#include "CacheDictionary.h" #include #include #include #include #include #include #include #include #include #include namespace ProfileEvents { extern const Event DictCacheKeysRequested; extern const Event DictCacheKeysRequestedMiss; extern const Event DictCacheKeysRequestedFound; extern const Event DictCacheKeysExpired; extern const Event DictCacheKeysNotFound; extern const Event DictCacheKeysHit; extern const Event DictCacheRequestTimeNs; extern const Event DictCacheRequests; extern const Event DictCacheLockWriteNs; extern const Event DictCacheLockReadNs; } namespace CurrentMetrics { extern const Metric DictCacheRequests; } namespace DB { namespace ErrorCodes { extern const int CACHE_DICTIONARY_UPDATE_FAIL; extern const int UNSUPPORTED_METHOD; } template CacheDictionary::CacheDictionary( const StorageID & dict_id_, const DictionaryStructure & dict_struct_, DictionarySourcePtr source_ptr_, CacheDictionaryStoragePtr cache_storage_ptr_, CacheDictionaryUpdateQueueConfiguration update_queue_configuration_, DictionaryLifetime dict_lifetime_, bool allow_read_expired_keys_) : IDictionary(dict_id_) , dict_struct(dict_struct_) , source_ptr{std::move(source_ptr_)} , cache_storage_ptr(cache_storage_ptr_) , update_queue( dict_id_.getNameForLogs(), update_queue_configuration_, [this](CacheDictionaryUpdateUnitPtr unit_to_update) { update(unit_to_update); }) , dict_lifetime(dict_lifetime_) , log(&Poco::Logger::get("ExternalDictionaries")) , allow_read_expired_keys(allow_read_expired_keys_) , rnd_engine(randomSeed()) { if (!source_ptr->supportsSelectiveLoad()) throw Exception(ErrorCodes::UNSUPPORTED_METHOD, "{}: source cannot be used with CacheDictionary", full_name); } template CacheDictionary::~CacheDictionary() { update_queue.stopAndWait(); } template size_t CacheDictionary::getElementCount() const { const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs}; return cache_storage_ptr->getSize(); } template size_t CacheDictionary::getBytesAllocated() const { /// In case of existing string arena we check the size of it. /// But the same appears in setAttributeValue() function, which is called from update() function /// which in turn is called from another thread. const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs}; return cache_storage_ptr->getBytesAllocated(); } template double CacheDictionary::getLoadFactor() const { const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs}; return cache_storage_ptr->getLoadFactor(); } template std::exception_ptr CacheDictionary::getLastException() const { const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs}; return last_exception; } template const IDictionarySource * CacheDictionary::getSource() const { /// Mutex required here because of the getSourceAndUpdateIfNeeded() function /// which is used from another thread. std::lock_guard lock(source_mutex); return source_ptr.get(); } template ColumnPtr CacheDictionary::getColumn( const std::string & attribute_name, const DataTypePtr & result_type, const Columns & key_columns, const DataTypes & key_types, const ColumnPtr & default_values_column) const { return getColumns({attribute_name}, {result_type}, key_columns, key_types, {default_values_column}).front(); } template Columns CacheDictionary::getColumns( const Strings & attribute_names, const DataTypes & result_types, const Columns & key_columns, const DataTypes & key_types, const Columns & default_values_columns) const { /** * Flow of getColumsImpl * 1. Get fetch result from storage * 2. If all keys are found in storage and not expired * 2.1. If storage returns fetched columns in order of keys then result is returned to client. * 2.2. If storage does not return fetched columns in order of keys then reorder * result columns and return result to client. * 3. If all keys are found in storage but some of them are expired and we allow to read expired keys * start async request to source and perform actions from step 2 for result returned from storage. * 4. If some keys are found and some are not, start sync update from source. * 5. Aggregate columns returned from storage and source, if key is not found in storage and in source * use default value. */ if (dictionary_key_type == DictionaryKeyType::complex) dict_struct.validateKeyTypes(key_types); DictionaryKeysArenaHolder arena_holder; DictionaryKeysExtractor extractor(key_columns, arena_holder.getComplexKeyArena()); auto keys = extractor.extractAllKeys(); DictionaryStorageFetchRequest request(dict_struct, attribute_names, result_types, default_values_columns); FetchResult result_of_fetch_from_storage; { const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockWriteNs}; result_of_fetch_from_storage = cache_storage_ptr->fetchColumnsForKeys(keys, request); } size_t found_keys_size = result_of_fetch_from_storage.found_keys_size; size_t expired_keys_size = result_of_fetch_from_storage.expired_keys_size; size_t not_found_keys_size = result_of_fetch_from_storage.not_found_keys_size; ProfileEvents::increment(ProfileEvents::DictCacheKeysHit, found_keys_size); ProfileEvents::increment(ProfileEvents::DictCacheKeysExpired, expired_keys_size); ProfileEvents::increment(ProfileEvents::DictCacheKeysNotFound, not_found_keys_size); query_count.fetch_add(keys.size(), std::memory_order_relaxed); hit_count.fetch_add(found_keys_size, std::memory_order_relaxed); found_count.fetch_add(found_keys_size, std::memory_order_relaxed); MutableColumns & fetched_columns_from_storage = result_of_fetch_from_storage.fetched_columns; const PaddedPODArray & key_index_to_state_from_storage = result_of_fetch_from_storage.key_index_to_state; bool source_returns_fetched_columns_in_order_of_keys = cache_storage_ptr->returnsFetchedColumnsInOrderOfRequestedKeys(); if (not_found_keys_size == 0 && expired_keys_size == 0) { /// All keys were found in storage if (source_returns_fetched_columns_in_order_of_keys) return request.filterRequestedColumns(fetched_columns_from_storage); else { /// Reorder result from storage to requested keys indexes MutableColumns aggregated_columns = aggregateColumnsInOrderOfKeys( keys, request, fetched_columns_from_storage, key_index_to_state_from_storage); return request.filterRequestedColumns(aggregated_columns); } } size_t keys_to_update_size = not_found_keys_size + expired_keys_size; auto update_unit = std::make_shared>(key_columns, key_index_to_state_from_storage, request, keys_to_update_size); HashMap requested_keys_to_fetched_columns_during_update_index; MutableColumns fetched_columns_during_update = request.makeAttributesResultColumns(); if (not_found_keys_size == 0 && expired_keys_size > 0 && allow_read_expired_keys) { /// Start async update only if allow read expired keys and all keys are found update_queue.tryPushToUpdateQueueOrThrow(update_unit); if (source_returns_fetched_columns_in_order_of_keys) return request.filterRequestedColumns(fetched_columns_from_storage); else { /// Reorder result from storage to requested keys indexes MutableColumns aggregated_columns = aggregateColumnsInOrderOfKeys( keys, request, fetched_columns_from_storage, key_index_to_state_from_storage); return request.filterRequestedColumns(aggregated_columns); } } else { /// Start sync update update_queue.tryPushToUpdateQueueOrThrow(update_unit); update_queue.waitForCurrentUpdateFinish(update_unit); requested_keys_to_fetched_columns_during_update_index = std::move(update_unit->requested_keys_to_fetched_columns_during_update_index); fetched_columns_during_update = std::move(update_unit->fetched_columns_during_update); } MutableColumns aggregated_columns = aggregateColumns( keys, request, fetched_columns_from_storage, key_index_to_state_from_storage, fetched_columns_during_update, requested_keys_to_fetched_columns_during_update_index); return request.filterRequestedColumns(aggregated_columns); } template ColumnUInt8::Ptr CacheDictionary::hasKeys(const Columns & key_columns, const DataTypes & key_types) const { /** * Flow of hasKeys. It is similar to getColumns. But there is an important detail, if key is identified with default value in storage * it means that in hasKeys result this key will be false. * * 1. Get fetch result from storage * 2. If all keys are found in storage and not expired and there are no default keys return that we have all keys. * Otherwise set allow_expired_keys_during_aggregation and go to step 5. * 3. If all keys are found in storage and some of them are expired and allow_read_expired keys is true return that we have all keys. * Otherwise set allow_expired_keys_during_aggregation and go to step 5. * 4. If not all keys are found in storage start sync update from source. * 5. Start aggregation of keys from source and storage. * If we allow read expired keys from step 2 or 3 then count them as founded in storage. * Check if key was found in storage not default for that key set true in result array. * Check that key was fetched during update for that key set true in result array. */ if (dictionary_key_type == DictionaryKeyType::complex) dict_struct.validateKeyTypes(key_types); DictionaryKeysArenaHolder arena_holder; DictionaryKeysExtractor extractor(key_columns, arena_holder.getComplexKeyArena()); const auto keys = extractor.extractAllKeys(); /// We make empty request just to fetch if keys exists DictionaryStorageFetchRequest request(dict_struct, {}, {}, {}); FetchResult result_of_fetch_from_storage; { /// Write lock on storage const ProfilingScopedWriteRWLock write_lock{rw_lock, ProfileEvents::DictCacheLockWriteNs}; result_of_fetch_from_storage = cache_storage_ptr->fetchColumnsForKeys(keys, request); } size_t found_keys_size = result_of_fetch_from_storage.found_keys_size; size_t expired_keys_size = result_of_fetch_from_storage.expired_keys_size; size_t not_found_keys_size = result_of_fetch_from_storage.not_found_keys_size; ProfileEvents::increment(ProfileEvents::DictCacheKeysHit, found_keys_size); ProfileEvents::increment(ProfileEvents::DictCacheKeysExpired, expired_keys_size); ProfileEvents::increment(ProfileEvents::DictCacheKeysNotFound, not_found_keys_size); query_count.fetch_add(keys.size(), std::memory_order_relaxed); hit_count.fetch_add(found_keys_size, std::memory_order_relaxed); found_count.fetch_add(found_keys_size, std::memory_order_relaxed); size_t keys_to_update_size = expired_keys_size + not_found_keys_size; auto update_unit = std::make_shared>(key_columns, result_of_fetch_from_storage.key_index_to_state, request, keys_to_update_size); HashMap requested_keys_to_fetched_columns_during_update_index; bool allow_expired_keys_during_aggregation = false; if (not_found_keys_size == 0 && expired_keys_size == 0) { /// All keys were found in storage if (result_of_fetch_from_storage.default_keys_size == 0) return ColumnUInt8::create(keys.size(), true); allow_expired_keys_during_aggregation = true; } else if (not_found_keys_size == 0 && expired_keys_size > 0 && allow_read_expired_keys) { /// Start async update only if allow read expired keys and all keys are found update_queue.tryPushToUpdateQueueOrThrow(update_unit); if (result_of_fetch_from_storage.default_keys_size == 0) return ColumnUInt8::create(keys.size(), true); allow_expired_keys_during_aggregation = true; } else { /// Start sync update update_queue.tryPushToUpdateQueueOrThrow(update_unit); update_queue.waitForCurrentUpdateFinish(update_unit); requested_keys_to_fetched_columns_during_update_index = std::move(update_unit->requested_keys_to_fetched_columns_during_update_index); } auto result = ColumnUInt8::create(keys.size(), false); auto & data = result->getData(); for (size_t key_index = 0; key_index < keys.size(); ++key_index) { auto key = keys[key_index]; bool valid_expired_key = allow_expired_keys_during_aggregation && result_of_fetch_from_storage.key_index_to_state[key_index].isExpired(); if (result_of_fetch_from_storage.key_index_to_state[key_index].isFound() || valid_expired_key) { /// Check if key was fetched from cache data[key_index] = !result_of_fetch_from_storage.key_index_to_state[key_index].isDefault(); } if (requested_keys_to_fetched_columns_during_update_index.has(key)) { /// Check if key was not in cache and was fetched during update data[key_index] = true; } } return result; } template ColumnPtr CacheDictionary::getHierarchy( ColumnPtr key_column [[maybe_unused]], const DataTypePtr & key_type [[maybe_unused]]) const { if (dictionary_key_type == DictionaryKeyType::simple) { size_t keys_found; auto result = getKeysHierarchyDefaultImplementation(this, key_column, key_type, keys_found); query_count.fetch_add(key_column->size(), std::memory_order_relaxed); found_count.fetch_add(keys_found, std::memory_order_relaxed); return result; } else return nullptr; } template ColumnUInt8::Ptr CacheDictionary::isInHierarchy( ColumnPtr key_column [[maybe_unused]], ColumnPtr in_key_column [[maybe_unused]], const DataTypePtr & key_type [[maybe_unused]]) const { if (dictionary_key_type == DictionaryKeyType::simple) { size_t keys_found; auto result = getKeysIsInHierarchyDefaultImplementation(this, key_column, in_key_column, key_type, keys_found); query_count.fetch_add(key_column->size(), std::memory_order_relaxed); found_count.fetch_add(keys_found, std::memory_order_relaxed); return result; } else return nullptr; } template MutableColumns CacheDictionary::aggregateColumnsInOrderOfKeys( const PaddedPODArray & keys, const DictionaryStorageFetchRequest & request, const MutableColumns & fetched_columns, const PaddedPODArray & key_index_to_state) { MutableColumns aggregated_columns = request.makeAttributesResultColumns(); /// If keys were returned not in order of keys, aggregate fetched columns in order of requested keys. for (size_t fetch_request_index = 0; fetch_request_index < request.attributesSize(); ++fetch_request_index) { if (!request.shouldFillResultColumnWithIndex(fetch_request_index)) continue; const auto & aggregated_column = aggregated_columns[fetch_request_index]; const auto & fetched_column = fetched_columns[fetch_request_index]; for (size_t key_index = 0; key_index < keys.size(); ++key_index) { auto state = key_index_to_state[key_index]; if (state.isNotFound()) continue; aggregated_column->insertFrom(*fetched_column, state.getFetchedColumnIndex()); } } return aggregated_columns; } template MutableColumns CacheDictionary::aggregateColumns( const PaddedPODArray & keys, const DictionaryStorageFetchRequest & request, const MutableColumns & fetched_columns_from_storage, const PaddedPODArray & key_index_to_fetched_columns_from_storage_result, const MutableColumns & fetched_columns_during_update, const HashMap & found_keys_to_fetched_columns_during_update_index) { /** * Aggregation of columns fetched from storage and from source during update. * If key was found in storage add it to result. * If key was found in source during update add it to result. * If key was not found in storage or in source during update add default value. */ MutableColumns aggregated_columns = request.makeAttributesResultColumns(); for (size_t fetch_request_index = 0; fetch_request_index < request.attributesSize(); ++fetch_request_index) { if (!request.shouldFillResultColumnWithIndex(fetch_request_index)) continue; const auto & aggregated_column = aggregated_columns[fetch_request_index]; const auto & fetched_column_from_storage = fetched_columns_from_storage[fetch_request_index]; const auto & fetched_column_during_update = fetched_columns_during_update[fetch_request_index]; const auto & default_value_provider = request.defaultValueProviderAtIndex(fetch_request_index); for (size_t key_index = 0; key_index < keys.size(); ++key_index) { auto key = keys[key_index]; auto key_state_from_storage = key_index_to_fetched_columns_from_storage_result[key_index]; if (key_state_from_storage.isFound()) { /// Check and insert value if key was fetched from cache aggregated_column->insertFrom(*fetched_column_from_storage, key_state_from_storage.getFetchedColumnIndex()); continue; } /// Check and insert value if key was not in cache and was fetched during update const auto * find_iterator_in_fetch_during_update = found_keys_to_fetched_columns_during_update_index.find(key); if (find_iterator_in_fetch_during_update) { aggregated_column->insertFrom(*fetched_column_during_update, find_iterator_in_fetch_during_update->getMapped()); continue; } /// Insert default value aggregated_column->insert(default_value_provider.getDefaultValue(key_index)); } } return aggregated_columns; } template BlockInputStreamPtr CacheDictionary::getBlockInputStream(const Names & column_names, size_t max_block_size) const { std::shared_ptr stream; { /// Write lock on storage const ProfilingScopedWriteRWLock write_lock{rw_lock, ProfileEvents::DictCacheLockWriteNs}; if constexpr (dictionary_key_type == DictionaryKeyType::simple) stream = std::make_shared(shared_from_this(), max_block_size, cache_storage_ptr->getCachedSimpleKeys(), column_names); else { auto keys = cache_storage_ptr->getCachedComplexKeys(); stream = std::make_shared(shared_from_this(), max_block_size, keys, column_names); } } return stream; } template void CacheDictionary::update(CacheDictionaryUpdateUnitPtr update_unit_ptr) { /** * Update has following flow. * 1. Filter only necessary keys to request, keys that are expired or not found. * And create not_found_keys hash_set including each requested key. * In case of simple_keys we need to fill requested_keys_vector with requested value key. * In case of complex_keys we need to fill requested_complex_key_rows with requested row. * 2. Create stream from source with necessary keys to request using method for simple or complex keys. * 3. Create fetched columns during update variable. This columns will aggregate columns that we fetch from source. * 4. When block is fetched from source. Split it into keys columns and attributes columns. * Insert attributes columns into associated fetched columns during update. * Create KeysExtractor and extract keys from keys columns. * Update map of requested found key to fetched column index. * Remove found key from not_found_keys. * 5. Add aggregated columns during update into storage. * 6. Add not found keys as default into storage. */ CurrentMetrics::Increment metric_increment{CurrentMetrics::DictCacheRequests}; Arena * complex_key_arena = update_unit_ptr->complex_keys_arena_holder.getComplexKeyArena(); DictionaryKeysExtractor requested_keys_extractor(update_unit_ptr->key_columns, complex_key_arena); auto requested_keys = requested_keys_extractor.extractAllKeys(); HashSet not_found_keys; std::vector requested_keys_vector; std::vector requested_complex_key_rows; if constexpr (dictionary_key_type == DictionaryKeyType::simple) requested_keys_vector.reserve(requested_keys.size()); else requested_complex_key_rows.reserve(requested_keys.size()); auto & key_index_to_state_from_storage = update_unit_ptr->key_index_to_state; for (size_t i = 0; i < key_index_to_state_from_storage.size(); ++i) { if (key_index_to_state_from_storage[i].isExpired() || key_index_to_state_from_storage[i].isNotFound()) { if constexpr (dictionary_key_type == DictionaryKeyType::simple) requested_keys_vector.emplace_back(requested_keys[i]); else requested_complex_key_rows.emplace_back(i); auto requested_key = requested_keys[i]; not_found_keys.insert(requested_key); } } size_t requested_keys_size = update_unit_ptr->keys_to_update_size; ProfileEvents::increment(ProfileEvents::DictCacheKeysRequested, requested_keys_size); const auto & fetch_request = update_unit_ptr->request; const auto now = std::chrono::system_clock::now(); if (now > backoff_end_time.load()) { try { auto current_source_ptr = getSourceAndUpdateIfNeeded(); Stopwatch watch; BlockInputStreamPtr stream; if constexpr (dictionary_key_type == DictionaryKeyType::simple) stream = current_source_ptr->loadIds(requested_keys_vector); else stream = current_source_ptr->loadKeys(update_unit_ptr->key_columns, requested_complex_key_rows); stream->readPrefix(); size_t skip_keys_size_offset = dict_struct.getKeysSize(); PaddedPODArray found_keys_in_source; Columns fetched_columns_during_update = fetch_request.makeAttributesResultColumnsNonMutable(); while (Block block = stream->read()) { Columns key_columns; key_columns.reserve(skip_keys_size_offset); auto block_columns = block.getColumns(); /// Split into keys columns and attribute columns for (size_t i = 0; i < skip_keys_size_offset; ++i) { key_columns.emplace_back(*block_columns.begin()); block_columns.erase(block_columns.begin()); } DictionaryKeysExtractor keys_extractor(key_columns, complex_key_arena); auto keys_extracted_from_block = keys_extractor.extractAllKeys(); for (size_t index_of_attribute = 0; index_of_attribute < fetched_columns_during_update.size(); ++index_of_attribute) { auto & column_to_update = fetched_columns_during_update[index_of_attribute]; auto column = block.safeGetByPosition(skip_keys_size_offset + index_of_attribute).column; column_to_update->assumeMutable()->insertRangeFrom(*column, 0, keys_extracted_from_block.size()); } for (size_t i = 0; i < keys_extracted_from_block.size(); ++i) { auto fetched_key_from_source = keys_extracted_from_block[i]; not_found_keys.erase(fetched_key_from_source); update_unit_ptr->requested_keys_to_fetched_columns_during_update_index[fetched_key_from_source] = found_keys_in_source.size(); found_keys_in_source.emplace_back(fetched_key_from_source); } } PaddedPODArray not_found_keys_in_source; not_found_keys_in_source.reserve(not_found_keys.size()); for (auto & cell : not_found_keys) not_found_keys_in_source.emplace_back(cell.getKey()); auto & update_unit_ptr_mutable_columns = update_unit_ptr->fetched_columns_during_update; for (const auto & fetched_column : fetched_columns_during_update) update_unit_ptr_mutable_columns.emplace_back(fetched_column->assumeMutable()); stream->readSuffix(); { /// Lock for cache modification ProfilingScopedWriteRWLock write_lock{rw_lock, ProfileEvents::DictCacheLockWriteNs}; cache_storage_ptr->insertColumnsForKeys(found_keys_in_source, fetched_columns_during_update); cache_storage_ptr->insertDefaultKeys(not_found_keys_in_source); error_count = 0; last_exception = std::exception_ptr{}; backoff_end_time = std::chrono::system_clock::time_point{}; } ProfileEvents::increment(ProfileEvents::DictCacheRequestTimeNs, watch.elapsed()); } catch (...) { /// Lock just for last_exception safety ProfilingScopedWriteRWLock write_lock{rw_lock, ProfileEvents::DictCacheLockWriteNs}; ++error_count; last_exception = std::current_exception(); backoff_end_time = now + std::chrono::seconds(calculateDurationWithBackoff(rnd_engine, error_count)); tryLogException(last_exception, log, "Could not update cache dictionary '" + getDictionaryID().getNameForLogs() + "', next update is scheduled at " + to_string(backoff_end_time.load())); try { std::rethrow_exception(last_exception); } catch (...) { throw DB::Exception(ErrorCodes::CACHE_DICTIONARY_UPDATE_FAIL, "Update failed for dictionary {} : {}", getDictionaryID().getNameForLogs(), getCurrentExceptionMessage(true /*with stack trace*/, true /*check embedded stack trace*/)); } } /// The underlying source can have duplicates, so count only unique keys this formula is used. size_t found_keys_size = requested_keys_size - not_found_keys.size(); ProfileEvents::increment(ProfileEvents::DictCacheKeysRequestedMiss, requested_keys_size - found_keys_size); ProfileEvents::increment(ProfileEvents::DictCacheKeysRequestedFound, found_keys_size); ProfileEvents::increment(ProfileEvents::DictCacheRequests); found_count.fetch_add(found_keys_size, std::memory_order_relaxed); } else { /// Won't request source for keys throw DB::Exception(ErrorCodes::CACHE_DICTIONARY_UPDATE_FAIL, "Query contains keys that are not present in cache or expired. Could not update cache dictionary {} now, because nearest update is scheduled at {}. Try again later.", getDictionaryID().getNameForLogs(), to_string(backoff_end_time.load())); } } template class CacheDictionary; template class CacheDictionary; }