Merge pull request #1945 from kuzudb/null-read-opt

Skip reading null data if the property is known to have no nulls

src/common/null_mask.cpp

@@ -96,6 +96,13 @@ bool NullMask::copyFromNullBits(const uint64_t* srcNullEntries, uint64_t srcOffs
     return false;
 }
 
+void NullMask::setNullFromRange(uint64_t offset, uint64_t numBitsToSet, bool isNull) {
+    if (isNull) {
+        this->mayContainNulls = true;
+    }
+    setNullRange(data, offset, numBitsToSet, isNull);
+}
+
 void NullMask::setNullRange(
     uint64_t* nullEntries, uint64_t offset, uint64_t numBitsToSet, bool isNull) {
     auto [firstEntryPos, firstBitPos] = getNullEntryAndBitPos(offset);

src/include/common/null_mask.h

@@ -139,6 +139,8 @@ class NullMask {
     static void setNullRange(
         uint64_t* nullEntries, uint64_t offset, uint64_t numBitsToSet, bool isNull);
 
+    void setNullFromRange(uint64_t offset, uint64_t numBitsToSet, bool isNull);
+
     void resize(uint64_t capacity);
 
 private:

src/include/common/vector/value_vector.h

@@ -36,9 +36,7 @@ class ValueVector {
     // On return true, there are no null. On return false, there may or may not be nulls.
     inline bool hasNoNullsGuarantee() const { return nullMask->hasNoNullsGuarantee(); }
     inline void setRangeNonNull(uint32_t startPos, uint32_t len) {
-        for (auto i = 0u; i < len; ++i) {
-            setNull(startPos + i, false);
-        }
+        nullMask->setNullFromRange(startPos, len, false);
     }
     inline const uint64_t* getNullMaskData() { return nullMask->getData(); }
     inline void setNull(uint32_t pos, bool isNull) { nullMask->setNull(pos, isNull); }

src/include/storage/copier/column_chunk.h

@@ -185,7 +185,7 @@ class BoolColumnChunk : public ColumnChunk {
         arrow::Array* array, common::offset_t startPosInChunk, uint32_t numValuesToAppend) final;
 
     void append(ColumnChunk* other, common::offset_t startPosInOtherChunk,
-        common::offset_t startPosInChunk, uint32_t numValuesToAppend) final;
+        common::offset_t startPosInChunk, uint32_t numValuesToAppend) override;
 
     void resize(uint64_t capacity) final;
 
@@ -200,12 +200,38 @@ class BoolColumnChunk : public ColumnChunk {
 
 class NullColumnChunk : public BoolColumnChunk {
 public:
-    NullColumnChunk() : BoolColumnChunk(nullptr /*copyDescription*/, false /*hasNullChunk*/) {}
+    NullColumnChunk()
+        : BoolColumnChunk(nullptr /*copyDescription*/, false /*hasNullChunk*/), mayHaveNullValue{
+                                                                                    false} {}
     // Maybe this should be combined with BoolColumnChunk if the only difference is these functions?
     inline bool isNull(common::offset_t pos) const { return getValue<bool>(pos); }
-    inline void setNull(common::offset_t pos, bool isNull) { setValue(isNull, pos); }
+    inline void setNull(common::offset_t pos, bool isNull) {
+        setValue(isNull, pos);
+        if (isNull) {
+            mayHaveNullValue = true;
+        }
+    }
+
+    inline bool mayHaveNull() const { return mayHaveNullValue; }
 
-    inline void resetNullBuffer() { memset(buffer.get(), 0 /* non null */, bufferSize); }
+    inline void resetNullBuffer() {
+        memset(buffer.get(), 0 /* non null */, bufferSize);
+        mayHaveNullValue = false;
+    }
+
+    inline void copyFromBuffer(uint64_t* srcBuffer, uint64_t srcOffset, uint64_t dstOffset,
+        uint64_t numBits, bool invert = false) {
+        if (common::NullMask::copyNullMask(
+                srcBuffer, srcOffset, (uint64_t*)buffer.get(), dstOffset, numBits, invert)) {
+            mayHaveNullValue = true;
+        }
+    }
+
+    void append(ColumnChunk* other, common::offset_t startPosInOtherChunk,
+        common::offset_t startPosInChunk, uint32_t numValuesToAppend) final;
+
+protected:
+    bool mayHaveNullValue;
 };
 
 class FixedListColumnChunk : public ColumnChunk {

src/include/storage/store/node_column.h

@@ -4,6 +4,7 @@
 #include "storage/copier/column_chunk.h"
 #include "storage/storage_structure/disk_array.h"
 #include "storage/storage_structure/storage_structure.h"
+#include "storage/store/property_statistics.h"
 
 namespace kuzu {
 namespace transaction {
@@ -57,10 +58,11 @@ class NodeColumn {
 public:
     NodeColumn(const catalog::Property& property, BMFileHandle* dataFH, BMFileHandle* metadataFH,
         BufferManager* bufferManager, WAL* wal, transaction::Transaction* transaction,
-        bool requireNullColumn = true);
+        RWPropertyStats propertyStatistics, bool requireNullColumn = true);
     NodeColumn(common::LogicalType dataType, const catalog::MetadataDAHInfo& metaDAHeaderInfo,
         BMFileHandle* dataFH, BMFileHandle* metadataFH, BufferManager* bufferManager, WAL* wal,
-        transaction::Transaction* transaction, bool requireNullColumn);
+        transaction::Transaction* transaction, RWPropertyStats PropertyStatistics,
+        bool requireNullColumn);
     virtual ~NodeColumn() = default;
 
     // Expose for feature store
@@ -140,13 +142,15 @@ class NodeColumn {
     std::vector<std::unique_ptr<NodeColumn>> childrenColumns;
     read_node_column_func_t readNodeColumnFunc;
     write_node_column_func_t writeNodeColumnFunc;
+    RWPropertyStats propertyStatistics;
 };
 
 class BoolNodeColumn : public NodeColumn {
 public:
     BoolNodeColumn(const catalog::MetadataDAHInfo& metaDAHeaderInfo, BMFileHandle* dataFH,
         BMFileHandle* metadataFH, BufferManager* bufferManager, WAL* wal,
-        transaction::Transaction* transaction, bool requireNullColumn = true);
+        transaction::Transaction* transaction, RWPropertyStats propertyStatistics,
+        bool requireNullColumn = true);
 
     void batchLookup(transaction::Transaction* transaction, const common::offset_t* nodeOffsets,
         size_t size, uint8_t* result) final;
@@ -158,10 +162,15 @@ class NullNodeColumn : public NodeColumn {
 public:
     NullNodeColumn(common::page_idx_t metaDAHPageIdx, BMFileHandle* dataFH,
         BMFileHandle* metadataFH, BufferManager* bufferManager, WAL* wal,
-        transaction::Transaction* transaction);
+        transaction::Transaction* transaction, RWPropertyStats propertyStatistics);
 
     void scan(transaction::Transaction* transaction, common::ValueVector* nodeIDVector,
         common::ValueVector* resultVector) final;
+    void scan(transaction::Transaction* transaction, common::node_group_idx_t nodeGroupIdx,
+        common::offset_t startOffsetInGroup, common::offset_t endOffsetInGroup,
+        common::ValueVector* resultVector, uint64_t offsetInVector = 0) final;
+    void scan(common::node_group_idx_t nodeGroupIdx, ColumnChunk* columnChunk) final;
+
     void lookup(transaction::Transaction* transaction, common::ValueVector* nodeIDVector,
         common::ValueVector* resultVector) final;
     common::page_idx_t append(
@@ -190,14 +199,14 @@ class SerialNodeColumn : public NodeColumn {
 struct NodeColumnFactory {
     static inline std::unique_ptr<NodeColumn> createNodeColumn(const catalog::Property& property,
         BMFileHandle* dataFH, BMFileHandle* metadataFH, BufferManager* bufferManager, WAL* wal,
-        transaction::Transaction* transaction) {
+        transaction::Transaction* transaction, RWPropertyStats propertyStatistics) {
         return createNodeColumn(*property.getDataType(), *property.getMetadataDAHInfo(), dataFH,
-            metadataFH, bufferManager, wal, transaction);
+            metadataFH, bufferManager, wal, transaction, propertyStatistics);
     }
     static std::unique_ptr<NodeColumn> createNodeColumn(const common::LogicalType& dataType,
         const catalog::MetadataDAHInfo& metaDAHeaderInfo, BMFileHandle* dataFH,
         BMFileHandle* metadataFH, BufferManager* bufferManager, WAL* wal,
-        transaction::Transaction* transaction);
+        transaction::Transaction* transaction, RWPropertyStats propertyStatistics);
 };
 
 } // namespace storage

src/include/storage/store/nodes_statistics_and_deleted_ids.h

@@ -11,19 +11,22 @@ namespace storage {
 class NodeStatisticsAndDeletedIDs : public TableStatistics {
 
 public:
-    NodeStatisticsAndDeletedIDs(common::table_id_t tableID, common::offset_t maxNodeOffset)
+    explicit NodeStatisticsAndDeletedIDs(const catalog::TableSchema& schema)
+        : TableStatistics{schema}, tableID{schema.tableID} {}
+
+    NodeStatisticsAndDeletedIDs(common::table_id_t tableID, common::offset_t maxNodeOffset,
+        std::unordered_map<common::property_id_t, std::unique_ptr<PropertyStatistics>>&&
+            propertyStatistics)
         : NodeStatisticsAndDeletedIDs(tableID, maxNodeOffset,
-              std::vector<
-                  common::offset_t>() /* no deleted node offsets during initial loading */) {}
+              std::vector<common::offset_t>() /* no deleted node offsets during initial loading */,
+              std::move(propertyStatistics)) {}
 
     NodeStatisticsAndDeletedIDs(common::table_id_t tableID, common::offset_t maxNodeOffset,
-        const std::vector<common::offset_t>& deletedNodeOffsets);
+        const std::vector<common::offset_t>& deletedNodeOffsets,
+        std::unordered_map<common::property_id_t, std::unique_ptr<PropertyStatistics>>&&
+            propertyStatistics);
 
-    NodeStatisticsAndDeletedIDs(const NodeStatisticsAndDeletedIDs& other)
-        : TableStatistics{other.getNumTuples()}, tableID{other.tableID},
-          adjListsAndColumns{other.adjListsAndColumns},
-          hasDeletedNodesPerMorsel{other.hasDeletedNodesPerMorsel},
-          deletedNodeOffsetsPerMorsel{other.deletedNodeOffsetsPerMorsel} {}
+    NodeStatisticsAndDeletedIDs(const NodeStatisticsAndDeletedIDs& other) = default;
 
     inline common::offset_t getMaxNodeOffset() {
         return getMaxNodeOffsetFromNumTuples(getNumTuples());
@@ -48,7 +51,7 @@ class NodeStatisticsAndDeletedIDs : public TableStatistics {
 
     std::vector<common::offset_t> getDeletedNodeOffsets();
 
-    static inline uint64_t geNumTuplesFromMaxNodeOffset(common::offset_t maxNodeOffset) {
+    static inline uint64_t getNumTuplesFromMaxNodeOffset(common::offset_t maxNodeOffset) {
         return (maxNodeOffset == UINT64_MAX) ? 0ull : maxNodeOffset + 1ull;
     }
 
@@ -86,7 +89,7 @@ class NodesStatisticsAndDeletedIDs : public TablesStatistics {
         logger->info("Initialized {}.", "NodesStatisticsAndDeletedIDs");
     }
 
-    // Should be used ony by tests;
+    // Should be used only by tests;
     explicit NodesStatisticsAndDeletedIDs(
         std::unordered_map<common::table_id_t, std::unique_ptr<NodeStatisticsAndDeletedIDs>>&
             nodesStatisticsAndDeletedIDs);
@@ -109,7 +112,7 @@ class NodesStatisticsAndDeletedIDs : public TablesStatistics {
     }
 
     inline void setNumTuplesForTable(common::table_id_t tableID, uint64_t numTuples) override {
-        initTableStatisticPerTableForWriteTrxIfNecessary();
+        initTableStatisticsForWriteTrx();
         ((NodeStatisticsAndDeletedIDs*)tablesStatisticsContentForWriteTrx
                 ->tableStatisticPerTable[tableID]
                 .get())
@@ -118,21 +121,18 @@ class NodesStatisticsAndDeletedIDs : public TablesStatistics {
 
     inline common::offset_t getMaxNodeOffset(
         transaction::Transaction* transaction, common::table_id_t tableID) {
-        return getMaxNodeOffset(transaction == nullptr || transaction->isReadOnly() ?
-                                    transaction::TransactionType::READ_ONLY :
-                                    transaction::TransactionType::WRITE,
-            tableID);
-    }
-
-    inline common::offset_t getMaxNodeOffset(
-        transaction::TransactionType transactionType, common::table_id_t tableID) {
-        return (transactionType == transaction::TransactionType::READ_ONLY ||
-                   tablesStatisticsContentForWriteTrx == nullptr) ?
-                   getNodeStatisticsAndDeletedIDs(tableID)->getMaxNodeOffset() :
-                   ((NodeStatisticsAndDeletedIDs*)tablesStatisticsContentForWriteTrx
-                           ->tableStatisticPerTable[tableID]
-                           .get())
-                       ->getMaxNodeOffset();
+        assert(transaction);
+        if (transaction->getType() == transaction::TransactionType::READ_ONLY) {
+            return getNodeStatisticsAndDeletedIDs(tableID)->getMaxNodeOffset();
+        } else {
+            std::unique_lock xLck{mtx};
+            return tablesStatisticsContentForWriteTrx == nullptr ?
+                       getNodeStatisticsAndDeletedIDs(tableID)->getMaxNodeOffset() :
+                       ((NodeStatisticsAndDeletedIDs*)tablesStatisticsContentForWriteTrx
+                               ->tableStatisticPerTable[tableID]
+                               .get())
+                           ->getMaxNodeOffset();
+        }
     }
 
     // This function is only used for testing purpose.
@@ -146,7 +146,7 @@ class NodesStatisticsAndDeletedIDs : public TablesStatistics {
     // keep the interface simple and no transaction is passed.
     common::offset_t addNode(common::table_id_t tableID) {
         lock_t lck{mtx};
-        initTableStatisticPerTableForWriteTrxIfNecessary();
+        initTableStatisticsForWriteTrxNoLock();
         return ((NodeStatisticsAndDeletedIDs*)tablesStatisticsContentForWriteTrx
                     ->tableStatisticPerTable[tableID]
                     .get())
@@ -156,7 +156,7 @@ class NodesStatisticsAndDeletedIDs : public TablesStatistics {
     // Refer to the comments for addNode.
     void deleteNode(common::table_id_t tableID, common::offset_t nodeOffset) {
         lock_t lck{mtx};
-        initTableStatisticPerTableForWriteTrxIfNecessary();
+        initTableStatisticsForWriteTrxNoLock();
         ((NodeStatisticsAndDeletedIDs*)tablesStatisticsContentForWriteTrx
                 ->tableStatisticPerTable[tableID]
                 .get())
@@ -179,10 +179,7 @@ class NodesStatisticsAndDeletedIDs : public TablesStatistics {
 
     inline std::unique_ptr<TableStatistics> constructTableStatistic(
         catalog::TableSchema* tableSchema) override {
-        // We use UINT64_MAX to represent an empty nodeTable which doesn't contain
-        // any nodes.
-        return std::make_unique<NodeStatisticsAndDeletedIDs>(
-            tableSchema->tableID, UINT64_MAX /* maxNodeOffset */);
+        return std::make_unique<NodeStatisticsAndDeletedIDs>(*tableSchema);
     }
 
     inline std::unique_ptr<TableStatistics> constructTableStatistic(
@@ -197,6 +194,8 @@ class NodesStatisticsAndDeletedIDs : public TablesStatistics {
     }
 
     std::unique_ptr<TableStatistics> deserializeTableStatistics(uint64_t numTuples,
+        std::unordered_map<common::property_id_t, std::unique_ptr<PropertyStatistics>>&&
+            propertyStats,
         uint64_t& offset, common::FileInfo* fileInfo, uint64_t tableID) override;
 
     void serializeTableStatistics(

src/include/storage/store/property_statistics.h

@@ -0,0 +1,52 @@
+#pragma once
+
+#include "common/types/types.h"
+#include "transaction/transaction.h"
+
+namespace kuzu {
+namespace storage {
+
+class PropertyStatistics {
+public:
+    PropertyStatistics() = default;
+    explicit PropertyStatistics(bool mayHaveNullValue) : mayHaveNullValue{mayHaveNullValue} {}
+    PropertyStatistics(const PropertyStatistics& other) {
+        this->mayHaveNullValue = other.mayHaveNullValue;
+    }
+
+    inline bool mayHaveNull() const { return mayHaveNullValue; }
+    PropertyStatistics(PropertyStatistics& other) = default;
+
+    void serialize(common::FileInfo* fileInfo, uint64_t& offset);
+    static std::unique_ptr<PropertyStatistics> deserialize(
+        common::FileInfo* fileInfo, uint64_t& offset);
+
+    inline void setHasNull() { mayHaveNullValue = true; }
+
+private:
+    // Stores whether or not the property is known to have contained a null value
+    // If false, the property is guaranteed to not contain any nulls
+    bool mayHaveNullValue = false;
+};
+
+class TablesStatistics;
+
+// Accessor used by NodeColumn, so that it doesn't need to handle the TableStatistics directly
+class RWPropertyStats {
+public:
+    RWPropertyStats(TablesStatistics* tablesStatistics, common::table_id_t tableID,
+        common::property_id_t propertyID);
+
+    static RWPropertyStats empty() { return RWPropertyStats(nullptr, 0, 0); }
+
+    bool mayHaveNull(const transaction::Transaction& transaction);
+    void setHasNull(const transaction::Transaction& transaction);
+
+private:
+    TablesStatistics* tablesStatistics;
+    common::table_id_t tableID;
+    common::property_id_t propertyID;
+};
+
+} // namespace storage
+} // namespace kuzu

src/include/storage/store/rels_statistics.h

@@ -13,9 +13,13 @@ class RelStatistics : public TableStatistics {
     friend class RelsStatistics;
 
 public:
-    RelStatistics() : TableStatistics{0 /* numTuples */}, nextRelOffset{0} {}
-    RelStatistics(uint64_t numRels, common::offset_t nextRelOffset)
-        : TableStatistics{numRels}, nextRelOffset{nextRelOffset} {}
+    RelStatistics(const catalog::TableSchema& tableSchema)
+        : TableStatistics{tableSchema}, nextRelOffset{0} {}
+    RelStatistics(uint64_t numRels,
+        std::unordered_map<common::property_id_t, std::unique_ptr<PropertyStatistics>>&&
+            propertyStats,
+        common::offset_t nextRelOffset)
+        : TableStatistics{numRels, std::move(propertyStats)}, nextRelOffset{nextRelOffset} {}
 
     inline common::offset_t getNextRelOffset() const { return nextRelOffset; }
 
@@ -64,7 +68,7 @@ class RelsStatistics : public TablesStatistics {
 
     inline std::unique_ptr<TableStatistics> constructTableStatistic(
         catalog::TableSchema* tableSchema) override {
-        return std::make_unique<RelStatistics>();
+        return std::make_unique<RelStatistics>(*tableSchema);
     }
 
     inline std::unique_ptr<TableStatistics> constructTableStatistic(
@@ -84,6 +88,8 @@ class RelsStatistics : public TablesStatistics {
     }
 
     std::unique_ptr<TableStatistics> deserializeTableStatistics(uint64_t numTuples,
+        std::unordered_map<common::property_id_t, std::unique_ptr<PropertyStatistics>>&&
+            propertyStats,
         uint64_t& offset, common::FileInfo* fileInfo, uint64_t tableID) override;
 
     void serializeTableStatistics(

src/include/storage/store/string_node_column.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "node_column.h"
+#include "storage/store/table_statistics.h"
 
 namespace kuzu {
 namespace storage {
@@ -14,7 +15,7 @@ class StringNodeColumn : public NodeColumn {
 public:
     StringNodeColumn(common::LogicalType dataType, const catalog::MetadataDAHInfo& metaDAHeaderInfo,
         BMFileHandle* dataFH, BMFileHandle* metadataFH, BufferManager* bufferManager, WAL* wal,
-        transaction::Transaction* transaction);
+        transaction::Transaction* transaction, RWPropertyStats propertyStatistics);
 
     void scan(transaction::Transaction* transaction, common::node_group_idx_t nodeGroupIdx,
         common::offset_t startOffsetInGroup, common::offset_t endOffsetInGroup,