#pragma once #include #include #include #include #include #include #include #include #include #include namespace DB { namespace ErrorCodes { extern const int NOT_IMPLEMENTED; } struct CacheDictionaryStorageConfiguration { /// Max size of storage in cells const size_t max_size_in_cells; /// Needed to perform check if cell is expired or not found. Default value is dictionary max lifetime. const size_t strict_max_lifetime_seconds; /// Lifetime of dictionary. Cell deadline is random value between lifetime min and max seconds. const DictionaryLifetime lifetime; }; /** ICacheDictionaryStorage implementation that keeps key in hash table with fixed collision length. * Value in hash table point to index in attributes arrays. */ template class CacheDictionaryStorage final : public ICacheDictionaryStorage { static constexpr size_t max_collision_length = 10; public: using KeyType = std::conditional_t; explicit CacheDictionaryStorage( const DictionaryStructure & dictionary_structure, CacheDictionaryStorageConfiguration & configuration_) : configuration(configuration_) , rnd_engine(randomSeed()) { size_t cells_size = roundUpToPowerOfTwoOrZero(std::max(configuration.max_size_in_cells, max_collision_length)); cells.resize_fill(cells_size); size_overlap_mask = cells_size - 1; createAttributes(dictionary_structure); } bool returnsFetchedColumnsInOrderOfRequestedKeys() const override { return true; } String getName() const override { if (dictionary_key_type == DictionaryKeyType::Simple) return "Cache"; else return "ComplexKeyCache"; } bool supportsSimpleKeys() const override { return dictionary_key_type == DictionaryKeyType::Simple; } SimpleKeysStorageFetchResult fetchColumnsForKeys( const PaddedPODArray & keys, const DictionaryStorageFetchRequest & fetch_request) override { if constexpr (dictionary_key_type == DictionaryKeyType::Simple) return fetchColumnsForKeysImpl(keys, fetch_request); else throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Method fetchColumnsForKeys is not supported for complex key storage"); } void insertColumnsForKeys(const PaddedPODArray & keys, Columns columns) override { if constexpr (dictionary_key_type == DictionaryKeyType::Simple) insertColumnsForKeysImpl(keys, columns); else throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Method insertColumnsForKeys is not supported for complex key storage"); } void insertDefaultKeys(const PaddedPODArray & keys) override { if constexpr (dictionary_key_type == DictionaryKeyType::Simple) insertDefaultKeysImpl(keys); else throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Method insertDefaultKeysImpl is not supported for complex key storage"); } PaddedPODArray getCachedSimpleKeys() const override { if constexpr (dictionary_key_type == DictionaryKeyType::Simple) return getCachedKeysImpl(); else throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Method getCachedSimpleKeys is not supported for complex key storage"); } bool supportsComplexKeys() const override { return dictionary_key_type == DictionaryKeyType::Complex; } ComplexKeysStorageFetchResult fetchColumnsForKeys( const PaddedPODArray & keys, const DictionaryStorageFetchRequest & column_fetch_requests) override { if constexpr (dictionary_key_type == DictionaryKeyType::Complex) return fetchColumnsForKeysImpl(keys, column_fetch_requests); else throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Method fetchColumnsForKeys is not supported for simple key storage"); } void insertColumnsForKeys(const PaddedPODArray & keys, Columns columns) override { if constexpr (dictionary_key_type == DictionaryKeyType::Complex) insertColumnsForKeysImpl(keys, columns); else throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Method insertColumnsForKeys is not supported for simple key storage"); } void insertDefaultKeys(const PaddedPODArray & keys) override { if constexpr (dictionary_key_type == DictionaryKeyType::Complex) insertDefaultKeysImpl(keys); else throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Method insertDefaultKeysImpl is not supported for simple key storage"); } PaddedPODArray getCachedComplexKeys() const override { if constexpr (dictionary_key_type == DictionaryKeyType::Complex) return getCachedKeysImpl(); else throw Exception(ErrorCodes::NOT_IMPLEMENTED, "Method getCachedComplexKeys is not supported for simple key storage"); } size_t getSize() const override { return size; } double getLoadFactor() const override { return static_cast(size) / configuration.max_size_in_cells; } size_t getBytesAllocated() const override { size_t attributes_size_in_bytes = 0; size_t attributes_size = attributes.size(); for (size_t attribute_index = 0; attribute_index < attributes_size; ++attribute_index) { getAttributeContainer(attribute_index, [&](const auto & container) { attributes_size_in_bytes += container.capacity() * sizeof(container[0]); }); } return arena.size() + sizeof(Cell) * configuration.max_size_in_cells + attributes_size_in_bytes; } private: struct FetchedKey { FetchedKey(size_t element_index_, bool is_default_) : element_index(element_index_) , is_default(is_default_) {} size_t element_index; bool is_default; }; template KeysStorageFetchResult fetchColumnsForKeysImpl( const PaddedPODArray & keys, const DictionaryStorageFetchRequest & fetch_request) { KeysStorageFetchResult result; result.fetched_columns = fetch_request.makeAttributesResultColumns(); result.key_index_to_state.resize_fill(keys.size()); const time_t now = std::chrono::system_clock::to_time_t(std::chrono::system_clock::now()); size_t fetched_columns_index = 0; size_t keys_size = keys.size(); PaddedPODArray fetched_keys; fetched_keys.resize_fill(keys_size); for (size_t key_index = 0; key_index < keys_size; ++key_index) { auto key = keys[key_index]; auto [key_state, cell_index] = getKeyStateAndCellIndex(key, now); if (unlikely(key_state == KeyState::not_found)) { result.key_index_to_state[key_index] = {KeyState::not_found}; ++result.not_found_keys_size; continue; } auto & cell = cells[cell_index]; result.expired_keys_size += static_cast(key_state == KeyState::expired); result.key_index_to_state[key_index] = {key_state, fetched_columns_index}; fetched_keys[fetched_columns_index] = FetchedKey(cell.element_index, cell.is_default); ++fetched_columns_index; result.key_index_to_state[key_index].setDefaultValue(cell.is_default); result.default_keys_size += cell.is_default; } result.found_keys_size = keys_size - (result.expired_keys_size + result.not_found_keys_size); for (size_t attribute_index = 0; attribute_index < fetch_request.attributesSize(); ++attribute_index) { if (!fetch_request.shouldFillResultColumnWithIndex(attribute_index)) continue; auto & attribute = attributes[attribute_index]; const auto & default_value_provider = fetch_request.defaultValueProviderAtIndex(attribute_index); auto & fetched_column = *result.fetched_columns[attribute_index]; fetched_column.reserve(fetched_columns_index); if (unlikely(attribute.is_nullable)) { getItemsForFetchedKeys( attribute, fetched_columns_index, fetched_keys, [&](Field & value) { fetched_column.insert(value); }, default_value_provider); } else { auto type_call = [&](const auto & dictionary_attribute_type) { using Type = std::decay_t; using AttributeType = typename Type::AttributeType; using ColumnProvider = DictionaryAttributeColumnProvider; using ColumnType = typename ColumnProvider::ColumnType; using ValueType = DictionaryValueType; ColumnType & column_typed = static_cast(fetched_column); if constexpr (std::is_same_v) { getItemsForFetchedKeys( attribute, fetched_columns_index, fetched_keys, [&](Array & value) { fetched_column.insert(value); }, default_value_provider); } else if constexpr (std::is_same_v) { getItemsForFetchedKeys( attribute, fetched_columns_index, fetched_keys, [&](StringRef value) { fetched_column.insertData(value.data, value.size); }, default_value_provider); } else { auto & data = column_typed.getData(); getItemsForFetchedKeys( attribute, fetched_columns_index, fetched_keys, [&](auto value) { data.push_back(value); }, default_value_provider); } }; callOnDictionaryAttributeType(attribute.type, type_call); } } return result; } void insertColumnsForKeysImpl(const PaddedPODArray & keys, Columns columns) { const auto now = std::chrono::system_clock::now(); Field column_value; for (size_t key_index = 0; key_index < keys.size(); ++key_index) { auto key = keys[key_index]; size_t cell_index = getCellIndexForInsert(key); auto & cell = cells[cell_index]; bool cell_was_default = cell.is_default; cell.is_default = false; bool was_inserted = cell.deadline == 0; if (was_inserted) { if constexpr (std::is_same_v) cell.key = copyStringInArena(key); else cell.key = key; for (size_t attribute_index = 0; attribute_index < columns.size(); ++attribute_index) { auto & column = columns[attribute_index]; getAttributeContainer(attribute_index, [&](auto & container) { container.emplace_back(); cell.element_index = container.size() - 1; using ElementType = std::decay_t; column->get(key_index, column_value); if constexpr (std::is_same_v) { container.back() = column_value; } else if constexpr (std::is_same_v) { const String & string_value = column_value.get(); StringRef string_value_ref = StringRef {string_value.data(), string_value.size()}; StringRef inserted_value = copyStringInArena(string_value_ref); container.back() = inserted_value; } else { container.back() = column_value.get>(); } }); } ++size; } else { if (cell.key != key) { if constexpr (std::is_same_v) { char * data = const_cast(cell.key.data); arena.free(data, cell.key.size); cell.key = copyStringInArena(key); } else cell.key = key; } /// Put values into existing index size_t index_to_use = cell.element_index; for (size_t attribute_index = 0; attribute_index < columns.size(); ++attribute_index) { auto & column = columns[attribute_index]; getAttributeContainer(attribute_index, [&](auto & container) { using ElementType = std::decay_t; column->get(key_index, column_value); if constexpr (std::is_same_v) { container[index_to_use] = column_value; } else if constexpr (std::is_same_v) { const String & string_value = column_value.get(); StringRef string_ref_value = StringRef {string_value.data(), string_value.size()}; StringRef inserted_value = copyStringInArena(string_ref_value); if (!cell_was_default) { StringRef previous_value = container[index_to_use]; arena.free(const_cast(previous_value.data), previous_value.size); } container[index_to_use] = inserted_value; } else { container[index_to_use] = column_value.get>(); } }); } } setCellDeadline(cell, now); } } void insertDefaultKeysImpl(const PaddedPODArray & keys) { const auto now = std::chrono::system_clock::now(); size_t keys_size = keys.size(); for (size_t key_index = 0; key_index < keys_size; ++key_index) { auto key = keys[key_index]; size_t cell_index = getCellIndexForInsert(key); auto & cell = cells[cell_index]; bool was_inserted = cell.deadline == 0; bool cell_was_default = cell.is_default; cell.is_default = true; if (was_inserted) { if constexpr (std::is_same_v) cell.key = copyStringInArena(key); else cell.key = key; for (size_t attribute_index = 0; attribute_index < attributes.size(); ++attribute_index) { getAttributeContainer(attribute_index, [&](auto & container) { container.emplace_back(); cell.element_index = container.size() - 1; }); } ++size; } else { for (size_t attribute_index = 0; attribute_index < attributes.size(); ++attribute_index) { getAttributeContainer(attribute_index, [&](const auto & container) { using ElementType = std::decay_t; if constexpr (std::is_same_v) { if (!cell_was_default) { StringRef previous_value = container[cell.element_index]; arena.free(const_cast(previous_value.data), previous_value.size); } } }); } if (cell.key != key) { if constexpr (std::is_same_v) { char * data = const_cast(cell.key.data); arena.free(data, cell.key.size); cell.key = copyStringInArena(key); } else cell.key = key; } } setCellDeadline(cell, now); } } PaddedPODArray getCachedKeysImpl() const { PaddedPODArray result; result.reserve(size); for (auto & cell : cells) { if (cell.deadline == 0) continue; if (cell.is_default) continue; result.emplace_back(cell.key); } return result; } template void getAttributeContainer(size_t attribute_index, GetContainerFunc && func) { auto & attribute = attributes[attribute_index]; auto & attribute_type = attribute.type; if (unlikely(attribute.is_nullable)) { auto & container = std::get>(attribute.attribute_container); std::forward(func)(container); } else { auto type_call = [&](const auto & dictionary_attribute_type) { using Type = std::decay_t; using AttributeType = typename Type::AttributeType; using ValueType = DictionaryValueType; auto & container = std::get>(attribute.attribute_container); std::forward(func)(container); }; callOnDictionaryAttributeType(attribute_type, type_call); } } template void getAttributeContainer(size_t attribute_index, GetContainerFunc && func) const { return const_cast *>(this)->template getAttributeContainer(attribute_index, std::forward(func)); } StringRef copyStringInArena(StringRef value_to_copy) { size_t value_to_copy_size = value_to_copy.size; char * place_for_key = arena.alloc(value_to_copy_size); memcpy(reinterpret_cast(place_for_key), reinterpret_cast(value_to_copy.data), value_to_copy_size); StringRef updated_value{place_for_key, value_to_copy_size}; return updated_value; } template using ContainerType = std::conditional_t< std::is_same_v || std::is_same_v, std::vector, PaddedPODArray>; struct Attribute { AttributeUnderlyingType type; bool is_nullable; std::variant< ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType, ContainerType> attribute_container; }; template void getItemsForFetchedKeys( Attribute & attribute, size_t fetched_keys_size, PaddedPODArray & fetched_keys, ValueSetter && value_setter, const DefaultValueProvider & default_value_provider) { auto & container = std::get>(attribute.attribute_container); for (size_t fetched_key_index = 0; fetched_key_index < fetched_keys_size; ++fetched_key_index) { auto fetched_key = fetched_keys[fetched_key_index]; if (unlikely(fetched_key.is_default)) { auto default_value = default_value_provider.getDefaultValue(fetched_key_index); if constexpr (std::is_same_v) { value_setter(default_value); } else if constexpr (std::is_same_v) { auto & value = default_value.get(); value_setter(value); } else { value_setter(default_value.get()); } } else { value_setter(container[fetched_key.element_index]); } } } void createAttributes(const DictionaryStructure & dictionary_structure) { /// For each dictionary attribute create storage attribute /// For simple attributes create PODArray, for complex vector of Fields attributes.reserve(dictionary_structure.attributes.size()); for (const auto & dictionary_attribute : dictionary_structure.attributes) { auto attribute_type = dictionary_attribute.underlying_type; auto type_call = [&](const auto & dictionary_attribute_type) { using Type = std::decay_t; using AttributeType = typename Type::AttributeType; using ValueType = DictionaryValueType; attributes.emplace_back(); auto & last_attribute = attributes.back(); last_attribute.type = attribute_type; last_attribute.is_nullable = dictionary_attribute.is_nullable; if (dictionary_attribute.is_nullable) last_attribute.attribute_container = ContainerType(); else last_attribute.attribute_container = ContainerType(); }; callOnDictionaryAttributeType(attribute_type, type_call); } } using TimePoint = std::chrono::system_clock::time_point; struct Cell { KeyType key; size_t element_index; bool is_default; time_t deadline; }; CacheDictionaryStorageConfiguration configuration; pcg64 rnd_engine; size_t size_overlap_mask = 0; size_t size = 0; PaddedPODArray cells; ArenaWithFreeLists arena; std::vector attributes; inline void setCellDeadline(Cell & cell, TimePoint now) { if (configuration.lifetime.min_sec == 0 && configuration.lifetime.max_sec == 0) { /// This maybe not obvious, but when we define is this cell is expired or expired permanently, we add strict_max_lifetime_seconds /// to the expiration time. And it overflows pretty well. auto deadline = std::chrono::time_point::max() - 2 * std::chrono::seconds(configuration.strict_max_lifetime_seconds); cell.deadline = std::chrono::system_clock::to_time_t(deadline); return; } size_t min_sec_lifetime = configuration.lifetime.min_sec; size_t max_sec_lifetime = configuration.lifetime.max_sec; std::uniform_int_distribution distribution{min_sec_lifetime, max_sec_lifetime}; auto deadline = now + std::chrono::seconds(distribution(rnd_engine)); cell.deadline = std::chrono::system_clock::to_time_t(deadline); } inline size_t getCellIndex(const KeyType key) const { const size_t hash = DefaultHash()(key); const size_t index = hash & size_overlap_mask; return index; } using KeyStateAndCellIndex = std::pair; inline KeyStateAndCellIndex getKeyStateAndCellIndex(const KeyType key, const time_t now) const { size_t place_value = getCellIndex(key); const size_t place_value_end = place_value + max_collision_length; time_t max_lifetime_seconds = static_cast(configuration.strict_max_lifetime_seconds); for (; place_value < place_value_end; ++place_value) { const auto cell_place_value = place_value & size_overlap_mask; const auto & cell = cells[cell_place_value]; if (cell.key != key) continue; if (unlikely(now > cell.deadline + max_lifetime_seconds)) return std::make_pair(KeyState::not_found, cell_place_value); if (unlikely(now > cell.deadline)) return std::make_pair(KeyState::expired, cell_place_value); return std::make_pair(KeyState::found, cell_place_value); } return std::make_pair(KeyState::not_found, place_value & size_overlap_mask); } inline size_t getCellIndexForInsert(const KeyType & key) const { size_t place_value = getCellIndex(key); const size_t place_value_end = place_value + max_collision_length; size_t oldest_place_value = place_value; time_t oldest_time = std::numeric_limits::max(); for (; place_value < place_value_end; ++place_value) { const size_t cell_place_value = place_value & size_overlap_mask; const Cell cell = cells[cell_place_value]; if (cell.deadline == 0) return cell_place_value; if (cell.key == key) return cell_place_value; if (cell.deadline < oldest_time) { oldest_time = cell.deadline; oldest_place_value = cell_place_value; } } return oldest_place_value; } }; }