Merge pull request #586 from Altinity/metadata_cache_for_parquet_24_12_2

Parquet File Metadata caching implementation

Author: Enmk

src/Common/ProfileEvents.cpp

@@ -968,7 +968,9 @@ The server successfully detected this situation and will download merged part fr
     M(FilterTransformPassedRows, "Number of rows that passed the filter in the query", ValueType::Number) \
     M(FilterTransformPassedBytes, "Number of bytes that passed the filter in the query", ValueType::Bytes) \
     M(QueryPreempted, "How many times tasks are paused and waiting due to 'priority' setting", ValueType::Number) \
-
+    M(ParquetFetchWaitTimeMicroseconds, "Time of waiting fetching parquet data", ValueType::Microseconds)                                                                                                                                                                                                                                               \
+    M(ParquetMetaDataCacheHits, "Number of times the read from filesystem cache hit the cache.", ValueType::Number) \
+    M(ParquetMetaDataCacheMisses, "Number of times the read from filesystem cache miss the cache.", ValueType::Number) \
 
 #ifdef APPLY_FOR_EXTERNAL_EVENTS
     #define APPLY_FOR_EVENTS(M) APPLY_FOR_BUILTIN_EVENTS(M) APPLY_FOR_EXTERNAL_EVENTS(M)

src/Core/FormatFactorySettings.h

@@ -1301,7 +1301,7 @@ Set the quoting rule for identifiers in SHOW CREATE query
     DECLARE(IdentifierQuotingStyle, show_create_query_identifier_quoting_style, IdentifierQuotingStyle::Backticks, R"(
 Set the quoting style for identifiers in SHOW CREATE query
 )", 0) \
-
+    DECLARE(Bool, input_format_parquet_use_metadata_cache, false, R"(Enable parquet file metadata caching)", 0) \
 // End of FORMAT_FACTORY_SETTINGS
 
 #define OBSOLETE_FORMAT_SETTINGS(M, ALIAS) \

src/Core/ServerSettings.cpp

@@ -1019,9 +1019,8 @@ namespace DB
     <wait_dictionaries_load_at_startup>true</wait_dictionaries_load_at_startup>
     ```
     )", 0) \
-    DECLARE(Bool, storage_shared_set_join_use_inner_uuid, false, "If enabled, an inner UUID is generated during the creation of SharedSet and SharedJoin. ClickHouse Cloud only", 0)
-
-
+    DECLARE(Bool, storage_shared_set_join_use_inner_uuid, false, "If enabled, an inner UUID is generated during the creation of SharedSet and SharedJoin. ClickHouse Cloud only", 0) \
+    DECLARE(UInt64, input_format_parquet_metadata_cache_max_size, 500000000, "Maximum size of parquet file metadata cache", 0) \
 // clang-format on
 
 /// If you add a setting which can be updated at runtime, please update 'changeable_settings' map in dumpToSystemServerSettingsColumns below

src/Processors/Formats/IInputFormat.h

@@ -67,6 +67,9 @@ class IInputFormat : public SourceWithKeyCondition
 
     void needOnlyCount() { need_only_count = true; }
 
+    /// Set additional info/key/id related to underlying storage of the ReadBuffer
+    virtual void setStorageRelatedUniqueKey(const ServerSettings & /* server_settings */, const Settings & /*settings*/, const String & /*key*/) {}
+
 protected:
     ReadBuffer & getReadBuffer() const { chassert(in); return *in; }
 

src/Processors/Formats/Impl/ParquetBlockInputFormat.cpp

@@ -3,6 +3,9 @@
 #if USE_PARQUET
 
 #include <Columns/ColumnNullable.h>
+#include <Core/Settings.h>
+#include <Core/ServerSettings.h>
+#include <Common/ProfileEvents.h>
 #include <Common/logger_useful.h>
 #include <Common/ThreadPool.h>
 #include <Formats/FormatFactory.h>
@@ -35,6 +38,8 @@
 namespace ProfileEvents
 {
     extern const Event ParquetFetchWaitTimeMicroseconds;
+    extern const Event ParquetMetaDataCacheHits;
+    extern const Event ParquetMetaDataCacheMisses;
 }
 
 namespace CurrentMetrics
@@ -51,6 +56,16 @@ namespace CurrentMetrics
 namespace DB
 {
 
+namespace Setting
+{
+    extern const SettingsBool input_format_parquet_use_metadata_cache;
+}
+
+namespace ServerSetting
+{
+    extern const ServerSettingsUInt64 input_format_parquet_metadata_cache_max_size;
+}
+
 namespace ErrorCodes
 {
     extern const int BAD_ARGUMENTS;
@@ -510,6 +525,58 @@ static std::vector<Range> getHyperrectangleForRowGroup(const parquet::FileMetaDa
     return hyperrectangle;
 }
 
+ParquetFileMetaDataCache::ParquetFileMetaDataCache(UInt64 max_size_bytes)
+    : CacheBase(max_size_bytes) {}
+
+ParquetFileMetaDataCache * ParquetFileMetaDataCache::instance(UInt64 max_size_bytes)
+{
+    static ParquetFileMetaDataCache instance(max_size_bytes);
+    return &instance;
+}
+
+std::shared_ptr<parquet::FileMetaData> ParquetBlockInputFormat::readMetadataFromFile()
+{
+    createArrowFileIfNotCreated();
+    return parquet::ReadMetaData(arrow_file);
+}
+
+std::shared_ptr<parquet::FileMetaData> ParquetBlockInputFormat::getFileMetaData()
+{
+    // in-memory cache is not implemented for local file operations, only for remote files
+    // there is a chance the user sets `input_format_parquet_use_metadata_cache=1` for a local file operation
+    // and the cache_key won't be set. Therefore, we also need to check for metadata_cache.key
+    if (!metadata_cache.use_cache || metadata_cache.key.empty())
+    {
+        return readMetadataFromFile();
+    }
+
+    auto [parquet_file_metadata, loaded] = ParquetFileMetaDataCache::instance(metadata_cache.max_size_bytes)->getOrSet(
+        metadata_cache.key,
+        [&]()
+        {
+            return readMetadataFromFile();
+        }
+    );
+    if (loaded)
+        ProfileEvents::increment(ProfileEvents::ParquetMetaDataCacheMisses);
+    else
+        ProfileEvents::increment(ProfileEvents::ParquetMetaDataCacheHits);
+    return parquet_file_metadata;
+}
+
+void ParquetBlockInputFormat::createArrowFileIfNotCreated()
+{
+    if (arrow_file)
+    {
+        return;
+    }
+
+    // Create arrow file adapter.
+    // TODO: Make the adapter do prefetching on IO threads, based on the full set of ranges that
+    //       we'll need to read (which we know in advance). Use max_download_threads for that.
+    arrow_file = asArrowFile(*in, format_settings, is_stopped, "Parquet", PARQUET_MAGIC_BYTES, /* avoid_buffering */ true);
+}
+
 std::unordered_set<std::size_t> getBloomFilterFilteringColumnKeys(const KeyCondition::RPN & rpn)
 {
     std::unordered_set<std::size_t> column_keys;
@@ -609,7 +676,7 @@ void ParquetBlockInputFormat::initializeIfNeeded()
     if (is_stopped)
         return;
 
-    metadata = parquet::ReadMetaData(arrow_file);
+    metadata = getFileMetaData();
     const bool prefetch_group = supportPrefetch();
 
     std::shared_ptr<arrow::Schema> schema;
@@ -709,6 +776,8 @@ void ParquetBlockInputFormat::initializeIfNeeded()
         }
     }
 
+    bool has_row_groups_to_read = false;
+
     auto skip_row_group_based_on_filters = [&](int row_group)
     {
         if (!format_settings.parquet.filter_push_down && !format_settings.parquet.bloom_filter_push_down)
@@ -758,9 +827,23 @@ void ParquetBlockInputFormat::initializeIfNeeded()
         row_group_batches.back().total_bytes_compressed += row_group_size;
         auto rows = adaptive_chunk_size(row_group);
         row_group_batches.back().adaptive_chunk_size = rows ? rows : format_settings.parquet.max_block_size;
+
+        has_row_groups_to_read = true;
+    }
+
+    if (has_row_groups_to_read)
+    {
+        createArrowFileIfNotCreated();
     }
 }
 
+void ParquetBlockInputFormat::setStorageRelatedUniqueKey(const ServerSettings & server_settings, const Settings & settings, const String & key_)
+{
+    metadata_cache.key = key_;
+    metadata_cache.use_cache = settings[Setting::input_format_parquet_use_metadata_cache];
+    metadata_cache.max_size_bytes = server_settings[ServerSetting::input_format_parquet_metadata_cache_max_size];
+}
+
 void ParquetBlockInputFormat::initializeRowGroupBatchReader(size_t row_group_batch_idx)
 {
     const bool row_group_prefetch = supportPrefetch();

src/Processors/Formats/Impl/ParquetBlockInputFormat.h

@@ -2,6 +2,7 @@
 #include "config.h"
 #if USE_PARQUET
 
+#include <Common/CacheBase.h>
 #include <Processors/Formats/IInputFormat.h>
 #include <Processors/Formats/ISchemaReader.h>
 #include <Formats/FormatSettings.h>
@@ -72,6 +73,8 @@ class ParquetBlockInputFormat : public IInputFormat
 
     size_t getApproxBytesReadForChunk() const override { return previous_approx_bytes_read_for_chunk; }
 
+    void setStorageRelatedUniqueKey(const ServerSettings & server_settings, const Settings & settings, const String & key_) override;
+
 private:
     Chunk read() override;
 
@@ -90,6 +93,11 @@ class ParquetBlockInputFormat : public IInputFormat
 
     void threadFunction(size_t row_group_batch_idx);
 
+    void createArrowFileIfNotCreated();
+    std::shared_ptr<parquet::FileMetaData> readMetadataFromFile();
+
+    std::shared_ptr<parquet::FileMetaData> getFileMetaData();
+
     inline bool supportPrefetch() const;
 
     // Data layout in the file:
@@ -338,6 +346,12 @@ class ParquetBlockInputFormat : public IInputFormat
     std::exception_ptr background_exception = nullptr;
     std::atomic<int> is_stopped{0};
     bool is_initialized = false;
+    struct Cache
+    {
+        String key;
+        bool use_cache = false;
+        UInt64 max_size_bytes{0};
+    } metadata_cache;
 };
 
 class ParquetSchemaReader : public ISchemaReader
@@ -356,6 +370,16 @@ class ParquetSchemaReader : public ISchemaReader
     std::shared_ptr<parquet::FileMetaData> metadata;
 };
 
+class ParquetFileMetaDataCache : public CacheBase<String, parquet::FileMetaData>
+{
+public:
+    static ParquetFileMetaDataCache *  instance(UInt64 max_size_bytes);
+    void clear() {}
+
+private:
+    ParquetFileMetaDataCache(UInt64 max_size_bytes);
+};
+
 }
 
 #endif

src/Storages/ObjectStorage/StorageObjectStorageSource.cpp

@@ -499,6 +499,9 @@ StorageObjectStorageSource::ReaderHolder StorageObjectStorageSource::createReade
         if (need_only_count)
             input_format->needOnlyCount();
 
+        if (!object_info->getPath().empty())
+            input_format->setStorageRelatedUniqueKey(context_->getServerSettings(), context_->getSettingsRef(), object_info->getPath() + ":" + object_info->metadata->etag);
+
         builder.init(Pipe(input_format));
 
         if (auto transformer = configuration->getSchemaTransformer(object_info->getPath()))

tests/queries/0_stateless/03299_parquet_object_storage_metadata_cache.reference

@@ -0,0 +1,3 @@
+10
+10
+10

tests/queries/0_stateless/03299_parquet_object_storage_metadata_cache.sql

@@ -0,0 +1,28 @@
+-- Tags: no-parallel, no-fasttest
+
+DROP TABLE IF EXISTS t_parquet_03262;
+
+CREATE TABLE t_parquet_03262 (a UInt64)
+ENGINE = S3(s3_conn, filename = 'test_03262_{_partition_id}', format = Parquet)
+PARTITION BY a;
+
+INSERT INTO t_parquet_03262 SELECT number FROM numbers(10) SETTINGS s3_truncate_on_insert=1;
+
+SELECT COUNT(*)
+FROM s3(s3_conn, filename = 'test_03262_*', format = Parquet)
+SETTINGS input_format_parquet_use_metadata_cache=1;
+
+SELECT COUNT(*)
+FROM s3(s3_conn, filename = 'test_03262_*', format = Parquet)
+SETTINGS input_format_parquet_use_metadata_cache=1, log_comment='test_03262_parquet_metadata_cache';
+
+SYSTEM FLUSH LOGS;
+
+SELECT ProfileEvents['ParquetMetaDataCacheHits']
+FROM system.query_log
+where log_comment = 'test_03262_parquet_metadata_cache'
+AND type = 'QueryFinish'
+ORDER BY event_time desc
+LIMIT 1;
+
+DROP TABLE t_parquet_03262;