Examples with DeltaLakeStatistics io.trino.plugin.deltalake.statistics.DeltaLakeStatistics used on opensource projects

Example 1 with DeltaLakeStatistics

use of io.trino.plugin.deltalake.statistics.DeltaLakeStatistics in project trino by trinodb.

the class HiveMetastoreBackedDeltaLakeMetastore method getTableStatistics.

@Override
public TableStatistics getTableStatistics(ConnectorSession session, DeltaLakeTableHandle tableHandle, Constraint constraint) {
    TableSnapshot tableSnapshot = getSnapshot(tableHandle.getSchemaTableName(), session);
    double numRecords = 0L;
    MetadataEntry metadata = transactionLogAccess.getMetadataEntry(tableSnapshot, session).orElseThrow(() -> new TrinoException(DELTA_LAKE_INVALID_SCHEMA, "Metadata not found in transaction log for " + tableHandle.getTableName()));
    List<ColumnMetadata> columnMetadata = DeltaLakeSchemaSupport.extractSchema(metadata, typeManager);
    List<DeltaLakeColumnHandle> columns = columnMetadata.stream().map(columnMeta -> new DeltaLakeColumnHandle(columnMeta.getName(), columnMeta.getType(), metadata.getCanonicalPartitionColumns().contains(columnMeta.getName()) ? PARTITION_KEY : REGULAR)).collect(toImmutableList());
    Map<DeltaLakeColumnHandle, Double> nullCounts = new HashMap<>();
    columns.forEach(column -> nullCounts.put(column, 0.0));
    Map<DeltaLakeColumnHandle, Double> minValues = new HashMap<>();
    Map<DeltaLakeColumnHandle, Double> maxValues = new HashMap<>();
    Map<DeltaLakeColumnHandle, Set<String>> partitioningColumnsDistinctValues = new HashMap<>();
    columns.stream().filter(column -> column.getColumnType() == PARTITION_KEY).forEach(column -> partitioningColumnsDistinctValues.put(column, new HashSet<>()));
    if (tableHandle.getEnforcedPartitionConstraint().isNone() || tableHandle.getNonPartitionConstraint().isNone() || constraint.getSummary().isNone()) {
        return createZeroStatistics(columns);
    }
    Set<String> predicatedColumnNames = tableHandle.getNonPartitionConstraint().getDomains().orElseThrow().keySet().stream().map(DeltaLakeColumnHandle::getName).collect(toImmutableSet());
    List<ColumnMetadata> predicatedColumns = columnMetadata.stream().filter(column -> predicatedColumnNames.contains(column.getName())).collect(toImmutableList());
    for (AddFileEntry addEntry : transactionLogAccess.getActiveFiles(tableSnapshot, session)) {
        Optional<? extends DeltaLakeFileStatistics> fileStatistics = addEntry.getStats();
        if (fileStatistics.isEmpty()) {
            // Open source Delta Lake does not collect stats
            return TableStatistics.empty();
        }
        DeltaLakeFileStatistics stats = fileStatistics.get();
        if (!partitionMatchesPredicate(addEntry.getCanonicalPartitionValues(), tableHandle.getEnforcedPartitionConstraint().getDomains().orElseThrow())) {
            continue;
        }
        TupleDomain<DeltaLakeColumnHandle> statisticsPredicate = createStatisticsPredicate(addEntry, predicatedColumns, tableHandle.getMetadataEntry().getCanonicalPartitionColumns());
        if (!tableHandle.getNonPartitionConstraint().overlaps(statisticsPredicate)) {
            continue;
        }
        if (stats.getNumRecords().isEmpty()) {
            // Not clear if it's possible for stats to be present with no row count, but bail out if that happens
            return TableStatistics.empty();
        }
        numRecords += stats.getNumRecords().get();
        for (DeltaLakeColumnHandle column : columns) {
            if (column.getColumnType() == PARTITION_KEY) {
                Optional<String> partitionValue = addEntry.getCanonicalPartitionValues().get(column.getName());
                if (partitionValue.isEmpty()) {
                    nullCounts.merge(column, (double) stats.getNumRecords().get(), Double::sum);
                } else {
                    // NULL is not counted as a distinct value
                    // Code below assumes that values returned by addEntry.getCanonicalPartitionValues() are normalized,
                    // it may not be true in case of real, doubles, timestamps etc
                    partitioningColumnsDistinctValues.get(column).add(partitionValue.get());
                }
            } else {
                Optional<Long> maybeNullCount = stats.getNullCount(column.getName());
                if (maybeNullCount.isPresent()) {
                    nullCounts.put(column, nullCounts.get(column) + maybeNullCount.get());
                } else {
                    // If any individual file fails to report null counts, fail to calculate the total for the table
                    nullCounts.put(column, NaN);
                }
            }
            // Math.min returns NaN if any operand is NaN
            stats.getMinColumnValue(column).map(parsedValue -> toStatsRepresentation(column.getType(), parsedValue)).filter(OptionalDouble::isPresent).map(OptionalDouble::getAsDouble).ifPresent(parsedValueAsDouble -> minValues.merge(column, parsedValueAsDouble, Math::min));
            stats.getMaxColumnValue(column).map(parsedValue -> toStatsRepresentation(column.getType(), parsedValue)).filter(OptionalDouble::isPresent).map(OptionalDouble::getAsDouble).ifPresent(parsedValueAsDouble -> maxValues.merge(column, parsedValueAsDouble, Math::max));
        }
    }
    if (numRecords == 0) {
        return createZeroStatistics(columns);
    }
    TableStatistics.Builder statsBuilder = new TableStatistics.Builder().setRowCount(Estimate.of(numRecords));
    Optional<DeltaLakeStatistics> statistics = Optional.empty();
    if (isExtendedStatisticsEnabled(session)) {
        statistics = statisticsAccess.readDeltaLakeStatistics(session, tableHandle.getLocation());
    }
    for (DeltaLakeColumnHandle column : columns) {
        ColumnStatistics.Builder columnStatsBuilder = new ColumnStatistics.Builder();
        Double nullCount = nullCounts.get(column);
        columnStatsBuilder.setNullsFraction(nullCount.isNaN() ? Estimate.unknown() : Estimate.of(nullCount / numRecords));
        Double maxValue = maxValues.get(column);
        Double minValue = minValues.get(column);
        if (isValidInRange(maxValue) && isValidInRange(minValue)) {
            columnStatsBuilder.setRange(new DoubleRange(minValue, maxValue));
        } else if (isValidInRange(maxValue)) {
            columnStatsBuilder.setRange(new DoubleRange(NEGATIVE_INFINITY, maxValue));
        } else if (isValidInRange(minValue)) {
            columnStatsBuilder.setRange(new DoubleRange(minValue, POSITIVE_INFINITY));
        }
        // extend statistics with NDV
        if (column.getColumnType() == PARTITION_KEY) {
            columnStatsBuilder.setDistinctValuesCount(Estimate.of(partitioningColumnsDistinctValues.get(column).size()));
        }
        if (statistics.isPresent()) {
            DeltaLakeColumnStatistics deltaLakeColumnStatistics = statistics.get().getColumnStatistics().get(column.getName());
            if (deltaLakeColumnStatistics != null && column.getColumnType() != PARTITION_KEY) {
                columnStatsBuilder.setDistinctValuesCount(Estimate.of(deltaLakeColumnStatistics.getNdvSummary().cardinality()));
            }
        }
        statsBuilder.setColumnStatistics(column, columnStatsBuilder.build());
    }
    return statsBuilder.build();
}

Example 2 with DeltaLakeStatistics

use of io.trino.plugin.deltalake.statistics.DeltaLakeStatistics in project trino by trinodb.

the class DeltaLakeMetadata method getTableHandleForStatisticsCollection.

@Nullable
@Override
public ConnectorTableHandle getTableHandleForStatisticsCollection(ConnectorSession session, SchemaTableName tableName, Map<String, Object> analyzeProperties) {
    Optional<Table> table = metastore.getTable(tableName.getSchemaName(), tableName.getTableName());
    if (table.isEmpty()) {
        return null;
    }
    if (!isExtendedStatisticsEnabled(session)) {
        throw new TrinoException(NOT_SUPPORTED, "ANALYZE not supported if extended statistics are disabled. Enable via delta.extended-statistics.enabled config property or extended_statistics_enabled session property.");
    }
    Optional<Instant> filesModifiedAfterFromProperties = DeltaLakeAnalyzeProperties.getFilesModifiedAfterProperty(analyzeProperties);
    TableSnapshot tableSnapshot = metastore.getSnapshot(tableName, session);
    long version = tableSnapshot.getVersion();
    String tableLocation = metastore.getTableLocation(tableName, session);
    Optional<DeltaLakeStatistics> statistics = statisticsAccess.readDeltaLakeStatistics(session, tableLocation);
    Optional<Instant> alreadyAnalyzedModifiedTimeMax = statistics.map(DeltaLakeStatistics::getAlreadyAnalyzedModifiedTimeMax);
    // determine list of files we want to read based on what caller requested via files_modified_after and what files were already analyzed in the past
    Optional<Instant> filesModifiedAfter = Optional.empty();
    if (filesModifiedAfterFromProperties.isPresent() || alreadyAnalyzedModifiedTimeMax.isPresent()) {
        filesModifiedAfter = Optional.of(Comparators.max(filesModifiedAfterFromProperties.orElse(Instant.ofEpochMilli(0)), alreadyAnalyzedModifiedTimeMax.orElse(Instant.ofEpochMilli(0))));
    }
    MetadataEntry metadata = metastore.getMetadata(tableSnapshot, session).orElseThrow(() -> new TrinoException(DELTA_LAKE_INVALID_SCHEMA, "Metadata not found in transaction log for " + table));
    Optional<Set<String>> analyzeColumnNames = DeltaLakeAnalyzeProperties.getColumnNames(analyzeProperties);
    if (analyzeColumnNames.isPresent()) {
        Set<String> columnNames = analyzeColumnNames.get();
        // validate that proper column names are passed via `columns` analyze property
        if (columnNames.isEmpty()) {
            throw new TrinoException(INVALID_ANALYZE_PROPERTY, "Cannot specify empty list of columns for analysis");
        }
        Set<String> allColumnNames = extractSchema(metadata, typeManager).stream().map(ColumnMetadata::getName).collect(toImmutableSet());
        if (!allColumnNames.containsAll(columnNames)) {
            throw new TrinoException(INVALID_ANALYZE_PROPERTY, format("Invalid columns specified for analysis: %s", Sets.difference(columnNames, allColumnNames)));
        }
    }
    // verify that we do not extend set of analyzed columns
    Optional<Set<String>> oldAnalyzeColumnNames = statistics.flatMap(DeltaLakeStatistics::getAnalyzedColumns);
    if (oldAnalyzeColumnNames.isPresent()) {
        if (analyzeColumnNames.isEmpty() || !oldAnalyzeColumnNames.get().containsAll(analyzeColumnNames.get())) {
            throw new TrinoException(INVALID_ANALYZE_PROPERTY, "List of columns to be analyzed must be a subset of previously used. To extend list of analyzed columns drop table statistics");
        }
    }
    AnalyzeHandle analyzeHandle = new AnalyzeHandle(version, statistics.isEmpty(), filesModifiedAfter, analyzeColumnNames);
    return new DeltaLakeTableHandle(tableName.getSchemaName(), tableName.getTableName(), tableLocation, Optional.of(metadata), TupleDomain.all(), TupleDomain.all(), Optional.empty(), Optional.empty(), Optional.empty(), Optional.empty(), Optional.of(analyzeHandle), version);
}

Example 3 with DeltaLakeStatistics

use of io.trino.plugin.deltalake.statistics.DeltaLakeStatistics in project trino by trinodb.

the class DeltaLakeMetadata method finishStatisticsCollection.

@Override
public void finishStatisticsCollection(ConnectorSession session, ConnectorTableHandle table, Collection<ComputedStatistics> computedStatistics) {
    DeltaLakeTableHandle tableHandle = (DeltaLakeTableHandle) table;
    AnalyzeHandle analyzeHandle = tableHandle.getAnalyzeHandle().orElseThrow(() -> new IllegalArgumentException("analyzeHandle not set"));
    String location = metastore.getTableLocation(tableHandle.getSchemaTableName(), session);
    Optional<DeltaLakeStatistics> oldStatistics = statisticsAccess.readDeltaLakeStatistics(session, location);
    // more elaborate logic for handling statistics model evaluation may need to be introduced in the future
    // for now let's have a simple check rejecting update
    oldStatistics.ifPresent(statistics -> checkArgument(statistics.getModelVersion() == DeltaLakeStatistics.CURRENT_MODEL_VERSION, "Existing table statistics are incompatible, run the drop statistics procedure on this table before re-analyzing"));
    Map<String, DeltaLakeColumnStatistics> oldColumnStatistics = oldStatistics.map(DeltaLakeStatistics::getColumnStatistics).orElseGet(ImmutableMap::of);
    Map<String, DeltaLakeColumnStatistics> newColumnStatistics = toDeltaLakeColumnStatistics(computedStatistics);
    Map<String, DeltaLakeColumnStatistics> mergedColumnStatistics = new HashMap<>();
    // only keep stats for existing columns
    Set<String> newColumns = newColumnStatistics.keySet();
    oldColumnStatistics.entrySet().stream().filter(entry -> newColumns.contains(entry.getKey())).forEach(entry -> mergedColumnStatistics.put(entry.getKey(), entry.getValue()));
    newColumnStatistics.forEach((columnName, columnStatistics) -> {
        mergedColumnStatistics.merge(columnName, columnStatistics, DeltaLakeColumnStatistics::update);
    });
    Optional<Instant> maxFileModificationTime = getMaxFileModificationTime(computedStatistics);
    // We do not want to hinder our future calls to ANALYZE if one of the files we analyzed have modification time far in the future.
    // Therefore we cap the value we store in extended_stats.json to current_time as observed on Trino coordinator.
    Instant finalAlreadyAnalyzedModifiedTimeMax = Instant.now();
    if (maxFileModificationTime.isPresent()) {
        finalAlreadyAnalyzedModifiedTimeMax = Comparators.min(maxFileModificationTime.get(), finalAlreadyAnalyzedModifiedTimeMax);
    }
    // also ensure that we are not traveling back in time
    if (oldStatistics.isPresent()) {
        finalAlreadyAnalyzedModifiedTimeMax = Comparators.max(oldStatistics.get().getAlreadyAnalyzedModifiedTimeMax(), finalAlreadyAnalyzedModifiedTimeMax);
    }
    if (analyzeHandle.getColumns().isPresent() && !mergedColumnStatistics.keySet().equals(analyzeHandle.getColumns().get())) {
        // sanity validation
        throw new IllegalStateException(format("Unexpected columns in in mergedColumnStatistics %s; expected %s", mergedColumnStatistics.keySet(), analyzeHandle.getColumns().get()));
    }
    DeltaLakeStatistics mergedDeltaLakeStatistics = new DeltaLakeStatistics(finalAlreadyAnalyzedModifiedTimeMax, mergedColumnStatistics, analyzeHandle.getColumns());
    statisticsAccess.updateDeltaLakeStatistics(session, location, mergedDeltaLakeStatistics);
}