#include "HashedDictionary.h" #include #include "DictionaryBlockInputStream.h" #include "DictionaryFactory.h" namespace { /// NOTE: Trailing return type is explicitly specified for SFINAE. /// google::sparse_hash_map template auto first(const T & value) -> decltype(value.first) { return value.first; } // NOLINT template auto second(const T & value) -> decltype(value.second) { return value.second; } // NOLINT /// HashMap template auto first(const T & value) -> decltype(value.getKey()) { return value.getKey(); } // NOLINT template auto second(const T & value) -> decltype(value.getMapped()) { return value.getMapped(); } // NOLINT } namespace DB { namespace ErrorCodes { extern const int TYPE_MISMATCH; extern const int BAD_ARGUMENTS; extern const int DICTIONARY_IS_EMPTY; extern const int UNSUPPORTED_METHOD; } HashedDictionary::HashedDictionary( const std::string & database_, const std::string & name_, const DictionaryStructure & dict_struct_, DictionarySourcePtr source_ptr_, const DictionaryLifetime dict_lifetime_, bool require_nonempty_, bool sparse_, BlockPtr saved_block_) : database(database_) , name(name_) , full_name{database_.empty() ? name_ : (database_ + "." + name_)} , dict_struct(dict_struct_) , source_ptr{std::move(source_ptr_)} , dict_lifetime(dict_lifetime_) , require_nonempty(require_nonempty_) , sparse(sparse_) , saved_block{std::move(saved_block_)} { createAttributes(); loadData(); calculateBytesAllocated(); } void HashedDictionary::toParent(const PaddedPODArray & ids, PaddedPODArray & out) const { const auto null_value = std::get(hierarchical_attribute->null_values); getItemsImpl( *hierarchical_attribute, ids, [&](const size_t row, const UInt64 value) { out[row] = value; }, [&](const size_t) { return null_value; }); } /// Allow to use single value in same way as array. static inline HashedDictionary::Key getAt(const PaddedPODArray & arr, const size_t idx) { return arr[idx]; } static inline HashedDictionary::Key getAt(const HashedDictionary::Key & value, const size_t) { return value; } template void HashedDictionary::isInAttrImpl(const AttrType & attr, const ChildType & child_ids, const AncestorType & ancestor_ids, PaddedPODArray & out) const { const auto null_value = std::get(hierarchical_attribute->null_values); const auto rows = out.size(); for (const auto row : ext::range(0, rows)) { auto id = getAt(child_ids, row); const auto ancestor_id = getAt(ancestor_ids, row); while (id != null_value && id != ancestor_id) { auto it = attr.find(id); if (it != std::end(attr)) id = second(*it); else break; } out[row] = id != null_value && id == ancestor_id; } query_count.fetch_add(rows, std::memory_order_relaxed); } template void HashedDictionary::isInImpl(const ChildType & child_ids, const AncestorType & ancestor_ids, PaddedPODArray & out) const { if (!sparse) return isInAttrImpl(*std::get>(hierarchical_attribute->maps), child_ids, ancestor_ids, out); return isInAttrImpl(*std::get>(hierarchical_attribute->sparse_maps), child_ids, ancestor_ids, out); } void HashedDictionary::isInVectorVector( const PaddedPODArray & child_ids, const PaddedPODArray & ancestor_ids, PaddedPODArray & out) const { isInImpl(child_ids, ancestor_ids, out); } void HashedDictionary::isInVectorConstant(const PaddedPODArray & child_ids, const Key ancestor_id, PaddedPODArray & out) const { isInImpl(child_ids, ancestor_id, out); } void HashedDictionary::isInConstantVector(const Key child_id, const PaddedPODArray & ancestor_ids, PaddedPODArray & out) const { isInImpl(child_id, ancestor_ids, out); } #define DECLARE(TYPE) \ void HashedDictionary::get##TYPE(const std::string & attribute_name, const PaddedPODArray & ids, ResultArrayType & out) \ const \ { \ const auto & attribute = getAttribute(attribute_name); \ checkAttributeType(full_name, attribute_name, attribute.type, AttributeUnderlyingType::ut##TYPE); \ \ const auto null_value = std::get(attribute.null_values); \ \ getItemsImpl( \ attribute, ids, [&](const size_t row, const auto value) { out[row] = value; }, [&](const size_t) { return null_value; }); \ } DECLARE(UInt8) DECLARE(UInt16) DECLARE(UInt32) DECLARE(UInt64) DECLARE(UInt128) DECLARE(Int8) DECLARE(Int16) DECLARE(Int32) DECLARE(Int64) DECLARE(Float32) DECLARE(Float64) DECLARE(Decimal32) DECLARE(Decimal64) DECLARE(Decimal128) #undef DECLARE void HashedDictionary::getString(const std::string & attribute_name, const PaddedPODArray & ids, ColumnString * out) const { const auto & attribute = getAttribute(attribute_name); checkAttributeType(full_name, attribute_name, attribute.type, AttributeUnderlyingType::utString); const auto & null_value = StringRef{std::get(attribute.null_values)}; getItemsImpl( attribute, ids, [&](const size_t, const StringRef value) { out->insertData(value.data, value.size); }, [&](const size_t) { return null_value; }); } #define DECLARE(TYPE) \ void HashedDictionary::get##TYPE( \ const std::string & attribute_name, \ const PaddedPODArray & ids, \ const PaddedPODArray & def, \ ResultArrayType & out) const \ { \ const auto & attribute = getAttribute(attribute_name); \ checkAttributeType(full_name, attribute_name, attribute.type, AttributeUnderlyingType::ut##TYPE); \ \ getItemsImpl( \ attribute, ids, [&](const size_t row, const auto value) { out[row] = value; }, [&](const size_t row) { return def[row]; }); \ } DECLARE(UInt8) DECLARE(UInt16) DECLARE(UInt32) DECLARE(UInt64) DECLARE(UInt128) DECLARE(Int8) DECLARE(Int16) DECLARE(Int32) DECLARE(Int64) DECLARE(Float32) DECLARE(Float64) DECLARE(Decimal32) DECLARE(Decimal64) DECLARE(Decimal128) #undef DECLARE void HashedDictionary::getString( const std::string & attribute_name, const PaddedPODArray & ids, const ColumnString * const def, ColumnString * const out) const { const auto & attribute = getAttribute(attribute_name); checkAttributeType(full_name, attribute_name, attribute.type, AttributeUnderlyingType::utString); getItemsImpl( attribute, ids, [&](const size_t, const StringRef value) { out->insertData(value.data, value.size); }, [&](const size_t row) { return def->getDataAt(row); }); } #define DECLARE(TYPE) \ void HashedDictionary::get##TYPE( \ const std::string & attribute_name, const PaddedPODArray & ids, const TYPE & def, ResultArrayType & out) const \ { \ const auto & attribute = getAttribute(attribute_name); \ checkAttributeType(full_name, attribute_name, attribute.type, AttributeUnderlyingType::ut##TYPE); \ \ getItemsImpl( \ attribute, ids, [&](const size_t row, const auto value) { out[row] = value; }, [&](const size_t) { return def; }); \ } DECLARE(UInt8) DECLARE(UInt16) DECLARE(UInt32) DECLARE(UInt64) DECLARE(UInt128) DECLARE(Int8) DECLARE(Int16) DECLARE(Int32) DECLARE(Int64) DECLARE(Float32) DECLARE(Float64) DECLARE(Decimal32) DECLARE(Decimal64) DECLARE(Decimal128) #undef DECLARE void HashedDictionary::getString( const std::string & attribute_name, const PaddedPODArray & ids, const String & def, ColumnString * const out) const { const auto & attribute = getAttribute(attribute_name); checkAttributeType(full_name, attribute_name, attribute.type, AttributeUnderlyingType::utString); getItemsImpl( attribute, ids, [&](const size_t, const StringRef value) { out->insertData(value.data, value.size); }, [&](const size_t) { return StringRef{def}; }); } void HashedDictionary::has(const PaddedPODArray & ids, PaddedPODArray & out) const { const auto & attribute = attributes.front(); switch (attribute.type) { case AttributeUnderlyingType::utUInt8: has(attribute, ids, out); break; case AttributeUnderlyingType::utUInt16: has(attribute, ids, out); break; case AttributeUnderlyingType::utUInt32: has(attribute, ids, out); break; case AttributeUnderlyingType::utUInt64: has(attribute, ids, out); break; case AttributeUnderlyingType::utUInt128: has(attribute, ids, out); break; case AttributeUnderlyingType::utInt8: has(attribute, ids, out); break; case AttributeUnderlyingType::utInt16: has(attribute, ids, out); break; case AttributeUnderlyingType::utInt32: has(attribute, ids, out); break; case AttributeUnderlyingType::utInt64: has(attribute, ids, out); break; case AttributeUnderlyingType::utFloat32: has(attribute, ids, out); break; case AttributeUnderlyingType::utFloat64: has(attribute, ids, out); break; case AttributeUnderlyingType::utString: has(attribute, ids, out); break; case AttributeUnderlyingType::utDecimal32: has(attribute, ids, out); break; case AttributeUnderlyingType::utDecimal64: has(attribute, ids, out); break; case AttributeUnderlyingType::utDecimal128: has(attribute, ids, out); break; } } void HashedDictionary::createAttributes() { const auto size = dict_struct.attributes.size(); attributes.reserve(size); for (const auto & attribute : dict_struct.attributes) { attribute_index_by_name.emplace(attribute.name, attributes.size()); attributes.push_back(createAttributeWithType(attribute.underlying_type, attribute.null_value)); if (attribute.hierarchical) { hierarchical_attribute = &attributes.back(); if (hierarchical_attribute->type != AttributeUnderlyingType::utUInt64) throw Exception{full_name + ": hierarchical attribute must be UInt64.", ErrorCodes::TYPE_MISMATCH}; } } } void HashedDictionary::blockToAttributes(const Block & block) { const auto & id_column = *block.safeGetByPosition(0).column; for (const size_t attribute_idx : ext::range(0, attributes.size())) { const IColumn & attribute_column = *block.safeGetByPosition(attribute_idx + 1).column; auto & attribute = attributes[attribute_idx]; for (const auto row_idx : ext::range(0, id_column.size())) if (setAttributeValue(attribute, id_column[row_idx].get(), attribute_column[row_idx])) ++element_count; } } void HashedDictionary::updateData() { if (!saved_block || saved_block->rows() == 0) { auto stream = source_ptr->loadUpdatedAll(); stream->readPrefix(); while (const auto block = stream->read()) { /// We are using this to keep saved data if input stream consists of multiple blocks if (!saved_block) saved_block = std::make_shared(block.cloneEmpty()); for (const auto attribute_idx : ext::range(0, attributes.size() + 1)) { const IColumn & update_column = *block.getByPosition(attribute_idx).column.get(); MutableColumnPtr saved_column = saved_block->getByPosition(attribute_idx).column->assumeMutable(); saved_column->insertRangeFrom(update_column, 0, update_column.size()); } } stream->readSuffix(); } else { auto stream = source_ptr->loadUpdatedAll(); stream->readPrefix(); while (Block block = stream->read()) { const auto & saved_id_column = *saved_block->safeGetByPosition(0).column; const auto & update_id_column = *block.safeGetByPosition(0).column; std::unordered_map> update_ids; for (size_t row = 0; row < update_id_column.size(); ++row) { const auto id = update_id_column.get64(row); update_ids[id].push_back(row); } const size_t saved_rows = saved_id_column.size(); IColumn::Filter filter(saved_rows); std::unordered_map>::iterator it; for (size_t row = 0; row < saved_id_column.size(); ++row) { auto id = saved_id_column.get64(row); it = update_ids.find(id); if (it != update_ids.end()) filter[row] = 0; else filter[row] = 1; } auto block_columns = block.mutateColumns(); for (const auto attribute_idx : ext::range(0, attributes.size() + 1)) { auto & column = saved_block->safeGetByPosition(attribute_idx).column; const auto & filtered_column = column->filter(filter, -1); block_columns[attribute_idx]->insertRangeFrom(*filtered_column.get(), 0, filtered_column->size()); } saved_block->setColumns(std::move(block_columns)); } stream->readSuffix(); } if (saved_block) blockToAttributes(*saved_block.get()); } void HashedDictionary::loadData() { if (!source_ptr->hasUpdateField()) { auto stream = source_ptr->loadAll(); stream->readPrefix(); while (const auto block = stream->read()) blockToAttributes(block); stream->readSuffix(); } else updateData(); if (require_nonempty && 0 == element_count) throw Exception{full_name + ": dictionary source is empty and 'require_nonempty' property is set.", ErrorCodes::DICTIONARY_IS_EMPTY}; } template void HashedDictionary::addAttributeSize(const Attribute & attribute) { if (!sparse) { const auto & map_ref = std::get>(attribute.maps); bytes_allocated += sizeof(CollectionType) + map_ref->getBufferSizeInBytes(); bucket_count = map_ref->getBufferSizeInCells(); } else { const auto & map_ref = std::get>(attribute.sparse_maps); bucket_count = map_ref->bucket_count(); /** TODO: more accurate calculation */ bytes_allocated += sizeof(SparseCollectionType); bytes_allocated += bucket_count; bytes_allocated += map_ref->size() * (sizeof(Key) + sizeof(T)); } } void HashedDictionary::calculateBytesAllocated() { bytes_allocated += attributes.size() * sizeof(attributes.front()); for (const auto & attribute : attributes) { switch (attribute.type) { case AttributeUnderlyingType::utUInt8: addAttributeSize(attribute); break; case AttributeUnderlyingType::utUInt16: addAttributeSize(attribute); break; case AttributeUnderlyingType::utUInt32: addAttributeSize(attribute); break; case AttributeUnderlyingType::utUInt64: addAttributeSize(attribute); break; case AttributeUnderlyingType::utUInt128: addAttributeSize(attribute); break; case AttributeUnderlyingType::utInt8: addAttributeSize(attribute); break; case AttributeUnderlyingType::utInt16: addAttributeSize(attribute); break; case AttributeUnderlyingType::utInt32: addAttributeSize(attribute); break; case AttributeUnderlyingType::utInt64: addAttributeSize(attribute); break; case AttributeUnderlyingType::utFloat32: addAttributeSize(attribute); break; case AttributeUnderlyingType::utFloat64: addAttributeSize(attribute); break; case AttributeUnderlyingType::utDecimal32: addAttributeSize(attribute); break; case AttributeUnderlyingType::utDecimal64: addAttributeSize(attribute); break; case AttributeUnderlyingType::utDecimal128: addAttributeSize(attribute); break; case AttributeUnderlyingType::utString: { addAttributeSize(attribute); bytes_allocated += sizeof(Arena) + attribute.string_arena->size(); break; } } } } template void HashedDictionary::createAttributeImpl(Attribute & attribute, const Field & null_value) { attribute.null_values = T(null_value.get>()); if (!sparse) attribute.maps = std::make_unique>(); else attribute.sparse_maps = std::make_unique>(); } HashedDictionary::Attribute HashedDictionary::createAttributeWithType(const AttributeUnderlyingType type, const Field & null_value) { Attribute attr{type, {}, {}, {}, {}}; switch (type) { case AttributeUnderlyingType::utUInt8: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utUInt16: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utUInt32: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utUInt64: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utUInt128: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utInt8: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utInt16: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utInt32: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utInt64: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utFloat32: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utFloat64: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utDecimal32: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utDecimal64: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utDecimal128: createAttributeImpl(attr, null_value); break; case AttributeUnderlyingType::utString: { attr.null_values = null_value.get(); if (!sparse) attr.maps = std::make_unique>(); else attr.sparse_maps = std::make_unique>(); attr.string_arena = std::make_unique(); break; } } return attr; } template void HashedDictionary::getItemsAttrImpl( const AttrType & attr, const PaddedPODArray & ids, ValueSetter && set_value, DefaultGetter && get_default) const { const auto rows = ext::size(ids); for (const auto i : ext::range(0, rows)) { const auto it = attr.find(ids[i]); set_value(i, it != attr.end() ? static_cast(second(*it)) : get_default(i)); } query_count.fetch_add(rows, std::memory_order_relaxed); } template void HashedDictionary::getItemsImpl( const Attribute & attribute, const PaddedPODArray & ids, ValueSetter && set_value, DefaultGetter && get_default) const { if (!sparse) return getItemsAttrImpl(*std::get>(attribute.maps), ids, set_value, get_default); return getItemsAttrImpl(*std::get>(attribute.sparse_maps), ids, set_value, get_default); } template bool HashedDictionary::setAttributeValueImpl(Attribute & attribute, const Key id, const T value) { if (!sparse) { auto & map = *std::get>(attribute.maps); return map.insert({id, value}).second; } else { auto & map = *std::get>(attribute.sparse_maps); return map.insert({id, value}).second; } } bool HashedDictionary::setAttributeValue(Attribute & attribute, const Key id, const Field & value) { switch (attribute.type) { case AttributeUnderlyingType::utUInt8: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utUInt16: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utUInt32: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utUInt64: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utUInt128: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utInt8: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utInt16: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utInt32: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utInt64: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utFloat32: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utFloat64: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utDecimal32: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utDecimal64: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utDecimal128: return setAttributeValueImpl(attribute, id, value.get()); case AttributeUnderlyingType::utString: { const auto & string = value.get(); const auto string_in_arena = attribute.string_arena->insert(string.data(), string.size()); if (!sparse) { auto & map = *std::get>(attribute.maps); return map.insert({id, StringRef{string_in_arena, string.size()}}).second; } else { auto & map = *std::get>(attribute.sparse_maps); return map.insert({id, StringRef{string_in_arena, string.size()}}).second; } } } throw Exception{"Invalid attribute type", ErrorCodes::BAD_ARGUMENTS}; } const HashedDictionary::Attribute & HashedDictionary::getAttribute(const std::string & attribute_name) const { const auto it = attribute_index_by_name.find(attribute_name); if (it == std::end(attribute_index_by_name)) throw Exception{full_name + ": no such attribute '" + attribute_name + "'", ErrorCodes::BAD_ARGUMENTS}; return attributes[it->second]; } template void HashedDictionary::has(const Attribute & attribute, const PaddedPODArray & ids, PaddedPODArray & out) const { const auto & attr = *std::get>(attribute.maps); const auto rows = ext::size(ids); for (const auto i : ext::range(0, rows)) out[i] = attr.find(ids[i]) != nullptr; query_count.fetch_add(rows, std::memory_order_relaxed); } template PaddedPODArray HashedDictionary::getIdsAttrImpl(const AttrType & attr) const { PaddedPODArray ids; ids.reserve(attr.size()); for (const auto & value : attr) ids.push_back(first(value)); return ids; } template PaddedPODArray HashedDictionary::getIds(const Attribute & attribute) const { if (!sparse) return getIdsAttrImpl(*std::get>(attribute.maps)); return getIdsAttrImpl(*std::get>(attribute.sparse_maps)); } PaddedPODArray HashedDictionary::getIds() const { const auto & attribute = attributes.front(); switch (attribute.type) { case AttributeUnderlyingType::utUInt8: return getIds(attribute); case AttributeUnderlyingType::utUInt16: return getIds(attribute); case AttributeUnderlyingType::utUInt32: return getIds(attribute); case AttributeUnderlyingType::utUInt64: return getIds(attribute); case AttributeUnderlyingType::utUInt128: return getIds(attribute); case AttributeUnderlyingType::utInt8: return getIds(attribute); case AttributeUnderlyingType::utInt16: return getIds(attribute); case AttributeUnderlyingType::utInt32: return getIds(attribute); case AttributeUnderlyingType::utInt64: return getIds(attribute); case AttributeUnderlyingType::utFloat32: return getIds(attribute); case AttributeUnderlyingType::utFloat64: return getIds(attribute); case AttributeUnderlyingType::utString: return getIds(attribute); case AttributeUnderlyingType::utDecimal32: return getIds(attribute); case AttributeUnderlyingType::utDecimal64: return getIds(attribute); case AttributeUnderlyingType::utDecimal128: return getIds(attribute); } return PaddedPODArray(); } BlockInputStreamPtr HashedDictionary::getBlockInputStream(const Names & column_names, size_t max_block_size) const { using BlockInputStreamType = DictionaryBlockInputStream; return std::make_shared(shared_from_this(), max_block_size, getIds(), column_names); } void registerDictionaryHashed(DictionaryFactory & factory) { auto create_layout = [=](const std::string & full_name, const DictionaryStructure & dict_struct, const Poco::Util::AbstractConfiguration & config, const std::string & config_prefix, DictionarySourcePtr source_ptr, bool sparse) -> DictionaryPtr { if (dict_struct.key) throw Exception{"'key' is not supported for dictionary of layout 'hashed'", ErrorCodes::UNSUPPORTED_METHOD}; if (dict_struct.range_min || dict_struct.range_max) throw Exception{full_name + ": elements .structure.range_min and .structure.range_max should be defined only " "for a dictionary of layout 'range_hashed'", ErrorCodes::BAD_ARGUMENTS}; const String database = config.getString(config_prefix + ".database", ""); const String name = config.getString(config_prefix + ".name"); const DictionaryLifetime dict_lifetime{config, config_prefix + ".lifetime"}; const bool require_nonempty = config.getBool(config_prefix + ".require_nonempty", false); return std::make_unique(database, name, dict_struct, std::move(source_ptr), dict_lifetime, require_nonempty, sparse); }; using namespace std::placeholders; factory.registerLayout("hashed", std::bind(create_layout, _1, _2, _3, _4, _5, /* sparse = */ false), false); factory.registerLayout("sparse_hashed", std::bind(create_layout, _1, _2, _3, _4, _5, /* sparse = */ true), false); } }