mirror of
https://github.com/ClickHouse/ClickHouse.git
synced 2024-11-22 23:52:03 +00:00
62 lines
2.9 KiB
C++
62 lines
2.9 KiB
C++
#pragma once
|
|
|
|
#include <mysqlxx/Connection.h>
|
|
#include <mysqlxx/Transaction.h>
|
|
#include <mysqlxx/Manip.h>
|
|
#include <mysqlxx/Pool.h>
|
|
#include <common/LocalDate.h>
|
|
#include <common/LocalDateTime.h>
|
|
#include <mysqlxx/Null.h>
|
|
|
|
|
|
/** 'mysqlxx' - very simple library for replacement of 'mysql++' library.
|
|
*
|
|
* For whatever reason, in Yandex.Metrica, back in 2008, 'mysql++' library was used.
|
|
* There are the following shortcomings of 'mysql++':
|
|
* 1. Too rich functionality: most of it is not used.
|
|
* 2. Low performance (when used for Yandex.Metrica).
|
|
*
|
|
* Low performance is caused by the following reasons:
|
|
*
|
|
* 1. Excessive copying: 'mysqlpp::Row' works like 'std::vector<std::string>'.
|
|
* Content of MYSQL_ROW is copied inside it.
|
|
* But MYSQL_ROW is a 'char**', that is allocated in single piece,
|
|
* where values are stored consecutively as (non-zero-terminated) strings.
|
|
*
|
|
* 2. Too slow methods for converting values to numbers.
|
|
* In mysql++, it is done through std::stringstream.
|
|
* This is slower than POSIX functions (strtoul, etc).
|
|
* In turn, this is slower than simple hand-coded functions,
|
|
* that doesn't respect locales and unused by MySQL number representations.
|
|
*
|
|
* 3. Too slow methods of escaping and quoting.
|
|
* In mysql++, 'mysql_real_escape_string' is used, that works correct
|
|
* even for charsets, that are not based on ASCII (examples: UTF-16, Shift-JIS).
|
|
* But when using charsets based on ASCII, as UTF-8,
|
|
* (in general, charsets, where escape characters are represented in same way as in ASCII,
|
|
* and where that codes cannot appear in representation of another characters)
|
|
* this function is redundant.
|
|
*
|
|
* 4. Too much garbage (dynamic_cast, typeid when converting values).
|
|
*
|
|
* Low performance cause the following effects:
|
|
* 1. In sequential read from MySQL table, the client CPU becomes a bottleneck, while MySQL server is not loaded.
|
|
* When using bare MySQL C API, this doesn't happen.
|
|
* 2. Sequential read from MySQL is lower than 30MB/s.
|
|
*
|
|
* Warning!
|
|
*
|
|
* mysqlxx is implemented as very simple wrapper around MySQL C API,
|
|
* and implements only limited subset of mysql++ interface, that we use.
|
|
* And for the sake of simplicity, some functions work only with certain assumptions,
|
|
* or with slightly different semantic than in mysql++.
|
|
* And we don't care about cross-platform usage of mysqlxx.
|
|
* These assumptions are specific for Yandex.Metrica. Your mileage may vary.
|
|
*
|
|
* mysqlxx could not be considered as separate full-featured library,
|
|
* because it is developed from the principle - "everything that we don't need is not implemented",
|
|
* and also the library depends on some other libraries from Yandex.Metrica code.
|
|
* (dependencied could be easily removed if necessary).
|
|
* It is assumed that the user will add all missing functionality that is needed.
|
|
*/
|