", schemaTableName, recursiveDeleteResult.getNotDeletedEligibleItems())); } }); } public synchronized Optional> getPartitionNames(String databaseName, String tableName) { return doGetPartitionNames(databaseName, tableName, ImmutableMap.of()); } public synchronized Optional> getPartitionNamesByFilter(String databaseName, String tableName, Map effectivePredicate) { return doGetPartitionNames(databaseName, tableName, effectivePredicate); } @GuardedBy("this") private Optional> doGetPartitionNames( String databaseName, String tableName, Map partitionPredicates) { checkHoldsLock(); checkReadable(); Optional
table = getTable(databaseName, tableName); if (!table.isPresent()) { return Optional.empty(); } List partitionNames; TableSource tableSource = getTableSource(databaseName, tableName); switch (tableSource) { case CREATED_IN_THIS_TRANSACTION: partitionNames = ImmutableList.of(); break; case PRE_EXISTING_TABLE: { Optional> partitionNameResult; List partitionColumns = table.get().getPartitionColumns(); if (!partitionPredicates.isEmpty()) { partitionNameResult = Optional.of(delegate.getPartitionNamesByFilter(databaseName, tableName, partitionPredicates)); } else { partitionNameResult = delegate.getPartitionNames(databaseName, tableName); } if (!partitionNameResult.isPresent()) { throw new PrestoException(TRANSACTION_CONFLICT, format("Table %s.%s was dropped by another transaction", databaseName, tableName)); } partitionNames = partitionNameResult.get(); break; } default: throw new UnsupportedOperationException("Unknown table source"); } Map, Action> partitionActionsOfTable = partitionActions.computeIfAbsent(new SchemaTableName(databaseName, tableName), k -> new HashMap<>()); ImmutableList.Builder resultBuilder = ImmutableList.builder(); // alter/remove newly-altered/dropped partitions from the results from underlying metastore for (String partitionName : partitionNames) { List partitionValues = toPartitionValues(partitionName); Action partitionAction = partitionActionsOfTable.get(partitionValues); if (partitionAction == null) { resultBuilder.add(partitionName); continue; } switch (partitionAction.getType()) { case ADD: throw new PrestoException(TRANSACTION_CONFLICT, format("Another transaction created partition %s in table %s.%s", partitionValues, databaseName, tableName)); case DROP: // do nothing break; case ALTER: case INSERT_EXISTING: resultBuilder.add(partitionName); break; default: throw new IllegalStateException("Unknown action type"); } } // add newly-added partitions to the results from underlying metastore if (!partitionActionsOfTable.isEmpty()) { List partitionColumns = table.get().getPartitionColumns(); List partitionColumnNames = partitionColumns.stream().map(Column::getName).collect(toList()); List parts = convertPredicateToParts(partitionPredicates); for (Action partitionAction : partitionActionsOfTable.values()) { if (partitionAction.getType() == ActionType.ADD) { List values = partitionAction.getData().getPartition().getValues(); if (parts.isEmpty() || partitionValuesMatch(values, parts)) { resultBuilder.add(makePartName(partitionColumnNames, values)); } } } } return Optional.of(resultBuilder.build()); } private static boolean partitionValuesMatch(List values, List pattern) { checkArgument(values.size() == pattern.size()); for (int i = 0; i < values.size(); i++) { if (pattern.get(i).isEmpty()) { // empty string match everything continue; } if (values.get(i).equals(pattern.get(i))) { return false; } } return true; } public synchronized Optional getPartition(String databaseName, String tableName, List partitionValues) { checkReadable(); TableSource tableSource = getTableSource(databaseName, tableName); Map, Action> partitionActionsOfTable = partitionActions.computeIfAbsent(new SchemaTableName(databaseName, tableName), k -> new HashMap<>()); Action partitionAction = partitionActionsOfTable.get(partitionValues); if (partitionAction != null) { return getPartitionFromPartitionAction(partitionAction); } switch (tableSource) { case PRE_EXISTING_TABLE: return delegate.getPartition(databaseName, tableName, partitionValues); case CREATED_IN_THIS_TRANSACTION: return Optional.empty(); default: throw new UnsupportedOperationException("unknown table source"); } } public synchronized Map> getPartitionsByNames(String databaseName, String tableName, List partitionNames) { checkReadable(); TableSource tableSource = getTableSource(databaseName, tableName); Map, Action> partitionActionsOfTable = partitionActions.computeIfAbsent(new SchemaTableName(databaseName, tableName), k -> new HashMap<>()); ImmutableList.Builder partitionNamesToQuery = ImmutableList.builder(); ImmutableMap.Builder> resultBuilder = ImmutableMap.builder(); for (String partitionName : partitionNames) { List partitionValues = toPartitionValues(partitionName); Action partitionAction = partitionActionsOfTable.get(partitionValues); if (partitionAction == null) { switch (tableSource) { case PRE_EXISTING_TABLE: partitionNamesToQuery.add(partitionName); break; case CREATED_IN_THIS_TRANSACTION: resultBuilder.put(partitionName, Optional.empty()); break; default: throw new UnsupportedOperationException("unknown table source"); } } else { resultBuilder.put(partitionName, getPartitionFromPartitionAction(partitionAction)); } } Map> delegateResult = delegate.getPartitionsByNames(databaseName, tableName, partitionNamesToQuery.build()); resultBuilder.putAll(delegateResult); return resultBuilder.build(); } private static Optional getPartitionFromPartitionAction(Action partitionAction) { switch (partitionAction.getType()) { case ADD: case ALTER: case INSERT_EXISTING: return Optional.of(partitionAction.getData().getAugmentedPartitionForInTransactionRead()); case DROP: return Optional.empty(); default: throw new IllegalStateException("Unknown action type"); } } public synchronized void addPartition( ConnectorSession session, String databaseName, String tableName, Partition partition, Path currentLocation, PartitionStatistics statistics) { setShared(); checkArgument(getPrestoQueryId(partition).isPresent()); Map, Action> partitionActionsOfTable = partitionActions.computeIfAbsent(new SchemaTableName(databaseName, tableName), k -> new HashMap<>()); Action oldPartitionAction = partitionActionsOfTable.get(partition.getValues()); HdfsContext context = new HdfsContext(session, databaseName, tableName); if (oldPartitionAction == null) { partitionActionsOfTable.put( partition.getValues(), new Action<>(ActionType.ADD, new PartitionAndMore(partition, currentLocation, Optional.empty(), statistics, statistics), context)); return; } switch (oldPartitionAction.getType()) { case DROP: { if (!oldPartitionAction.getContext().getIdentity().getUser().equals(session.getUser())) { throw new PrestoException(TRANSACTION_CONFLICT, "Operation on the same partition with different user in the same transaction is not supported"); } partitionActionsOfTable.put( partition.getValues(), new Action<>(ActionType.ALTER, new PartitionAndMore(partition, currentLocation, Optional.empty(), statistics, statistics), context)); break; } case ADD: case ALTER: case INSERT_EXISTING: throw new PrestoException(ALREADY_EXISTS, format("Partition already exists for table '%s.%s': %s", databaseName, tableName, partition.getValues())); default: throw new IllegalStateException("Unknown action type"); } } public synchronized void dropPartition(ConnectorSession session, String databaseName, String tableName, List partitionValues) { setShared(); Map, Action> partitionActionsOfTable = partitionActions.computeIfAbsent(new SchemaTableName(databaseName, tableName), k -> new HashMap<>()); Action oldPartitionAction = partitionActionsOfTable.get(partitionValues); if (oldPartitionAction == null) { HdfsContext context = new HdfsContext(session, databaseName, tableName); partitionActionsOfTable.put(partitionValues, new Action<>(ActionType.DROP, null, context)); return; } switch (oldPartitionAction.getType()) { case DROP: throw new PartitionNotFoundException(new SchemaTableName(databaseName, tableName), partitionValues); case ADD: case ALTER: case INSERT_EXISTING: throw new PrestoException( NOT_SUPPORTED, format("dropping a partition added in the same transaction is not supported: %s %s %s", databaseName, tableName, partitionValues)); default: throw new IllegalStateException("Unknown action type"); } } public synchronized void finishInsertIntoExistingPartition( ConnectorSession session, String databaseName, String tableName, List partitionValues, Path currentLocation, List fileNames, PartitionStatistics statisticsUpdate) { setShared(); SchemaTableName schemaTableName = new SchemaTableName(databaseName, tableName); Map, Action> partitionActionsOfTable = partitionActions.computeIfAbsent(schemaTableName, k -> new HashMap<>()); Action oldPartitionAction = partitionActionsOfTable.get(partitionValues); if (oldPartitionAction == null) { Partition partition = delegate.getPartition(databaseName, tableName, partitionValues) .orElseThrow(() -> new PartitionNotFoundException(schemaTableName, partitionValues)); String partitionName = getPartitionName(databaseName, tableName, partitionValues); PartitionStatistics currentStatistics = delegate.getPartitionStatistics(databaseName, tableName, ImmutableSet.of(partitionName)).get(partitionName); if (currentStatistics == null) { throw new PrestoException(HIVE_METASTORE_ERROR, "currentStatistics is null"); } HdfsContext context = new HdfsContext(session, databaseName, tableName); partitionActionsOfTable.put( partitionValues, new Action<>( ActionType.INSERT_EXISTING, new PartitionAndMore( partition, currentLocation, Optional.of(fileNames), merge(currentStatistics, statisticsUpdate), statisticsUpdate), context)); return; } switch (oldPartitionAction.getType()) { case DROP: throw new PartitionNotFoundException(schemaTableName, partitionValues); case ADD: case ALTER: case INSERT_EXISTING: throw new UnsupportedOperationException("Inserting into a partition that were added, altered, or inserted into in the same transaction is not supported"); default: throw new IllegalStateException("Unknown action type"); } } private String getPartitionName(String databaseName, String tableName, List partitionValues) { Table table = getTable(databaseName, tableName) .orElseThrow(() -> new TableNotFoundException(new SchemaTableName(databaseName, tableName))); return getPartitionName(table, partitionValues); } private String getPartitionName(Table table, List partitionValues) { List columnNames = table.getPartitionColumns().stream() .map(Column::getName) .collect(toImmutableList()); return makePartName(columnNames, partitionValues); } public synchronized void createRole(String role, String grantor) { setExclusive((delegate, hdfsEnvironment) -> delegate.createRole(role, grantor)); } public synchronized void dropRole(String role) { setExclusive((delegate, hdfsEnvironment) -> delegate.dropRole(role)); } public synchronized Set listRoles() { checkReadable(); return delegate.listRoles(); } public synchronized void grantRoles(Set roles, Set grantees, boolean withAdminOption, PrestoPrincipal grantor) { setExclusive((delegate, hdfsEnvironment) -> delegate.grantRoles(roles, grantees, withAdminOption, grantor)); } public synchronized void revokeRoles(Set roles, Set grantees, boolean adminOptionFor, PrestoPrincipal grantor) { setExclusive((delegate, hdfsEnvironment) -> delegate.revokeRoles(roles, grantees, adminOptionFor, grantor)); } public synchronized Set listRoleGrants(PrestoPrincipal principal) { checkReadable(); return delegate.listRoleGrants(principal); } public synchronized Set listTablePrivileges(String databaseName, String tableName, PrestoPrincipal principal) { checkReadable(); SchemaTableName schemaTableName = new SchemaTableName(databaseName, tableName); Action tableAction = tableActions.get(schemaTableName); if (tableAction == null) { return delegate.listTablePrivileges(databaseName, tableName, principal); } switch (tableAction.getType()) { case ADD: case ALTER: { if (principal.getType() == PrincipalType.ROLE) { return ImmutableSet.of(); } if (!principal.getName().equals(tableAction.getData().getTable().getOwner())) { return ImmutableSet.of(); } Collection privileges = tableAction.getData().getPrincipalPrivileges().getUserPrivileges().get(principal.getName()); return ImmutableSet.builder() .addAll(privileges) .add(new HivePrivilegeInfo(OWNERSHIP, true, new PrestoPrincipal(USER, principal.getName()), new PrestoPrincipal(USER, principal.getName()))) .build(); } case INSERT_EXISTING: return delegate.listTablePrivileges(databaseName, tableName, principal); case DROP: throw new TableNotFoundException(schemaTableName); default: throw new IllegalStateException("Unknown action type"); } } public synchronized void grantTablePrivileges(String databaseName, String tableName, PrestoPrincipal grantee, Set privileges) { setExclusive((delegate, hdfsEnvironment) -> delegate.grantTablePrivileges(databaseName, tableName, grantee, privileges)); } public synchronized void revokeTablePrivileges(String databaseName, String tableName, PrestoPrincipal grantee, Set privileges) { setExclusive((delegate, hdfsEnvironment) -> delegate.revokeTablePrivileges(databaseName, tableName, grantee, privileges)); } public synchronized void declareIntentionToWrite( ConnectorSession session, WriteMode writeMode, Path stagingPathRoot, Optional tempPathRoot, String filePrefix, SchemaTableName schemaTableName, boolean temporaryTable) { setShared(); if (writeMode == WriteMode.DIRECT_TO_TARGET_EXISTING_DIRECTORY) { Map, Action> partitionActionsOfTable = partitionActions.get(schemaTableName); if (partitionActionsOfTable != null && !partitionActionsOfTable.isEmpty()) { throw new PrestoException(NOT_SUPPORTED, "Can not insert into a table with a partition that has been modified in the same transaction when Presto is configured to skip temporary directories."); } } HdfsContext context = new HdfsContext(session, schemaTableName.getSchemaName(), schemaTableName.getTableName()); declaredIntentionsToWrite.add(new DeclaredIntentionToWrite(writeMode, context, stagingPathRoot, tempPathRoot, filePrefix, schemaTableName, temporaryTable)); } public synchronized void commit() { try { switch (state) { case EMPTY: break; case SHARED_OPERATION_BUFFERED: commitShared(); break; case EXCLUSIVE_OPERATION_BUFFERED: requireNonNull(bufferedExclusiveOperation, "bufferedExclusiveOperation is null"); bufferedExclusiveOperation.execute(delegate, hdfsEnvironment); break; case FINISHED: throw new IllegalStateException("Tried to commit buffered metastore operations after transaction has been committed/aborted"); default: throw new IllegalStateException("Unknown state"); } } finally { state = State.FINISHED; } } public synchronized void rollback() { try { switch (state) { case EMPTY: case EXCLUSIVE_OPERATION_BUFFERED: break; case SHARED_OPERATION_BUFFERED: rollbackShared(); break; case FINISHED: throw new IllegalStateException("Tried to rollback buffered metastore operations after transaction has been committed/aborted"); default: throw new IllegalStateException("Unknown state"); } } finally { state = State.FINISHED; } } @GuardedBy("this") private void commitShared() { checkHoldsLock(); Committer committer = new Committer(); try { for (Map.Entry> entry : tableActions.entrySet()) { SchemaTableName schemaTableName = entry.getKey(); Action action = entry.getValue(); switch (action.getType()) { case DROP: committer.prepareDropTable(schemaTableName); break; case ALTER: committer.prepareAlterTable(); break; case ADD: committer.prepareAddTable(action.getContext(), action.getData()); break; case INSERT_EXISTING: committer.prepareInsertExistingTable(action.getContext(), action.getData()); break; default: throw new IllegalStateException("Unknown action type"); } } for (Map.Entry, Action>> tableEntry : partitionActions.entrySet()) { SchemaTableName schemaTableName = tableEntry.getKey(); for (Map.Entry, Action> partitionEntry : tableEntry.getValue().entrySet()) { List partitionValues = partitionEntry.getKey(); Action action = partitionEntry.getValue(); switch (action.getType()) { case DROP: committer.prepareDropPartition(schemaTableName, partitionValues); break; case ALTER: committer.prepareAlterPartition(action.getContext(), action.getData()); break; case ADD: committer.prepareAddPartition(action.getContext(), action.getData()); break; case INSERT_EXISTING: committer.prepareInsertExistingPartition(action.getContext(), action.getData()); break; default: throw new IllegalStateException("Unknown action type"); } } } // Wait for all renames submitted for "INSERT_EXISTING" action to finish ListenableFuture listenableFutureAggregate = whenAllSucceed(committer.getFileRenameFutures()).call(() -> null, directExecutor()); try { getFutureValue(listenableFutureAggregate, PrestoException.class); } catch (RuntimeException e) { listenableFutureAggregate.cancel(true); throw e; } // At this point, all file system operations, whether asynchronously issued or not, have completed successfully. // We are moving on to metastore operations now. committer.executeAddTableOperations(); committer.executeAlterPartitionOperations(); committer.executeAddPartitionOperations(); committer.executeUpdateStatisticsOperations(); } catch (Throwable t) { committer.cancelUnstartedAsyncRenames(); committer.undoUpdateStatisticsOperations(); committer.undoAddPartitionOperations(); committer.undoAddTableOperations(); committer.waitForAsyncRenamesSuppressThrowables(); // fileRenameFutures must all come back before any file system cleanups are carried out. // Otherwise, files that should be deleted may be created after cleanup is done. committer.executeCleanupTasksForAbort(committer.extractFilePrefixes(declaredIntentionsToWrite)); committer.executeRenameTasksForAbort(); // Partition directory must be put back before relevant metastore operation can be undone committer.undoAlterPartitionOperations(); rollbackShared(); throw t; } try { // After this line, operations are no longer reversible. // The next section will deal with "dropping table/partition". Commit may still fail in // this section. Even if commit fails, cleanups, instead of rollbacks, will be executed. if (!committer.metastoreDeleteOperations.isEmpty()) { committer.executeMetastoreDeleteOperations(); } // If control flow reached this point, this commit is considered successful no matter // what happens later. The only kind of operations that haven't been carried out yet // are cleanups. // The program control flow will go to finally next. And cleanup will run because // moveForwardInFinally has been set to false. } finally { // In this method, all operations are best-effort clean up operations. // If any operation fails, the error will be logged and ignored. // Additionally, other clean up operations should still be attempted. // Execute deletion tasks committer.executeDeletionTasksForFinish(); // Clean up temporary tables deleteTemporaryTableDirectories(declaredIntentionsToWrite, hdfsEnvironment); // Clean up empty staging directories (that may recursively contain empty directories) committer.deleteEmptyStagingDirectories(declaredIntentionsToWrite); // Clean up root temp directories deleteTempPathRootDirectory(declaredIntentionsToWrite, hdfsEnvironment); } } private class Committer { private final AtomicBoolean fileRenameCancelled = new AtomicBoolean(false); private final List> fileRenameFutures = new ArrayList<>(); // File system // For file system changes, only operations outside of writing paths (as specified in declared intentions to write) // need to MOVE_BACKWARD tasks scheduled. Files in writing paths are handled by rollbackShared(). private final List deletionTasksForFinish = new ArrayList<>(); private final List cleanUpTasksForAbort = new ArrayList<>(); private final List renameTasksForAbort = new ArrayList<>(); // Metastore private final List addTableOperations = new ArrayList<>(); private final Map partitionAdders = new HashMap<>(); private final List alterPartitionOperations = new ArrayList<>(); private final List updateStatisticsOperations = new ArrayList<>(); private final List metastoreDeleteOperations = new ArrayList<>(); // Flag for better error message private boolean deleteOnly = true; private List> getFileRenameFutures() { return ImmutableList.copyOf(fileRenameFutures); } private void prepareDropTable(SchemaTableName schemaTableName) { metastoreDeleteOperations.add(new IrreversibleMetastoreOperation( format("drop table %s", schemaTableName), () -> delegate.dropTable(schemaTableName.getSchemaName(), schemaTableName.getTableName(), true))); } private void prepareAlterTable() { deleteOnly = false; // Currently, ALTER action is never constructed for tables. Dropping a table and then re-creating it // in the same transaction is not supported now. The following line should be replaced with actual // implementation when create after drop support is introduced for a table. throw new UnsupportedOperationException("Dropping and then creating a table with the same name is not supported"); } private void prepareAddTable(HdfsContext context, TableAndMore tableAndMore) { deleteOnly = false; Table table = tableAndMore.getTable(); if (table.getTableType().equals(TEMPORARY_TABLE)) { // do not commit a temporary table to the metastore return; } if (table.getTableType().equals(MANAGED_TABLE)) { String targetLocation = table.getStorage().getLocation(); checkArgument(!targetLocation.isEmpty(), "target location is empty"); Optional currentPath = tableAndMore.getCurrentLocation(); Path targetPath = new Path(targetLocation); if (table.getPartitionColumns().isEmpty() && currentPath.isPresent()) { // CREATE TABLE AS SELECT unpartitioned table if (targetPath.equals(currentPath.get())) { // Target path and current path are the same. Therefore, directory move is not needed. } else { renameDirectory( context, hdfsEnvironment, currentPath.get(), targetPath, () -> cleanUpTasksForAbort.add(new DirectoryCleanUpTask(context, targetPath, true))); } } else { // CREATE TABLE AS SELECT partitioned table, or // CREATE TABLE partitioned/unpartitioned table (without data) if (pathExists(context, hdfsEnvironment, targetPath)) { if (currentPath.isPresent() && currentPath.get().equals(targetPath)) { // It is okay to skip directory creation when currentPath is equal to targetPath // because the directory may have been created when creating partition directories. // However, it is important to note that the two being equal does not guarantee // a directory had been created. } else { throw new PrestoException( HIVE_PATH_ALREADY_EXISTS, format("Unable to create directory %s: target directory already exists", targetPath)); } } else { cleanUpTasksForAbort.add(new DirectoryCleanUpTask(context, targetPath, true)); createDirectory(context, hdfsEnvironment, targetPath); } } } addTableOperations.add(new CreateTableOperation(table, tableAndMore.getPrincipalPrivileges(), tableAndMore.isIgnoreExisting())); if (!isPrestoView(table)) { updateStatisticsOperations.add(new UpdateStatisticsOperation( new SchemaTableName(table.getDatabaseName(), table.getTableName()), Optional.empty(), tableAndMore.getStatisticsUpdate(), false)); } } private void prepareInsertExistingTable(HdfsContext context, TableAndMore tableAndMore) { deleteOnly = false; Table table = tableAndMore.getTable(); if (table.getTableType().equals(TEMPORARY_TABLE)) { // do not commit a temporary table to the metastore return; } Path targetPath = new Path(table.getStorage().getLocation()); Path currentPath = tableAndMore.getCurrentLocation().get(); cleanUpTasksForAbort.add(new DirectoryCleanUpTask(context, targetPath, false)); if (!targetPath.equals(currentPath)) { asyncRename(hdfsEnvironment, renameExecutor, fileRenameCancelled, fileRenameFutures, context, currentPath, targetPath, tableAndMore.getFileNames().get()); } updateStatisticsOperations.add(new UpdateStatisticsOperation( new SchemaTableName(table.getDatabaseName(), table.getTableName()), Optional.empty(), tableAndMore.getStatisticsUpdate(), true)); } private void prepareDropPartition(SchemaTableName schemaTableName, List partitionValues) { metastoreDeleteOperations.add(new IrreversibleMetastoreOperation( format("drop partition %s.%s %s", schemaTableName.getSchemaName(), schemaTableName.getTableName(), partitionValues), () -> delegate.dropPartition(schemaTableName.getSchemaName(), schemaTableName.getTableName(), partitionValues, true))); } private void prepareAlterPartition(HdfsContext context, PartitionAndMore partitionAndMore) { deleteOnly = false; Partition partition = partitionAndMore.getPartition(); String targetLocation = partition.getStorage().getLocation(); Optional oldPartition = delegate.getPartition(partition.getDatabaseName(), partition.getTableName(), partition.getValues()); if (!oldPartition.isPresent()) { throw new PrestoException( TRANSACTION_CONFLICT, format("The partition that this transaction modified was deleted in another transaction. %s %s", partition.getTableName(), partition.getValues())); } String partitionName = getPartitionName(partition.getDatabaseName(), partition.getTableName(), partition.getValues()); PartitionStatistics oldPartitionStatistics = getExistingPartitionStatistics(partition, partitionName); String oldPartitionLocation = oldPartition.get().getStorage().getLocation(); Path oldPartitionPath = new Path(oldPartitionLocation); // Location of the old partition and the new partition can be different because we allow arbitrary directories through LocationService. // If the location of the old partition is the same as the location of the new partition: // * Rename the old data directory to a temporary path with a special suffix // * Remember we will need to delete that directory at the end if transaction successfully commits // * Remember we will need to undo the rename if transaction aborts // Otherwise, // * Remember we will need to delete the location of the old partition at the end if transaction successfully commits if (targetLocation.equals(oldPartitionLocation)) { Path oldPartitionStagingPath = new Path(oldPartitionPath.getParent(), "_temp_" + oldPartitionPath.getName() + "_" + context.getQueryId()); renameDirectory( context, hdfsEnvironment, oldPartitionPath, oldPartitionStagingPath, () -> renameTasksForAbort.add(new DirectoryRenameTask(context, oldPartitionStagingPath, oldPartitionPath))); if (!skipDeletionForAlter) { deletionTasksForFinish.add(new DirectoryDeletionTask(context, oldPartitionStagingPath)); } } else { if (!skipDeletionForAlter) { deletionTasksForFinish.add(new DirectoryDeletionTask(context, oldPartitionPath)); } } Path currentPath = partitionAndMore.getCurrentLocation(); Path targetPath = new Path(targetLocation); if (!targetPath.equals(currentPath)) { renameDirectory( context, hdfsEnvironment, currentPath, targetPath, () -> cleanUpTasksForAbort.add(new DirectoryCleanUpTask(context, targetPath, true))); } // Partition alter must happen regardless of whether original and current location is the same // because metadata might change: e.g. storage format, column types, etc alterPartitionOperations.add(new AlterPartitionOperation( new PartitionWithStatistics(partition, partitionName, partitionAndMore.getStatisticsUpdate()), new PartitionWithStatistics(oldPartition.get(), partitionName, oldPartitionStatistics))); } private PartitionStatistics getExistingPartitionStatistics(Partition partition, String partitionName) { try { PartitionStatistics statistics = delegate.getPartitionStatistics(partition.getDatabaseName(), partition.getTableName(), ImmutableSet.of(partitionName)) .get(partitionName); if (statistics == null) { throw new PrestoException( TRANSACTION_CONFLICT, format("The partition that this transaction modified was deleted in another transaction. %s %s", partition.getTableName(), partition.getValues())); } return statistics; } catch (PrestoException e) { if (e.getErrorCode().equals(HIVE_CORRUPTED_COLUMN_STATISTICS.toErrorCode())) { log.warn( e, "Corrupted statistics found when altering partition. Table: %s.%s. Partition: %s", partition.getDatabaseName(), partition.getTableName(), partition.getValues()); return PartitionStatistics.empty(); } throw e; } } private void prepareAddPartition(HdfsContext context, PartitionAndMore partitionAndMore) { deleteOnly = false; Partition partition = partitionAndMore.getPartition(); String targetLocation = partition.getStorage().getLocation(); Path currentPath = partitionAndMore.getCurrentLocation(); Path targetPath = new Path(targetLocation); SchemaTableName schemaTableName = new SchemaTableName(partition.getDatabaseName(), partition.getTableName()); PartitionAdder partitionAdder = partitionAdders.computeIfAbsent( schemaTableName, ignored -> new PartitionAdder(partition.getDatabaseName(), partition.getTableName(), delegate, PARTITION_COMMIT_BATCH_SIZE)); if (pathExists(context, hdfsEnvironment, currentPath)) { if (!targetPath.equals(currentPath)) { renameDirectory( context, hdfsEnvironment, currentPath, targetPath, () -> cleanUpTasksForAbort.add(new DirectoryCleanUpTask(context, targetPath, true))); } } else { cleanUpTasksForAbort.add(new DirectoryCleanUpTask(context, targetPath, true)); createDirectory(context, hdfsEnvironment, targetPath); } String partitionName = getPartitionName(partition.getDatabaseName(), partition.getTableName(), partition.getValues()); partitionAdder.addPartition(new PartitionWithStatistics(partition, partitionName, partitionAndMore.getStatisticsUpdate())); } private void prepareInsertExistingPartition(HdfsContext context, PartitionAndMore partitionAndMore) { deleteOnly = false; Partition partition = partitionAndMore.getPartition(); Path targetPath = new Path(partition.getStorage().getLocation()); Path currentPath = partitionAndMore.getCurrentLocation(); cleanUpTasksForAbort.add(new DirectoryCleanUpTask(context, targetPath, false)); if (!targetPath.equals(currentPath)) { asyncRename(hdfsEnvironment, renameExecutor, fileRenameCancelled, fileRenameFutures, context, currentPath, targetPath, partitionAndMore.getFileNames()); } updateStatisticsOperations.add(new UpdateStatisticsOperation( new SchemaTableName(partition.getDatabaseName(), partition.getTableName()), Optional.of(getPartitionName(partition.getDatabaseName(), partition.getTableName(), partition.getValues())), partitionAndMore.getStatisticsUpdate(), true)); } private void executeCleanupTasksForAbort(List filePrefixes) { for (DirectoryCleanUpTask cleanUpTask : cleanUpTasksForAbort) { recursiveDeleteFilesAndLog(cleanUpTask.getContext(), cleanUpTask.getPath(), filePrefixes, cleanUpTask.isDeleteEmptyDirectory(), "temporary directory commit abort"); } } private void executeDeletionTasksForFinish() { for (DirectoryDeletionTask deletionTask : deletionTasksForFinish) { if (!deleteRecursivelyIfExists(deletionTask.getContext(), hdfsEnvironment, deletionTask.getPath())) { logCleanupFailure("Error deleting directory %s", deletionTask.getPath().toString()); } } } private void executeRenameTasksForAbort() { for (DirectoryRenameTask directoryRenameTask : renameTasksForAbort) { try { // Ignore the task if the source directory doesn't exist. // This is probably because the original rename that we are trying to undo here never succeeded. if (pathExists(directoryRenameTask.getContext(), hdfsEnvironment, directoryRenameTask.getRenameFrom())) { renameDirectory(directoryRenameTask.getContext(), hdfsEnvironment, directoryRenameTask.getRenameFrom(), directoryRenameTask.getRenameTo(), () -> {}); } } catch (Throwable throwable) { logCleanupFailure(throwable, "failed to undo rename of partition directory: %s to %s", directoryRenameTask.getRenameFrom(), directoryRenameTask.getRenameTo()); } } } private void deleteEmptyStagingDirectories(List declaredIntentionsToWrite) { for (DeclaredIntentionToWrite declaredIntentionToWrite : declaredIntentionsToWrite) { if (declaredIntentionToWrite.getMode() != WriteMode.STAGE_AND_MOVE_TO_TARGET_DIRECTORY) { continue; } Path path = declaredIntentionToWrite.getStagingPathRoot(); recursiveDeleteFilesAndLog(declaredIntentionToWrite.getContext(), path, ImmutableList.of(), true, "staging directory cleanup"); } } private void waitForAsyncRenamesSuppressThrowables() { for (ListenableFuture future : fileRenameFutures) { try { future.get(); } catch (InterruptedException e) { Thread.currentThread().interrupt(); } catch (Throwable t) { // ignore } } } private void cancelUnstartedAsyncRenames() { fileRenameCancelled.set(true); } private void executeAddTableOperations() { for (CreateTableOperation addTableOperation : addTableOperations) { addTableOperation.run(delegate); } } private void executeAlterPartitionOperations() { for (AlterPartitionOperation alterPartitionOperation : alterPartitionOperations) { alterPartitionOperation.run(delegate); } } private void executeAddPartitionOperations() { for (PartitionAdder partitionAdder : partitionAdders.values()) { partitionAdder.execute(); } } private void executeUpdateStatisticsOperations() { for (UpdateStatisticsOperation operation : updateStatisticsOperations) { operation.run(delegate); } } private void undoAddPartitionOperations() { for (PartitionAdder partitionAdder : partitionAdders.values()) { List> partitionsFailedToRollback = partitionAdder.rollback(); if (!partitionsFailedToRollback.isEmpty()) { logCleanupFailure("Failed to rollback: add_partition for partitions %s.%s %s", partitionAdder.getSchemaName(), partitionAdder.getTableName(), partitionsFailedToRollback.stream()); } } } private void undoAddTableOperations() { for (CreateTableOperation addTableOperation : addTableOperations) { try { addTableOperation.undo(delegate); } catch (Throwable throwable) { logCleanupFailure(throwable, "failed to rollback: %s", addTableOperation.getDescription()); } } } private void undoAlterPartitionOperations() { for (AlterPartitionOperation alterPartitionOperation : alterPartitionOperations) { try { alterPartitionOperation.undo(delegate); } catch (Throwable throwable) { logCleanupFailure(throwable, "failed to rollback: %s", alterPartitionOperation.getDescription()); } } } private void undoUpdateStatisticsOperations() { for (UpdateStatisticsOperation operation : updateStatisticsOperations) { try { operation.undo(delegate); } catch (Throwable throwable) { logCleanupFailure(throwable, "failed to rollback: %s", operation.getDescription()); } } } private void executeMetastoreDeleteOperations() { List failedDeletionDescriptions = new ArrayList<>(); List suppressedExceptions = new ArrayList<>(); boolean anySucceeded = false; for (IrreversibleMetastoreOperation deleteOperation : metastoreDeleteOperations) { try { deleteOperation.run(); anySucceeded = true; } catch (Throwable t) { if (metastoreDeleteOperations.size() == 1 && t instanceof TableNotFoundException) { throw new PrestoException(HIVE_TABLE_DROPPED_DURING_QUERY, "The metastore delete operation failed: " + deleteOperation.getDescription()); } failedDeletionDescriptions.add(deleteOperation.getDescription()); // A limit is needed to avoid having a huge exception object. 5 was chosen arbitrarily. if (suppressedExceptions.size() < 5) { suppressedExceptions.add(t); } } } if (!suppressedExceptions.isEmpty()) { StringBuilder message = new StringBuilder(); if (deleteOnly && !anySucceeded) { message.append("The following metastore delete operations failed: "); } else { message.append("The transaction didn't commit cleanly. All operations other than the following delete operations were completed: "); } Joiner.on("; ").appendTo(message, failedDeletionDescriptions); PrestoException prestoException = new PrestoException(HIVE_METASTORE_ERROR, message.toString()); suppressedExceptions.forEach(prestoException::addSuppressed); throw prestoException; } } private List extractFilePrefixes(List declaredIntentionsToWrite) { Set filePrefixSet = new HashSet<>(); for (DeclaredIntentionToWrite declaredIntentionToWrite : declaredIntentionsToWrite) { filePrefixSet.add(declaredIntentionToWrite.getFilePrefix()); } return ImmutableList.copyOf(filePrefixSet); } } @GuardedBy("this") private void rollbackShared() { checkHoldsLock(); deleteTemporaryTableDirectories(declaredIntentionsToWrite, hdfsEnvironment); deleteTempPathRootDirectory(declaredIntentionsToWrite, hdfsEnvironment); for (DeclaredIntentionToWrite declaredIntentionToWrite : declaredIntentionsToWrite) { switch (declaredIntentionToWrite.getMode()) { case STAGE_AND_MOVE_TO_TARGET_DIRECTORY: case DIRECT_TO_TARGET_NEW_DIRECTORY: { if (skipTargetCleanupOnRollback && declaredIntentionToWrite.getMode() == DIRECT_TO_TARGET_NEW_DIRECTORY) { break; } // Note: For STAGE_AND_MOVE_TO_TARGET_DIRECTORY there is no need to cleanup the target directory as it will only be written // to during the commit call and the commit call cleans up after failures. Path rootPath = declaredIntentionToWrite.getStagingPathRoot(); // In the case of DIRECT_TO_TARGET_NEW_DIRECTORY, if the directory is not guaranteed to be unique // for the query, it is possible that another query or compute engine may see the directory, wrote // data to it, and exported it through metastore. Therefore it may be argued that cleanup of staging // directories must be carried out conservatively. To be safe, we only delete files that start with // the unique prefix for queries in this transaction. recursiveDeleteFilesAndLog( declaredIntentionToWrite.getContext(), rootPath, ImmutableList.of(declaredIntentionToWrite.getFilePrefix()), true, format("staging/target_new directory rollback for table %s", declaredIntentionToWrite.getSchemaTableName())); break; } case DIRECT_TO_TARGET_EXISTING_DIRECTORY: { Set pathsToClean = new HashSet<>(); // Check the base directory of the declared intention // * existing partition may also be in this directory // * this is where new partitions are created Path baseDirectory = declaredIntentionToWrite.getStagingPathRoot(); pathsToClean.add(baseDirectory); SchemaTableName schemaTableName = declaredIntentionToWrite.getSchemaTableName(); Optional
table = delegate.getTable(schemaTableName.getSchemaName(), schemaTableName.getTableName()); if (table.isPresent()) { // check every existing partition that is outside for the base directory if (!table.get().getPartitionColumns().isEmpty()) { List partitionNames = delegate.getPartitionNames(schemaTableName.getSchemaName(), schemaTableName.getTableName()) .orElse(ImmutableList.of()); for (List partitionNameBatch : Iterables.partition(partitionNames, 10)) { Collection> partitions = delegate.getPartitionsByNames(schemaTableName.getSchemaName(), schemaTableName.getTableName(), partitionNameBatch).values(); partitions.stream() .filter(Optional::isPresent) .map(Optional::get) .map(partition -> partition.getStorage().getLocation()) .map(Path::new) .filter(path -> !isSameOrParent(baseDirectory, path)) .forEach(pathsToClean::add); } } } else { logCleanupFailure( "Error rolling back write to table %s.%s. Data directory may contain temporary data. Table was dropped in another transaction.", schemaTableName.getSchemaName(), schemaTableName.getTableName()); } // delete any file that starts with the unique prefix of this query for (Path path : pathsToClean) { // TODO: It is a known deficiency that some empty directory does not get cleaned up in S3. // We can not delete any of the directories here since we do not know who created them. recursiveDeleteFilesAndLog( declaredIntentionToWrite.getContext(), path, ImmutableList.of(declaredIntentionToWrite.getFilePrefix()), false, format("target_existing directory rollback for table %s", schemaTableName)); } break; } default: throw new UnsupportedOperationException("Unknown write mode"); } } } private static void deleteTemporaryTableDirectories(List declaredIntentionsToWrite, HdfsEnvironment hdfsEnvironment) { for (DeclaredIntentionToWrite declaredIntentionToWrite : declaredIntentionsToWrite) { if (declaredIntentionToWrite.isTemporaryTable()) { deleteRecursivelyIfExists(declaredIntentionToWrite.getContext(), hdfsEnvironment, declaredIntentionToWrite.getStagingPathRoot()); } } } private static void deleteTempPathRootDirectory(List declaredIntentionsToWrite, HdfsEnvironment hdfsEnvironment) { for (DeclaredIntentionToWrite declaredIntentionToWrite : declaredIntentionsToWrite) { if (declaredIntentionToWrite.getTempPathRoot().isPresent()) { deleteRecursivelyIfExists(declaredIntentionToWrite.getContext(), hdfsEnvironment, declaredIntentionToWrite.getTempPathRoot().get()); } } } @VisibleForTesting public synchronized void testOnlyCheckIsReadOnly() { if (state != State.EMPTY) { throw new AssertionError("Test did not commit or rollback"); } } @VisibleForTesting public void testOnlyThrowOnCleanupFailures() { throwOnCleanupFailure = true; } @GuardedBy("this") private void checkReadable() { checkHoldsLock(); switch (state) { case EMPTY: case SHARED_OPERATION_BUFFERED: return; case EXCLUSIVE_OPERATION_BUFFERED: throw new PrestoException(NOT_SUPPORTED, "Unsupported combination of operations in a single transaction"); case FINISHED: throw new IllegalStateException("Tried to access metastore after transaction has been committed/aborted"); } } @GuardedBy("this") private void setShared() { checkHoldsLock(); checkReadable(); state = State.SHARED_OPERATION_BUFFERED; } @GuardedBy("this") private void setExclusive(ExclusiveOperation exclusiveOperation) { checkHoldsLock(); if (state != State.EMPTY) { throw new PrestoException(StandardErrorCode.NOT_SUPPORTED, "Unsupported combination of operations in a single transaction"); } state = State.EXCLUSIVE_OPERATION_BUFFERED; bufferedExclusiveOperation = exclusiveOperation; } @GuardedBy("this") private void checkNoPartitionAction(String databaseName, String tableName) { checkHoldsLock(); Map, Action> partitionActionsOfTable = partitionActions.get(new SchemaTableName(databaseName, tableName)); if (partitionActionsOfTable != null && !partitionActionsOfTable.isEmpty()) { throw new PrestoException(NOT_SUPPORTED, "Cannot make schema changes to a table/view with modified partitions in the same transaction"); } } private static boolean isSameOrParent(Path parent, Path child) { int parentDepth = parent.depth(); int childDepth = child.depth(); if (parentDepth > childDepth) { return false; } for (int i = childDepth; i > parentDepth; i--) { child = child.getParent(); } return parent.equals(child); } private void logCleanupFailure(String format, Object... args) { if (throwOnCleanupFailure) { throw new RuntimeException(format(format, args)); } log.warn(format, args); } private void logCleanupFailure(Throwable t, String format, Object... args) { if (throwOnCleanupFailure) { throw new RuntimeException(format(format, args), t); } log.warn(t, format, args); } private static void asyncRename( HdfsEnvironment hdfsEnvironment, ListeningExecutorService executor, AtomicBoolean cancelled, List> fileRenameFutures, HdfsContext context, Path currentPath, Path targetPath, List fileNames) { FileSystem fileSystem = getFileSystem(hdfsEnvironment, context, currentPath); for (String fileName : fileNames) { Path source = new Path(currentPath, fileName); Path target = new Path(targetPath, fileName); fileRenameFutures.add(executor.submit(() -> { if (cancelled.get()) { return; } renameFile(fileSystem, source, target); })); } } private void recursiveDeleteFilesAndLog(HdfsContext context, Path directory, List filePrefixes, boolean deleteEmptyDirectories, String reason) { RecursiveDeleteResult recursiveDeleteResult = recursiveDeleteFiles( hdfsEnvironment, context, directory, filePrefixes, deleteEmptyDirectories); if (!recursiveDeleteResult.getNotDeletedEligibleItems().isEmpty()) { logCleanupFailure( "Error deleting directory %s for %s. Some eligible items can not be deleted: %s.", directory.toString(), reason, recursiveDeleteResult.getNotDeletedEligibleItems()); } else if (deleteEmptyDirectories && !recursiveDeleteResult.isDirectoryNoLongerExists()) { logCleanupFailure( "Error deleting directory %s for %s. Can not delete the directory.", directory.toString(), reason); } } /** * Attempt to recursively remove eligible files and/or directories in {@code directory}. *

* When {@code filePrefixes} is not present, all files (but not necessarily directories) will be * ineligible. If all files shall be deleted, you can use an empty string as {@code filePrefixes}. *

* When {@code deleteEmptySubDirectory} is true, any empty directory (including directories that * were originally empty, and directories that become empty after files prefixed with * {@code filePrefixes} are deleted) will be eligible. *

* This method will not delete anything that's neither a directory nor a file. * * @param filePrefixes prefix of files that should be deleted * @param deleteEmptyDirectories whether empty directories should be deleted */ private static RecursiveDeleteResult recursiveDeleteFiles(HdfsEnvironment hdfsEnvironment, HdfsContext context, Path directory, List filePrefixes, boolean deleteEmptyDirectories) { FileSystem fileSystem; try { fileSystem = hdfsEnvironment.getFileSystem(context, directory); if (!fileSystem.exists(directory)) { return new RecursiveDeleteResult(true, ImmutableList.of()); } } catch (IOException e) { ImmutableList.Builder notDeletedItems = ImmutableList.builder(); notDeletedItems.add(directory.toString() + "/**"); return new RecursiveDeleteResult(false, notDeletedItems.build()); } return doRecursiveDeleteFiles(fileSystem, directory, filePrefixes, deleteEmptyDirectories); } private static RecursiveDeleteResult doRecursiveDeleteFiles(FileSystem fileSystem, Path directory, List filePrefixes, boolean deleteEmptyDirectories) { // don't delete hidden presto directories if (directory.getName().startsWith(".presto")) { return new RecursiveDeleteResult(false, ImmutableList.of()); } FileStatus[] allFiles; try { allFiles = fileSystem.listStatus(directory); } catch (IOException e) { ImmutableList.Builder notDeletedItems = ImmutableList.builder(); notDeletedItems.add(directory.toString() + "/**"); return new RecursiveDeleteResult(false, notDeletedItems.build()); } boolean allDescendentsDeleted = true; ImmutableList.Builder notDeletedEligibleItems = ImmutableList.builder(); for (FileStatus fileStatus : allFiles) { if (fileStatus.isFile()) { Path filePath = fileStatus.getPath(); String fileName = filePath.getName(); boolean eligible = false; // never delete presto dot files if (!fileName.startsWith(".presto")) { // file name that starts with ".tmp.presto" is staging file, see HiveWriterFactory#createWriter. eligible = filePrefixes.stream().anyMatch(prefix -> fileName.startsWith(prefix) || fileName.startsWith(".tmp.presto." + prefix)); } if (eligible) { if (!deleteIfExists(fileSystem, filePath, false)) { allDescendentsDeleted = false; notDeletedEligibleItems.add(filePath.toString()); } } else { allDescendentsDeleted = false; } } else if (fileStatus.isDirectory()) { RecursiveDeleteResult subResult = doRecursiveDeleteFiles(fileSystem, fileStatus.getPath(), filePrefixes, deleteEmptyDirectories); if (!subResult.isDirectoryNoLongerExists()) { allDescendentsDeleted = false; } if (!subResult.getNotDeletedEligibleItems().isEmpty()) { notDeletedEligibleItems.addAll(subResult.getNotDeletedEligibleItems()); } } else { allDescendentsDeleted = false; notDeletedEligibleItems.add(fileStatus.getPath().toString()); } } if (allDescendentsDeleted && deleteEmptyDirectories) { verify(notDeletedEligibleItems.build().isEmpty()); if (!deleteIfExists(fileSystem, directory, false)) { return new RecursiveDeleteResult(false, ImmutableList.of(directory.toString() + "/")); } return new RecursiveDeleteResult(true, ImmutableList.of()); } return new RecursiveDeleteResult(false, notDeletedEligibleItems.build()); } /** * Attempts to remove the file or empty directory. * * @return true if the location no longer exists */ private static boolean deleteIfExists(FileSystem fileSystem, Path path, boolean recursive) { try { // attempt to delete the path if (fileSystem.delete(path, recursive)) { return true; } // delete failed // check if path still exists return !fileSystem.exists(path); } catch (FileNotFoundException ignored) { // path was already removed or never existed return true; } catch (IOException ignored) { } return false; } /** * Attempts to remove the file or empty directory. * * @return true if the location no longer exists */ private static boolean deleteRecursivelyIfExists(HdfsContext context, HdfsEnvironment hdfsEnvironment, Path path) { FileSystem fileSystem; try { fileSystem = hdfsEnvironment.getFileSystem(context, path); } catch (IOException ignored) { return false; } return deleteIfExists(fileSystem, path, true); } private static void renameDirectory(HdfsContext context, HdfsEnvironment hdfsEnvironment, Path source, Path target, Runnable runWhenPathDoesntExist) { if (pathExists(context, hdfsEnvironment, target)) { throw new PrestoException(HIVE_PATH_ALREADY_EXISTS, format("Unable to rename from %s to %s: target directory already exists", source, target)); } if (!pathExists(context, hdfsEnvironment, target.getParent())) { createDirectory(context, hdfsEnvironment, target.getParent()); } // The runnable will assume that if rename fails, it will be okay to delete the directory (if the directory is empty). // This is not technically true because a race condition still exists. runWhenPathDoesntExist.run(); try { if (!hdfsEnvironment.getFileSystem(context, source).rename(source, target)) { throw new PrestoException(HIVE_FILESYSTEM_ERROR, format("Failed to rename %s to %s: rename returned false", source, target)); } } catch (IOException e) { throw new PrestoException(HIVE_FILESYSTEM_ERROR, format("Failed to rename %s to %s", source, target), e); } } private static Optional getPrestoQueryId(Table table) { return Optional.ofNullable(table.getParameters().get(PRESTO_QUERY_ID_NAME)); } private static Optional getPrestoQueryId(Partition partition) { return Optional.ofNullable(partition.getParameters().get(PRESTO_QUERY_ID_NAME)); } private void checkHoldsLock() { // This method serves a similar purpose at runtime as GuardedBy on method serves during static analysis. // This method should not have significant performance impact. If it does, it may be reasonably to remove this method. // This intentionally does not use checkState. if (!Thread.holdsLock(this)) { throw new IllegalStateException(format("Thread must hold a lock on the %s", getClass().getSimpleName())); } } private enum State { EMPTY, SHARED_OPERATION_BUFFERED, EXCLUSIVE_OPERATION_BUFFERED, FINISHED, } private enum ActionType { DROP, ADD, ALTER, INSERT_EXISTING } private enum TableSource { CREATED_IN_THIS_TRANSACTION, PRE_EXISTING_TABLE, // RECREATED_IN_THIS_TRANSACTION is a possible case, but it is not supported with the current implementation } public static class Action { private final ActionType type; private final T data; private final HdfsContext context; public Action(ActionType type, T data, HdfsContext context) { this.type = requireNonNull(type, "type is null"); if (type == ActionType.DROP) { checkArgument(data == null, "data is not null"); } else { requireNonNull(data, "data is null"); } this.data = data; this.context = requireNonNull(context, "context is null"); } public ActionType getType() { return type; } public T getData() { checkState(type != ActionType.DROP); return data; } public HdfsContext getContext() { return context; } @Override public String toString() { return toStringHelper(this) .add("type", type) .add("data", data) .toString(); } } private static class TableAndMore { private final Table table; private final Optional principalPrivileges; private final Optional currentLocation; // unpartitioned table only private final Optional> fileNames; private final boolean ignoreExisting; private final PartitionStatistics statistics; private final PartitionStatistics statisticsUpdate; public TableAndMore( Table table, Optional principalPrivileges, Optional currentLocation, Optional> fileNames, boolean ignoreExisting, PartitionStatistics statistics, PartitionStatistics statisticsUpdate) { this.table = requireNonNull(table, "table is null"); this.principalPrivileges = requireNonNull(principalPrivileges, "principalPrivileges is null"); this.currentLocation = requireNonNull(currentLocation, "currentLocation is null"); this.fileNames = requireNonNull(fileNames, "fileNames is null"); this.ignoreExisting = ignoreExisting; this.statistics = requireNonNull(statistics, "statistics is null"); this.statisticsUpdate = requireNonNull(statisticsUpdate, "statisticsUpdate is null"); checkArgument(!table.getTableType().equals(VIRTUAL_VIEW) || !currentLocation.isPresent(), "currentLocation can not be supplied for view"); checkArgument(!fileNames.isPresent() || currentLocation.isPresent(), "fileNames can be supplied only when currentLocation is supplied"); } public boolean isIgnoreExisting() { return ignoreExisting; } public Table getTable() { return table; } public PrincipalPrivileges getPrincipalPrivileges() { checkState(principalPrivileges.isPresent()); return principalPrivileges.get(); } public Optional getCurrentLocation() { return currentLocation; } public Optional> getFileNames() { return fileNames; } public PartitionStatistics getStatistics() { return statistics; } public PartitionStatistics getStatisticsUpdate() { return statisticsUpdate; } public Table getAugmentedTableForInTransactionRead() { // Don't augment the location for partitioned tables, // as data is never read directly from the partitioned table location if (!table.getPartitionColumns().isEmpty()) { return table; } // view don't have currentLocation if (!currentLocation.isPresent()) { return table; } // For unpartitioned table, this method augments the location field of the table // to the staging location. // This way, if the table is accessed in an ongoing transaction, staged data // can be found and accessed. String currentLocation = this.currentLocation.get().toString(); if (!currentLocation.equals(table.getStorage().getLocation())) { return Table.builder(table) .withStorage(storage -> storage.setLocation(currentLocation)) .build(); } return table; } @Override public String toString() { return toStringHelper(this) .add("table", table) .add("principalPrivileges", principalPrivileges) .add("currentLocation", currentLocation) .add("fileNames", fileNames) .add("ignoreExisting", ignoreExisting) .add("statistics", statistics) .add("statisticsUpdate", statisticsUpdate) .toString(); } } private static class PartitionAndMore { private final Partition partition; private final Path currentLocation; private final Optional> fileNames; private final PartitionStatistics statistics; private final PartitionStatistics statisticsUpdate; public PartitionAndMore(Partition partition, Path currentLocation, Optional> fileNames, PartitionStatistics statistics, PartitionStatistics statisticsUpdate) { this.partition = requireNonNull(partition, "partition is null"); this.currentLocation = requireNonNull(currentLocation, "currentLocation is null"); this.fileNames = requireNonNull(fileNames, "fileNames is null"); this.statistics = requireNonNull(statistics, "statistics is null"); this.statisticsUpdate = requireNonNull(statisticsUpdate, "statisticsUpdate is null"); } public Partition getPartition() { return partition; } public Path getCurrentLocation() { return currentLocation; } public List getFileNames() { checkState(fileNames.isPresent()); return fileNames.get(); } public PartitionStatistics getStatistics() { return statistics; } public PartitionStatistics getStatisticsUpdate() { return statisticsUpdate; } public Partition getAugmentedPartitionForInTransactionRead() { // This method augments the location field of the partition to the staging location. // This way, if the partition is accessed in an ongoing transaction, staged data // can be found and accessed. String currentLocation = this.currentLocation.toString(); if (!currentLocation.equals(partition.getStorage().getLocation())) { return Partition.builder(partition) .withStorage(storage -> storage.setLocation(currentLocation)) .build(); } return partition; } @Override public String toString() { return toStringHelper(this) .add("partition", partition) .add("currentLocation", currentLocation) .add("fileNames", fileNames) .toString(); } } private static class DeclaredIntentionToWrite { private final WriteMode mode; private final HdfsContext context; private final String filePrefix; private final Path stagingPathRoot; private final Optional tempPathRoot; private final SchemaTableName schemaTableName; private final boolean temporaryTable; public DeclaredIntentionToWrite( WriteMode mode, HdfsContext context, Path stagingPathRoot, Optional tempPathRoot, String filePrefix, SchemaTableName schemaTableName, boolean temporaryTable) { this.mode = requireNonNull(mode, "mode is null"); this.context = requireNonNull(context, "context is null"); this.stagingPathRoot = requireNonNull(stagingPathRoot, "stagingPathRoot is null"); this.tempPathRoot = requireNonNull(tempPathRoot, "tempPathRoot is null"); this.filePrefix = requireNonNull(filePrefix, "filePrefix is null"); this.schemaTableName = requireNonNull(schemaTableName, "schemaTableName is null"); this.temporaryTable = temporaryTable; } public WriteMode getMode() { return mode; } public HdfsContext getContext() { return context; } public String getFilePrefix() { return filePrefix; } public Path getStagingPathRoot() { return stagingPathRoot; } public Optional getTempPathRoot() { return tempPathRoot; } public SchemaTableName getSchemaTableName() { return schemaTableName; } public boolean isTemporaryTable() { return temporaryTable; } @Override public String toString() { return toStringHelper(this) .add("mode", mode) .add("context", context) .add("filePrefix", filePrefix) .add("stagingPathRoot", stagingPathRoot) .add("tempPathRoot", tempPathRoot) .add("schemaTableName", schemaTableName) .add("temporaryTable", temporaryTable) .toString(); } } private static class DirectoryCleanUpTask { private final HdfsContext context; private final Path path; private final boolean deleteEmptyDirectory; public DirectoryCleanUpTask(HdfsContext context, Path path, boolean deleteEmptyDirectory) { this.context = context; this.path = path; this.deleteEmptyDirectory = deleteEmptyDirectory; } public HdfsContext getContext() { return context; } public Path getPath() { return path; } public boolean isDeleteEmptyDirectory() { return deleteEmptyDirectory; } @Override public String toString() { return toStringHelper(this) .add("context", context) .add("path", path) .add("deleteEmptyDirectory", deleteEmptyDirectory) .toString(); } } private static class DirectoryDeletionTask { private final HdfsContext context; private final Path path; public DirectoryDeletionTask(HdfsContext context, Path path) { this.context = context; this.path = path; } public HdfsContext getContext() { return context; } public Path getPath() { return path; } @Override public String toString() { return toStringHelper(this) .add("context", context) .add("path", path) .toString(); } } private static class DirectoryRenameTask { private final HdfsContext context; private final Path renameFrom; private final Path renameTo; public DirectoryRenameTask(HdfsContext context, Path renameFrom, Path renameTo) { this.context = requireNonNull(context, "context is null"); this.renameFrom = requireNonNull(renameFrom, "renameFrom is null"); this.renameTo = requireNonNull(renameTo, "renameTo is null"); } public HdfsContext getContext() { return context; } public Path getRenameFrom() { return renameFrom; } public Path getRenameTo() { return renameTo; } @Override public String toString() { return toStringHelper(this) .add("context", context) .add("renameFrom", renameFrom) .add("renameTo", renameTo) .toString(); } } private static class IrreversibleMetastoreOperation { private final String description; private final Runnable action; public IrreversibleMetastoreOperation(String description, Runnable action) { this.description = requireNonNull(description, "description is null"); this.action = requireNonNull(action, "action is null"); } public String getDescription() { return description; } public void run() { action.run(); } } private static class CreateTableOperation { private final Table newTable; private final PrincipalPrivileges privileges; private boolean tableCreated; private final boolean ignoreExisting; private final String queryId; public CreateTableOperation(Table newTable, PrincipalPrivileges privileges, boolean ignoreExisting) { requireNonNull(newTable, "newTable is null"); this.newTable = newTable; this.privileges = requireNonNull(privileges, "privileges is null"); this.ignoreExisting = ignoreExisting; this.queryId = getPrestoQueryId(newTable).orElseThrow(() -> new IllegalArgumentException("Query id is not present")); } public String getDescription() { return format("add table %s.%s", newTable.getDatabaseName(), newTable.getTableName()); } public void run(ExtendedHiveMetastore metastore) { boolean done = false; try { metastore.createTable(newTable, privileges); done = true; } catch (RuntimeException e) { try { Optional

existingTable = metastore.getTable(newTable.getDatabaseName(), newTable.getTableName()); if (existingTable.isPresent()) { Table table = existingTable.get(); Optional existingTableQueryId = getPrestoQueryId(table); if (existingTableQueryId.isPresent() && existingTableQueryId.get().equals(queryId)) { // ignore table if it was already created by the same query during retries done = true; } else { // If the table definition in the metastore is different than what this tx wants to create // then there is a conflict (e.g., current tx wants to create T(a: bigint), // but another tx already created T(a: varchar)). // This may be a problem if there is an insert after this step. if (!hasTheSameSchema(newTable, table)) { e = new PrestoException(TRANSACTION_CONFLICT, format("Table already exists with a different schema: '%s'", newTable.getTableName())); } else { done = ignoreExisting; } } } } catch (RuntimeException ignored) { // When table could not be fetched from metastore, it is not known whether the table was added. // Deleting the table when aborting commit has the risk of deleting table not added in this transaction. // Not deleting the table may leave garbage behind. The former is much more dangerous than the latter. // Therefore, the table is not considered added. } if (!done) { throw e; } } tableCreated = true; } private boolean hasTheSameSchema(Table newTable, Table existingTable) { List newTableColumns = newTable.getDataColumns(); List existingTableColumns = existingTable.getDataColumns(); if (newTableColumns.size() != existingTableColumns.size()) { return false; } for (Column existingColumn : existingTableColumns) { if (newTableColumns.stream() .noneMatch(newColumn -> newColumn.getName().equals(existingColumn.getName()) && newColumn.getType().equals(existingColumn.getType()))) { return false; } } return true; } public void undo(ExtendedHiveMetastore metastore) { if (!tableCreated) { return; } metastore.dropTable(newTable.getDatabaseName(), newTable.getTableName(), false); } } private static class AlterPartitionOperation { private final PartitionWithStatistics newPartition; private final PartitionWithStatistics oldPartition; private boolean undo; public AlterPartitionOperation(PartitionWithStatistics newPartition, PartitionWithStatistics oldPartition) { this.newPartition = requireNonNull(newPartition, "newPartition is null"); this.oldPartition = requireNonNull(oldPartition, "oldPartition is null"); checkArgument(newPartition.getPartition().getDatabaseName().equals(oldPartition.getPartition().getDatabaseName())); checkArgument(newPartition.getPartition().getTableName().equals(oldPartition.getPartition().getTableName())); checkArgument(newPartition.getPartition().getValues().equals(oldPartition.getPartition().getValues())); } public String getDescription() { return format( "alter partition %s.%s %s", newPartition.getPartition().getDatabaseName(), newPartition.getPartition().getTableName(), newPartition.getPartition().getValues()); } public void run(ExtendedHiveMetastore metastore) { undo = true; metastore.alterPartition(newPartition.getPartition().getDatabaseName(), newPartition.getPartition().getTableName(), newPartition); } public void undo(ExtendedHiveMetastore metastore) { if (!undo) { return; } metastore.alterPartition(oldPartition.getPartition().getDatabaseName(), oldPartition.getPartition().getTableName(), oldPartition); } } private static class UpdateStatisticsOperation { private final SchemaTableName tableName; private final Optional partitionName; private final PartitionStatistics statistics; private final boolean merge; private boolean done; public UpdateStatisticsOperation(SchemaTableName tableName, Optional partitionName, PartitionStatistics statistics, boolean merge) { this.tableName = requireNonNull(tableName, "tableName is null"); this.partitionName = requireNonNull(partitionName, "partitionValues is null"); this.statistics = requireNonNull(statistics, "statistics is null"); this.merge = merge; } public void run(ExtendedHiveMetastore metastore) { if (partitionName.isPresent()) { metastore.updatePartitionStatistics(tableName.getSchemaName(), tableName.getTableName(), partitionName.get(), this::updateStatistics); } else { metastore.updateTableStatistics(tableName.getSchemaName(), tableName.getTableName(), this::updateStatistics); } done = true; } public void undo(ExtendedHiveMetastore metastore) { if (!done) { return; } if (partitionName.isPresent()) { metastore.updatePartitionStatistics(tableName.getSchemaName(), tableName.getTableName(), partitionName.get(), this::resetStatistics); } else { metastore.updateTableStatistics(tableName.getSchemaName(), tableName.getTableName(), this::resetStatistics); } } public String getDescription() { if (partitionName.isPresent()) { return format("replace partition parameters %s %s", tableName, partitionName.get()); } return format("replace table parameters %s", tableName); } private PartitionStatistics updateStatistics(PartitionStatistics currentStatistics) { return merge ? merge(currentStatistics, statistics) : statistics; } private PartitionStatistics resetStatistics(PartitionStatistics currentStatistics) { return new PartitionStatistics(reduce(currentStatistics.getBasicStatistics(), statistics.getBasicStatistics(), SUBTRACT), ImmutableMap.of()); } } private static class PartitionAdder { private final String schemaName; private final String tableName; private final ExtendedHiveMetastore metastore; private final int batchSize; private final List partitions; private List> createdPartitionValues = new ArrayList<>(); public PartitionAdder(String schemaName, String tableName, ExtendedHiveMetastore metastore, int batchSize) { this.schemaName = schemaName; this.tableName = tableName; this.metastore = metastore; this.batchSize = batchSize; this.partitions = new ArrayList<>(batchSize); } public String getSchemaName() { return schemaName; } public String getTableName() { return tableName; } public void addPartition(PartitionWithStatistics partition) { checkArgument(getPrestoQueryId(partition.getPartition()).isPresent()); partitions.add(partition); } public void execute() { List> batchedPartitions = Lists.partition(partitions, batchSize); for (List batch : batchedPartitions) { try { metastore.addPartitions(schemaName, tableName, batch); for (PartitionWithStatistics partition : batch) { createdPartitionValues.add(partition.getPartition().getValues()); } } catch (Throwable t) { // Add partition to the created list conservatively. // Some metastore implementations are known to violate the "all or none" guarantee for add_partitions call. boolean batchCompletelyAdded = true; for (PartitionWithStatistics partition : batch) { try { Optional remotePartition = metastore.getPartition(schemaName, tableName, partition.getPartition().getValues()); // getPrestoQueryId(partition) is guaranteed to be non-empty. It is asserted in PartitionAdder.addPartition. if (remotePartition.isPresent() && getPrestoQueryId(remotePartition.get()).equals(getPrestoQueryId(partition.getPartition()))) { createdPartitionValues.add(partition.getPartition().getValues()); } else { batchCompletelyAdded = false; } } catch (Throwable ignored) { // When partition could not be fetched from metastore, it is not known whether the partition was added. // Deleting the partition when aborting commit has the risk of deleting partition not added in this transaction. // Not deleting the partition may leave garbage behind. The former is much more dangerous than the latter. // Therefore, the partition is not added to the createdPartitionValues list here. batchCompletelyAdded = false; } } // If all the partitions were added successfully, the add_partition operation was actually successful. // For some reason, it threw an exception (communication failure, retry failure after communication failure, etc). // But we would consider it successful anyways. if (!batchCompletelyAdded) { if (t instanceof TableNotFoundException) { throw new PrestoException(HIVE_TABLE_DROPPED_DURING_QUERY, t); } throw t; } } } partitions.clear(); } public List> rollback() { // drop created partitions List> partitionsFailedToRollback = new ArrayList<>(); for (List createdPartitionValue : createdPartitionValues) { try { metastore.dropPartition(schemaName, tableName, createdPartitionValue, false); } catch (PartitionNotFoundException e) { // Maybe some one deleted the partition we added. // Anyways, we are good because the partition is not there anymore. } catch (Throwable t) { partitionsFailedToRollback.add(createdPartitionValue); } } createdPartitionValues = partitionsFailedToRollback; return partitionsFailedToRollback; } } private static class RecursiveDeleteResult { private final boolean directoryNoLongerExists; private final List notDeletedEligibleItems; public RecursiveDeleteResult(boolean directoryNoLongerExists, List notDeletedEligibleItems) { this.directoryNoLongerExists = directoryNoLongerExists; this.notDeletedEligibleItems = notDeletedEligibleItems; } public boolean isDirectoryNoLongerExists() { return directoryNoLongerExists; } public List getNotDeletedEligibleItems() { return notDeletedEligibleItems; } } private interface ExclusiveOperation { void execute(ExtendedHiveMetastore delegate, HdfsEnvironment hdfsEnvironment); } }