ClickHouse/dbms/Storages/StorageLog.cpp

658 lines
21 KiB
C++
Raw Normal View History

#include <Storages/StorageLog.h>
#include <Storages/StorageFactory.h>
2011-11-05 23:31:19 +00:00
#include <Common/Exception.h>
#include <Common/StringUtils/StringUtils.h>
2019-12-23 16:57:16 +00:00
#include <Common/typeid_cast.h>
2010-03-18 19:32:14 +00:00
2020-02-14 14:28:33 +00:00
#include <Interpreters/evaluateConstantExpression.h>
#include <IO/ReadBufferFromFileBase.h>
#include <IO/WriteBufferFromFileBase.h>
2018-12-28 18:15:26 +00:00
#include <Compression/CompressedReadBuffer.h>
#include <Compression/CompressedWriteBuffer.h>
#include <IO/ReadHelpers.h>
#include <IO/WriteHelpers.h>
2012-01-09 19:20:48 +00:00
#include <DataTypes/NestedUtils.h>
2012-08-29 20:07:24 +00:00
#include <DataStreams/IBlockInputStream.h>
#include <DataStreams/IBlockOutputStream.h>
2015-01-18 08:25:56 +00:00
#include <Columns/ColumnArray.h>
2012-08-29 20:07:24 +00:00
#include <Interpreters/Context.h>
2020-02-14 14:28:33 +00:00
#include <Parsers/ASTLiteral.h>
#include "StorageLogSettings.h"
#include <Processors/Sources/SourceWithProgress.h>
#include <Processors/Pipe.h>
2010-03-18 19:32:14 +00:00
#define DBMS_STORAGE_LOG_DATA_FILE_EXTENSION ".bin"
#define DBMS_STORAGE_LOG_MARKS_FILE_NAME "__marks.mrk"
2012-01-09 19:20:48 +00:00
2010-03-18 19:32:14 +00:00
namespace DB
{
namespace ErrorCodes
{
extern const int LOGICAL_ERROR;
extern const int DUPLICATE_COLUMN;
extern const int SIZES_OF_MARKS_FILES_ARE_INCONSISTENT;
extern const int NUMBER_OF_ARGUMENTS_DOESNT_MATCH;
extern const int INCORRECT_FILE_NAME;
}
2010-03-18 19:32:14 +00:00
class LogSource final : public SourceWithProgress
2010-03-18 19:32:14 +00:00
{
2015-01-18 08:25:56 +00:00
public:
static Block getHeader(const NamesAndTypesList & columns)
{
Block res;
for (const auto & name_type : columns)
res.insert({ name_type.type->createColumn(), name_type.type, name_type.name });
return Nested::flatten(res);
}
LogSource(
size_t block_size_, const NamesAndTypesList & columns_, StorageLog & storage_,
size_t mark_number_, size_t rows_limit_, size_t max_read_buffer_size_)
: SourceWithProgress(getHeader(columns_)),
block_size(block_size_),
columns(columns_),
storage(storage_),
mark_number(mark_number_),
rows_limit(rows_limit_),
max_read_buffer_size(max_read_buffer_size_)
{
}
String getName() const override { return "Log"; }
2015-01-18 08:25:56 +00:00
protected:
Chunk generate() override;
2016-08-24 00:39:38 +00:00
2015-01-18 08:25:56 +00:00
private:
size_t block_size;
NamesAndTypesList columns;
StorageLog & storage;
size_t mark_number; /// from what mark to read data
size_t rows_limit; /// The maximum number of rows that can be read
size_t rows_read = 0;
size_t max_read_buffer_size;
struct Stream
{
2019-12-23 16:57:16 +00:00
Stream(const DiskPtr & disk, const String & data_path, size_t offset, size_t max_read_buffer_size_)
: plain(disk->readFile(data_path, std::min(max_read_buffer_size_, disk->getFileSize(data_path)))),
compressed(*plain)
{
if (offset)
plain->seek(offset, SEEK_SET);
}
2020-02-14 14:28:33 +00:00
std::unique_ptr<ReadBufferFromFileBase> plain;
CompressedReadBuffer compressed;
};
using FileStreams = std::map<String, Stream>;
FileStreams streams;
using DeserializeState = IDataType::DeserializeBinaryBulkStatePtr;
using DeserializeStates = std::map<String, DeserializeState>;
DeserializeStates deserialize_states;
void readData(const String & name, const IDataType & type, IColumn & column, size_t max_rows_to_read);
2015-01-18 08:25:56 +00:00
};
class LogBlockOutputStream final : public IBlockOutputStream
{
2015-01-18 08:25:56 +00:00
public:
explicit LogBlockOutputStream(StorageLog & storage_)
: storage(storage_),
lock(storage.rwlock),
marks_stream(storage.disk->writeFile(storage.marks_file_path, 4096, WriteMode::Rewrite))
{
}
~LogBlockOutputStream() override
{
try
{
writeSuffix();
}
catch (...)
{
tryLogCurrentException(__PRETTY_FUNCTION__);
}
}
Block getHeader() const override { return storage.getSampleBlock(); }
void write(const Block & block) override;
void writeSuffix() override;
2015-04-02 23:58:26 +00:00
2015-01-18 08:25:56 +00:00
private:
StorageLog & storage;
std::unique_lock<std::shared_mutex> lock;
bool done = false;
struct Stream
{
2019-12-23 16:57:16 +00:00
Stream(const DiskPtr & disk, const String & data_path, CompressionCodecPtr codec, size_t max_compress_block_size) :
plain(disk->writeFile(data_path, max_compress_block_size, WriteMode::Append)),
compressed(*plain, std::move(codec), max_compress_block_size),
2019-12-23 16:57:16 +00:00
plain_offset(disk->getFileSize(data_path))
{
}
std::unique_ptr<WriteBuffer> plain;
CompressedWriteBuffer compressed;
size_t plain_offset; /// How many bytes were in the file at the time the LogBlockOutputStream was created.
void finalize()
{
compressed.next();
plain->next();
}
};
2017-11-26 19:22:33 +00:00
using Mark = StorageLog::Mark;
2017-08-30 18:13:32 +00:00
using MarksForColumns = std::vector<std::pair<size_t, Mark>>;
using FileStreams = std::map<String, Stream>;
FileStreams streams;
using WrittenStreams = std::set<String>;
std::unique_ptr<WriteBuffer> marks_stream; /// Declared below `lock` to make the file open when rwlock is captured.
using SerializeState = IDataType::SerializeBinaryBulkStatePtr;
using SerializeStates = std::map<String, SerializeState>;
SerializeStates serialize_states;
IDataType::OutputStreamGetter createStreamGetter(const String & name, WrittenStreams & written_streams);
void writeData(const String & name, const IDataType & type, const IColumn & column,
MarksForColumns & out_marks,
WrittenStreams & written_streams);
void writeMarks(MarksForColumns && marks);
2015-01-18 08:25:56 +00:00
};
Chunk LogSource::generate()
2010-03-18 19:32:14 +00:00
{
Block res;
if (rows_read == rows_limit)
return {};
if (storage.disk->isDirectoryEmpty(storage.table_path))
return {};
/// How many rows to read for the next block.
size_t max_rows_to_read = std::min(block_size, rows_limit - rows_read);
for (const auto & name_type : columns)
{
MutableColumnPtr column = name_type.type->createColumn();
try
{
readData(name_type.name, *name_type.type, *column, max_rows_to_read);
}
catch (Exception & e)
{
e.addMessage("while reading column " + name_type.name + " at " + fullPath(storage.disk, storage.table_path));
throw;
}
if (!column->empty())
res.insert(ColumnWithTypeAndName(std::move(column), name_type.type, name_type.name));
}
if (res)
rows_read += res.rows();
if (!res || rows_read == rows_limit)
{
/** Close the files (before destroying the object).
* When many sources are created, but simultaneously reading only a few of them,
* buffers don't waste memory.
*/
streams.clear();
}
res = Nested::flatten(res);
UInt64 num_rows = res.rows();
return Chunk(res.getColumns(), num_rows);
2010-03-18 19:32:14 +00:00
}
void LogSource::readData(const String & name, const IDataType & type, IColumn & column, size_t max_rows_to_read)
2012-08-29 20:07:24 +00:00
{
IDataType::DeserializeBinaryBulkSettings settings; /// TODO Use avg_value_size_hint.
auto create_string_getter = [&](bool stream_for_prefix)
{
2018-09-02 02:27:56 +00:00
return [&, stream_for_prefix] (const IDataType::SubstreamPath & path) -> ReadBuffer *
{
String stream_name = IDataType::getFileNameForStream(name, path);
const auto & file_it = storage.files.find(stream_name);
if (storage.files.end() == file_it)
throw Exception("Logical error: no information about file " + stream_name + " in StorageLog", ErrorCodes::LOGICAL_ERROR);
UInt64 offset = 0;
if (!stream_for_prefix && mark_number)
offset = file_it->second.marks[mark_number].offset;
auto & data_file_path = file_it->second.data_file_path;
2019-12-23 16:57:16 +00:00
auto it = streams.try_emplace(stream_name, storage.disk, data_file_path, offset, max_read_buffer_size).first;
return &it->second.compressed;
};
};
if (deserialize_states.count(name) == 0)
{
settings.getter = create_string_getter(true);
type.deserializeBinaryBulkStatePrefix(settings, deserialize_states[name]);
}
settings.getter = create_string_getter(false);
type.deserializeBinaryBulkWithMultipleStreams(column, max_rows_to_read, settings, deserialize_states[name]);
2012-08-29 20:07:24 +00:00
}
2010-03-18 19:32:14 +00:00
void LogBlockOutputStream::write(const Block & block)
{
storage.check(block, true);
/// The set of written offset columns so that you do not write shared offsets of columns for nested structures multiple times
WrittenStreams written_streams;
2010-03-18 19:32:14 +00:00
MarksForColumns marks;
marks.reserve(storage.file_count);
for (size_t i = 0; i < block.columns(); ++i)
{
const ColumnWithTypeAndName & column = block.safeGetByPosition(i);
writeData(column.name, *column.type, *column.column, marks, written_streams);
}
writeMarks(std::move(marks));
2012-08-29 20:07:24 +00:00
}
2013-09-15 01:40:29 +00:00
void LogBlockOutputStream::writeSuffix()
{
if (done)
return;
done = true;
2015-04-02 23:58:26 +00:00
WrittenStreams written_streams;
IDataType::SerializeBinaryBulkSettings settings;
for (const auto & column : getHeader())
{
auto it = serialize_states.find(column.name);
if (it != serialize_states.end())
{
settings.getter = createStreamGetter(column.name, written_streams);
column.type->serializeBinaryBulkStateSuffix(settings, it->second);
}
}
/// Finish write.
marks_stream->next();
2013-09-15 01:40:29 +00:00
2017-11-26 19:22:33 +00:00
for (auto & name_stream : streams)
name_stream.second.finalize();
2013-09-15 01:40:29 +00:00
Strings column_files;
2017-11-26 19:22:33 +00:00
for (const auto & name_stream : streams)
column_files.push_back(storage.files[name_stream.first].data_file_path);
column_files.push_back(storage.marks_file_path);
storage.file_checker.update(column_files.begin(), column_files.end());
streams.clear();
2013-09-15 01:40:29 +00:00
}
IDataType::OutputStreamGetter LogBlockOutputStream::createStreamGetter(const String & name,
WrittenStreams & written_streams)
{
return [&] (const IDataType::SubstreamPath & path) -> WriteBuffer *
{
String stream_name = IDataType::getFileNameForStream(name, path);
if (written_streams.count(stream_name))
return nullptr;
auto it = streams.find(stream_name);
if (streams.end() == it)
throw Exception("Logical error: stream was not created when writing data in LogBlockOutputStream",
ErrorCodes::LOGICAL_ERROR);
return &it->second.compressed;
};
}
void LogBlockOutputStream::writeData(const String & name, const IDataType & type, const IColumn & column,
2019-12-23 16:57:16 +00:00
MarksForColumns & out_marks, WrittenStreams & written_streams)
2012-08-29 20:07:24 +00:00
{
IDataType::SerializeBinaryBulkSettings settings;
type.enumerateStreams([&] (const IDataType::SubstreamPath & path)
{
String stream_name = IDataType::getFileNameForStream(name, path);
if (written_streams.count(stream_name))
return;
const auto & columns = storage.getColumns();
streams.try_emplace(
stream_name,
2019-12-23 16:57:16 +00:00
storage.disk,
storage.files[stream_name].data_file_path,
columns.getCodecOrDefault(name),
storage.max_compress_block_size);
}, settings.path);
settings.getter = createStreamGetter(name, written_streams);
if (serialize_states.count(name) == 0)
type.serializeBinaryBulkStatePrefix(settings, serialize_states[name]);
type.enumerateStreams([&] (const IDataType::SubstreamPath & path)
2017-11-26 19:22:33 +00:00
{
String stream_name = IDataType::getFileNameForStream(name, path);
if (written_streams.count(stream_name))
return;
2017-11-26 19:22:33 +00:00
const auto & file = storage.files[stream_name];
const auto stream_it = streams.find(stream_name);
Mark mark;
mark.rows = (file.marks.empty() ? 0 : file.marks.back().rows) + column.size();
mark.offset = stream_it->second.plain_offset + stream_it->second.plain->count();
out_marks.emplace_back(file.column_index, mark);
}, settings.path);
2017-11-26 19:22:33 +00:00
type.serializeBinaryBulkWithMultipleStreams(column, 0, 0, settings, serialize_states[name]);
type.enumerateStreams([&] (const IDataType::SubstreamPath & path)
{
String stream_name = IDataType::getFileNameForStream(name, path);
if (!written_streams.emplace(stream_name).second)
return;
2017-11-26 19:22:33 +00:00
auto it = streams.find(stream_name);
if (streams.end() == it)
throw Exception("Logical error: stream was not created when writing data in LogBlockOutputStream", ErrorCodes::LOGICAL_ERROR);
it->second.compressed.next();
}, settings.path);
2010-03-18 19:32:14 +00:00
}
void LogBlockOutputStream::writeMarks(MarksForColumns && marks)
{
if (marks.size() != storage.file_count)
throw Exception("Wrong number of marks generated from block. Makes no sense.", ErrorCodes::LOGICAL_ERROR);
std::sort(marks.begin(), marks.end(), [](const auto & a, const auto & b) { return a.first < b.first; });
2017-11-26 19:22:33 +00:00
for (const auto & mark : marks)
{
writeIntBinary(mark.second.rows, *marks_stream);
writeIntBinary(mark.second.offset, *marks_stream);
2017-11-26 19:22:33 +00:00
size_t column_index = mark.first;
storage.files[storage.column_names_by_idx[column_index]].marks.push_back(mark.second);
}
}
StorageLog::StorageLog(
DiskPtr disk_,
const String & relative_path_,
2019-12-04 16:06:55 +00:00
const StorageID & table_id_,
const ColumnsDescription & columns_,
2019-08-24 21:20:20 +00:00
const ConstraintsDescription & constraints_,
size_t max_compress_block_size_)
2019-12-04 16:06:55 +00:00
: IStorage(table_id_)
, disk(std::move(disk_))
, table_path(relative_path_)
2019-12-04 16:06:55 +00:00
, max_compress_block_size(max_compress_block_size_)
, file_checker(disk, table_path + "sizes.json")
2010-03-18 19:32:14 +00:00
{
2019-08-24 21:20:20 +00:00
setColumns(columns_);
setConstraints(constraints_);
2019-10-25 19:07:47 +00:00
if (relative_path_.empty())
throw Exception("Storage " + getName() + " requires data path", ErrorCodes::INCORRECT_FILE_NAME);
/// create directories if they do not exist
disk->createDirectories(table_path);
2012-01-10 22:11:51 +00:00
2018-03-13 15:00:28 +00:00
for (const auto & column : getColumns().getAllPhysical())
addFiles(column.name, *column.type);
marks_file_path = table_path + DBMS_STORAGE_LOG_MARKS_FILE_NAME;
2012-08-29 20:07:24 +00:00
}
void StorageLog::addFiles(const String & column_name, const IDataType & type)
2012-08-29 20:07:24 +00:00
{
if (files.end() != files.find(column_name))
throw Exception("Duplicate column with name " + column_name + " in constructor of StorageLog.",
ErrorCodes::DUPLICATE_COLUMN);
IDataType::StreamCallback stream_callback = [&] (const IDataType::SubstreamPath & substream_path)
{
String stream_name = IDataType::getFileNameForStream(column_name, substream_path);
if (!files.count(stream_name))
{
ColumnData & column_data = files[stream_name];
column_data.column_index = file_count;
column_data.data_file_path = table_path + stream_name + DBMS_STORAGE_LOG_DATA_FILE_EXTENSION;
column_names_by_idx.push_back(stream_name);
++file_count;
}
};
IDataType::SubstreamPath substream_path;
type.enumerateStreams(stream_callback, substream_path);
2012-06-21 16:33:00 +00:00
}
2012-01-10 22:11:51 +00:00
2012-06-21 16:33:00 +00:00
void StorageLog::loadMarks()
{
std::unique_lock<std::shared_mutex> lock(rwlock);
if (loaded_marks)
return;
using FilesByIndex = std::vector<Files::iterator>;
FilesByIndex files_by_index(file_count);
for (Files::iterator it = files.begin(); it != files.end(); ++it)
files_by_index[it->second.column_index] = it;
if (disk->exists(marks_file_path))
{
size_t file_size = disk->getFileSize(marks_file_path);
if (file_size % (file_count * sizeof(Mark)) != 0)
throw Exception("Size of marks file is inconsistent", ErrorCodes::SIZES_OF_MARKS_FILES_ARE_INCONSISTENT);
size_t marks_count = file_size / (file_count * sizeof(Mark));
2017-11-26 19:22:33 +00:00
for (auto & file : files_by_index)
file->second.marks.reserve(marks_count);
2019-12-26 14:28:22 +00:00
std::unique_ptr<ReadBuffer> marks_rb = disk->readFile(marks_file_path, 32768);
while (!marks_rb->eof())
{
2020-03-09 01:22:33 +00:00
for (auto & file : files_by_index)
{
Mark mark;
readIntBinary(mark.rows, *marks_rb);
readIntBinary(mark.offset, *marks_rb);
2020-03-09 01:22:33 +00:00
file->second.marks.push_back(mark);
}
}
}
loaded_marks = true;
2010-03-18 19:32:14 +00:00
}
void StorageLog::rename(const String & new_path_to_table_data, const String & new_database_name, const String & new_table_name, TableStructureWriteLockHolder &)
2012-06-18 06:19:13 +00:00
{
std::unique_lock<std::shared_mutex> lock(rwlock);
disk->moveDirectory(table_path, new_path_to_table_data);
table_path = new_path_to_table_data;
file_checker.setPath(table_path + "sizes.json");
2012-06-18 06:19:13 +00:00
2017-11-26 19:22:33 +00:00
for (auto & file : files)
file.second.data_file_path = table_path + fileName(file.second.data_file_path);
marks_file_path = table_path + DBMS_STORAGE_LOG_MARKS_FILE_NAME;
2019-12-03 16:25:32 +00:00
renameInMemory(new_database_name, new_table_name);
2012-06-18 06:19:13 +00:00
}
2019-08-27 20:43:08 +00:00
void StorageLog::truncate(const ASTPtr &, const Context &, TableStructureWriteLockHolder &)
2018-04-21 00:35:20 +00:00
{
std::shared_lock<std::shared_mutex> lock(rwlock);
2018-06-09 15:48:22 +00:00
files.clear();
file_count = 0;
loaded_marks = false;
2018-04-21 00:35:20 +00:00
disk->clearDirectory(table_path);
2018-04-21 00:35:20 +00:00
2018-11-26 00:56:50 +00:00
for (const auto & column : getColumns().getAllPhysical())
2018-04-21 00:35:20 +00:00
addFiles(column.name, *column.type);
file_checker = FileChecker{disk, table_path + "sizes.json"};
marks_file_path = table_path + DBMS_STORAGE_LOG_MARKS_FILE_NAME;
2018-04-21 00:35:20 +00:00
}
2012-06-18 06:19:13 +00:00
2017-11-26 19:22:33 +00:00
const StorageLog::Marks & StorageLog::getMarksWithRealRowCount() const
{
/// There should be at least one physical column
2020-02-20 15:40:23 +00:00
const String column_name = getColumns().getAllPhysical().begin()->name;
const auto column_type = getColumns().getAllPhysical().begin()->type;
String filename;
/** We take marks from first column.
* If this is a data type with multiple stream, get the first stream, that we assume have real row count.
* (Example: for Array data type, first stream is array sizes; and number of array sizes is the number of arrays).
*/
IDataType::SubstreamPath substream_root_path;
2020-02-20 15:40:23 +00:00
column_type->enumerateStreams([&](const IDataType::SubstreamPath & substream_path)
{
if (filename.empty())
filename = IDataType::getFileNameForStream(column_name, substream_path);
}, substream_root_path);
Files::const_iterator it = files.find(filename);
if (files.end() == it)
throw Exception("Cannot find file " + filename, ErrorCodes::LOGICAL_ERROR);
return it->second.marks;
}
Pipes StorageLog::read(
const Names & column_names,
2017-12-01 21:13:25 +00:00
const SelectQueryInfo & /*query_info*/,
const Context & context,
QueryProcessingStage::Enum /*processed_stage*/,
size_t max_block_size,
2017-06-02 15:54:39 +00:00
unsigned num_streams)
2010-03-18 19:32:14 +00:00
{
check(column_names);
loadMarks();
2018-03-13 15:00:28 +00:00
NamesAndTypesList all_columns = Nested::collect(getColumns().getAllPhysical().addTypes(column_names));
std::shared_lock<std::shared_mutex> lock(rwlock);
Pipes pipes;
const Marks & marks = getMarksWithRealRowCount();
size_t marks_size = marks.size();
2017-06-02 15:54:39 +00:00
if (num_streams > marks_size)
num_streams = marks_size;
size_t max_read_buffer_size = context.getSettingsRef().max_read_buffer_size;
for (size_t stream = 0; stream < num_streams; ++stream)
{
size_t mark_begin = stream * marks_size / num_streams;
size_t mark_end = (stream + 1) * marks_size / num_streams;
size_t rows_begin = mark_begin ? marks[mark_begin - 1].rows : 0;
size_t rows_end = mark_end ? marks[mark_end - 1].rows : 0;
pipes.emplace_back(std::make_shared<LogSource>(
max_block_size,
all_columns,
*this,
mark_begin,
rows_end - rows_begin,
max_read_buffer_size));
}
return pipes;
2010-03-18 19:32:14 +00:00
}
2011-08-28 02:22:23 +00:00
BlockOutputStreamPtr StorageLog::write(
const ASTPtr & /*query*/, const Context & /*context*/)
2010-03-18 19:32:14 +00:00
{
loadMarks();
return std::make_shared<LogBlockOutputStream>(*this);
2010-03-18 19:32:14 +00:00
}
2019-07-03 13:17:19 +00:00
CheckResults StorageLog::checkData(const ASTPtr & /* query */, const Context & /* context */)
{
std::shared_lock<std::shared_mutex> lock(rwlock);
return file_checker.check();
}
2011-11-05 23:31:19 +00:00
void registerStorageLog(StorageFactory & factory)
{
StorageFactory::StorageFeatures features{
.supports_settings = true
};
factory.registerStorage("Log", [](const StorageFactory::Arguments & args)
{
if (!args.engine_args.empty())
throw Exception(
"Engine " + args.engine_name + " doesn't support any arguments (" + toString(args.engine_args.size()) + " given)",
ErrorCodes::NUMBER_OF_ARGUMENTS_DOESNT_MATCH);
String disk_name = getDiskName(*args.storage_def);
DiskPtr disk = args.context.getDisk(disk_name);
2020-02-14 14:28:33 +00:00
return StorageLog::create(
2020-02-14 14:28:33 +00:00
disk, args.relative_data_path, args.table_id, args.columns, args.constraints,
args.context.getSettings().max_compress_block_size);
}, features);
}
2010-03-18 19:32:14 +00:00
}