database/doris
2
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

[opt](FileReader) InMemoryReader is only used in s3 (#23486)

Browse Source 
If file size < 8MB, the file will be read into memory, and this idea is from https://github.com/apache/hadoop/blob/trunk/hadoop-tools/hadoop-aws/src/site/markdown/tools/hadoop-aws/prefetching.md#s3inmemoryinputstream. However, in some cases, we only read one or two columns in a file, and the actually required bytes is only 1%, resulting in a multiple fold increase in the amount of data read. Therefore, `InMemoryReader` can only used in object storage, and reduce the threshold.
This commit is contained in:
Ashin Gau
2023-08-30 20:43:39 +08:00
committed by GitHub
parent b7404896fa
commit 449c595f9d
4 changed files with 13 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
be/src/common/config.cpp
View File

@ -850,6 +850,9 @@ DEFINE_Validator(jsonb_type_length_soft_limit_bytes,
// is greater than object_pool_buffer_size, release the object in the unused_object_pool.
DEFINE_Int32(object_pool_buffer_size, "100");
// Threshold of reading a small file into memory
DEFINE_mInt32(in_memory_file_size, "1048576"); // 1MB
// ParquetReaderWrap prefetch buffer size
DEFINE_Int32(parquet_reader_max_buffer_size, "50");
// Max size of parquet page header in bytes

3
be/src/common/config.h
View File

@ -897,6 +897,9 @@ DECLARE_mInt32(jsonb_type_length_soft_limit_bytes);
// is greater than object_pool_buffer_size, release the object in the unused_object_pool.
DECLARE_Int32(object_pool_buffer_size);
// Threshold fo reading a small file into memory
DECLARE_mInt32(in_memory_file_size);
// ParquetReaderWrap prefetch buffer size
DECLARE_Int32(parquet_reader_max_buffer_size);
// Max size of parquet page header in bytes

8
be/src/io/fs/buffered_reader.cpp
View File

@ -778,8 +778,12 @@ Status DelegateReader::create_file_reader(RuntimeProfile* profile,
io::FileReaderSPtr reader;
RETURN_IF_ERROR(FileFactory::create_file_reader(system_properties, file_description,
reader_options, file_system, &reader, profile));
if (reader->size() < IN_MEMORY_FILE_SIZE) {
*file_reader = std::make_shared<InMemoryFileReader>(reader);
if (reader->size() < config::in_memory_file_size) {
if (typeid_cast<io::S3FileReader*>(reader.get())) {
*file_reader = std::make_shared<InMemoryFileReader>(reader);
} else {
*file_reader = std::move(reader);
}
} else if (access_mode == AccessMode::SEQUENTIAL) {
bool is_thread_safe = false;
if (typeid_cast<io::S3FileReader*>(reader.get())) {

4
be/src/io/fs/buffered_reader.h
View File

@ -238,7 +238,7 @@ private:
/**
* Create a file reader suitable for accessing scenarios:
* 1. When file size < 8MB, create InMemoryFileReader file reader
* 1. When file size < config::in_memory_file_size, create InMemoryFileReader file reader
* 2. When reading sequential file(csv/json), create PrefetchBufferedReader
* 3. When reading random access file(parquet/orc), create normal file reader
*/
@ -246,8 +246,6 @@ class DelegateReader {
public:
enum AccessMode { SEQUENTIAL, RANDOM };
static constexpr size_t IN_MEMORY_FILE_SIZE = 8 * 1024 * 1024;
static Status create_file_reader(
RuntimeProfile* profile, const FileSystemProperties& system_properties,
const FileDescription& file_description, const io::FileReaderOptions& reader_options,