为什么通过ORM 5-8x加载SQLAlchemy对象要比通过原始MySQLdb光标加载行慢？

Question

我注意到SQLAlchemy获取（和ORMing）一些数据的速度很慢，使用裸SQL获取数据的速度相当快。首先，我创建了一个包含一百万条记录的数据库：

mysql> use foo
mysql> describe Foo;
+-------+---------+------+-----+---------+-------+
| Field | Type    | Null | Key | Default | Extra |
+-------+---------+------+-----+---------+-------+
| id    | int(11) | NO   | PRI | NULL    |       |
| A     | int(11) | NO   |     | NULL    |       |
| B     | int(11) | NO   |     | NULL    |       |
| C     | int(11) | NO   |     | NULL    |       |
+-------+---------+------+-----+---------+-------+
mysql> SELECT COUNT(*) FROM Foo;
+----------+
| COUNT(*) |
+----------+
|  1000000 |
+----------+
mysql> 

作为一个粗略的测试，查询所有Foo大约需要2秒：

herbert@dev0 ~ $ date; echo 'use foo; select * from Foo;' | mysql -uroot -pxxx > /dev/null; date
zo apr 20 18:48:49 CEST 2014
zo apr 20 18:48:51 CEST 2014

如果我使用MySQLdb在python中执行此操作，大约需要3秒，包括Foo对象的构造：

herbert@dev0 ~ $ python BareORM.py 
query execution time:  0:00:02.198986
total time:  0:00:03.403084

其输出为：

#!/usr/bin/python
# -*- coding: utf-8 -*-

import MySQLdb
import sys
import time
import datetime

class Foo:
    def __init__(self, a, b, c):
        self.a=a; self.b=b; self.c=c;

try:
    start = datetime.datetime.now()
    con = MySQLdb.connect('localhost', 'root', 'xxx', 'foo')
    cur = con.cursor();

    cur.execute("""SELECT * FROM Foo LIMIT 1000000""")
    print "query execution time: ", datetime.datetime.now()-start
    foos = [];
    for elem in cur:
        foos.append(Foo(elem[1], elem[2], elem[3]))
    con.commit()

except MySQLdb.Error, e:
    print "Error %d: %s" % (e.args[0], e.args[1])
    sys.exit(1)

finally:
    if con: con.close()
    print "total time: ",  datetime.datetime.now()-start

但是，使用SQLAlchemy来减少样板代码，大约需要25秒来完成相同的工作：

herbert@dev0 ~ $ python AlchemyORM.py 
total time:  0:00:24.649279

使用此代码：

import sqlalchemy
import datetime
import MySQLdb

from sqlalchemy import Column, Integer, create_engine
from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy.orm import sessionmaker, relationship, backref

Base = declarative_base()

class Foo(Base):
    __tablename__ = 'Foo'
    id = Column(Integer, primary_key=True)
    A  = Column(Integer(unsigned=False), nullable=False)
    B  = Column(Integer(unsigned=False), nullable=False)
    C  = Column(Integer(unsigned=False), nullable=False)

engine  = create_engine('mysql+mysqldb://root:xxx@localhost/foo')
Session = sessionmaker(bind=engine)
session = Session()
start = datetime.datetime.now()
foos  = session.query(Foo).limit(1000000).all()
print "total time: ", datetime.datetime.now()-start

假设SQLAlchemy应该做大致相同的事情，为什么SQLAlchemy的运行速度比裸SQL解决方案慢10倍？我能加快速度吗？

这是一个更复杂查询的最小工作示例，它使用紧急加载连接多个表。我正在考虑只对一个表进行简单的查询，然后使用字典创建id->；对象映射并整理一对多关系。但是在这样做之前，我想确定SQLAlchemy不能表现得更好，因为从软件设计的角度来看，编写自己的ORM是一个坏主意。嗯，2倍的减速是可以接受的（也许）。

如果您知道其他（更快的）python SQL ORM，或者类似BigTable的解决方案（已经是ORM了），可以在评论中提到它们。

编辑：还尝试了Peewee，结果约15秒

from peewee import *
import datetime;
database = MySQLDatabase("foo", host="localhost", port=3306, user="root", passwd="xxx")

class Foo(Model):
        id = IntegerField()
        A  = IntegerField()
        B  = IntegerField()
        C  = IntegerField()

        class Meta:
                db_table = 'Foo'
                database = database

start = datetime.datetime.now()
foos = Foo.select()
cnt=0;
for i in foos: cnt=cnt+1
print "total time: ", datetime.datetime.now() - start

编辑：作为对Matthias的回应，我尝试在Java中使用Hibernate做同样的事情，结果大约是8到10秒，不是很快，但比25秒快得多。代码，从一些类开始，以一些配置结束：

package herbert.hibernateorm;

import java.util.List;

import org.hibernate.Session; 
import org.hibernate.Transaction;
import org.hibernate.SessionFactory;
import org.hibernate.cfg.Configuration;

public class App {
   public static void main(String[] args) throws Exception {
      SessionFactory factory = new Configuration().configure().buildSessionFactory();
      Session session = factory.openSession();
      Transaction tx = session.beginTransaction();
      long start = System.currentTimeMillis();
      List foos = session.createQuery("FROM Foo").list(); 
      System.out.println(foos.size());
      System.out.printf("total time: %d\n", System.currentTimeMillis() - start);
      session.close();
   }
}

package herbert.hibernateorm;

public class Foo {
    private int id, a, b, c;
    public Foo() {}
    public Foo(int A, int B, int C) { this.a=A; this.b=B; this.c=C; }

    public int getId() { return id; }
    public void setId(int id) { this.id = id; }
    public int getA() { return a; }
    public void setA(int a) { this.a = a; }
    public int getB() { return b; }
    public void setB(int b) { this.b = b; }
    public int getC() { return c; }
    public void setC(int c) { this.c = c; }
}

配置（分别为hibernate.cfg.xml和hibernate.hbm.xml）

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE hibernate-configuration PUBLIC "-//Hibernate/Hibernate Configuration DTD 3.0//EN" "http://hibernate.sourceforge.net/hibernate-configuration-3.0.dtd">
<hibernate-configuration>
  <session-factory>
    <property name="hibernate.dialect">org.hibernate.dialect.MySQLDialect</property>
    <property name="hibernate.connection.driver_class">com.mysql.jdbc.Driver</property>
    <property name="hibernate.connection.url">jdbc:mysql://localhost:3306/foo?zeroDateTimeBehavior=convertToNull</property>
    <property name="hibernate.connection.username">root</property>
    <property name="hibernate.connection.password">xxx</property>
    <mapping resource="hibernate.hbm.xml"/>
  </session-factory>
</hibernate-configuration>

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE hibernate-mapping PUBLIC "-//Hibernate/Hibernate Mapping DTD 3.0//EN" "http://hibernate.sourceforge.net/hibernate-mapping-3.0.dtd">
<hibernate-mapping>
    <class name="herbert.hibernateorm.Foo" table="Foo" catalog="foo">
        <id name="id" type="int">
            <column name="id" />
            <generator class="assigned" />
        </id>
        <property name="a" type="int">
            <column name="A" not-null="true" />
        </property>
        <property name="b" type="int">
            <column name="B" not-null="true" />
        </property>
        <property name="c" type="int">
            <column name="C" not-null="true" />
        </property>
    </class>
</hibernate-mapping>

最后是在maven中运行的pom文件：

<project xmlns="http://maven.apache.org/POM/4.0.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
         xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd">
    <modelVersion>4.0.0</modelVersion>

    <groupId>herbert</groupId>
    <artifactId>hibernateORM</artifactId>
    <version>1.0-SNAPSHOT</version>
    <packaging>jar</packaging>

    <name>hibernateORM</name>
    <url>http://maven.apache.org</url>
    <repositories>
        <repository>
            <id>unknown-jars-temp-repo</id>
            <name>A temporary repository created by NetBeans for libraries and jars it could not identify. Please replace the dependencies in this repository with correct ones and delete this repository.</name>
            <url>file:${project.basedir}/lib</url>
        </repository>
    </repositories>
    <properties>
        <project.build.sourceEncoding>UTF-8</project.build.sourceEncoding>
    </properties>

    <dependencies>
        <dependency>
            <groupId>junit</groupId>
            <artifactId>junit</artifactId>
            <version>3.8.1</version>
            <scope>test</scope>
        </dependency>
        <dependency>
            <groupId>mysql</groupId>
            <artifactId>mysql-connector-java</artifactId>
            <version>5.1.21</version>
        </dependency>
        <dependency>
            <groupId>org.hibernate</groupId>
            <artifactId>hibernate-core</artifactId>
            <version>4.0.1.Final</version>
        </dependency>
        <dependency>
            <groupId>org.hibernate</groupId>
            <artifactId>hibernate-entitymanager</artifactId>
            <version>4.0.1.Final</version>
        </dependency>
        <dependency>
            <groupId>org.hibernate.common</groupId>
            <artifactId>hibernate-commons-annotations</artifactId>
            <version>4.0.1.Final</version>
        </dependency>   
        <dependency>
            <groupId>nz.ac.waikato.cms.weka</groupId>
            <artifactId>weka-dev</artifactId>
            <version>3.7.10</version>
        </dependency>
        <dependency>
            <groupId>commons-configuration</groupId>
            <artifactId>commons-configuration</artifactId>
            <version>1.9</version>
        </dependency>
        <dependency>
            <groupId>commons-net</groupId>
            <artifactId>commons-net</artifactId>
            <version>3.1</version>
            <classifier>examples</classifier>
        </dependency>
        <dependency>
            <groupId>com.google.code.gson</groupId>
            <artifactId>gson</artifactId>
            <version>2.2.2</version>
        </dependency>
        <dependency>
            <groupId>maven</groupId>
            <artifactId>maven-jetty-plugin</artifactId>
            <version>1.1</version>
            <type>plugin</type>
        </dependency>
        <dependency>
            <groupId>commons-io</groupId>
            <artifactId>commons-io</artifactId>
            <version>2.4</version>
        </dependency>
        <dependency>
                <groupId>com.kenai.nbpwr</groupId>
                <artifactId>org-slf4j-jdk14</artifactId>
                <version>1.6.1-201106101300</version>
                <type>nbm</type>
        </dependency>

    </dependencies>
</project>

Answer 1

这不是我问题的答案，但可能有助于公众解决大型数据集的速度问题。我发现，选择一百万条记录通常可以在3秒钟内完成，但是连接可能会减慢进程。在这种情况下，一个有大约150k个Foo，这个Foo与1M条有1-many关系，那么选择那些使用JOIN的Foo可能会很慢，因为每个Foo返回大约6.5次。我发现，在python中分别选择两个表并使用dict连接它们比SQLAlchemy（大约25秒）快3倍，比使用连接的“裸”python代码快2倍（大约17秒）。在我的用例中，代码花费了8秒。选择1M个不带关系的记录（如上面的条形图示例）需要3秒钟。我用了这个代码：

#!/usr/bin/python
# -*- coding: utf-8 -*-

import MySQLdb
import sys
import time
import datetime
import inspect
from operator import itemgetter, attrgetter

# fetch all objects of class Class, where the fields are determined as the
# arguments of the __init__ constructor (not flexible, but fairly simple ;))
def fetch(Class, cursor, tablename, ids=["id"], where=None):
    arguments = inspect.getargspec(Class.__init__).args; del arguments[0];
    fields = ", ".join(["`" + tablename + "`.`" + column + "`" for column in arguments])
    sql = "SELECT " + fields + " FROM `" + tablename + "`"
    if where != None: sql = sql + " WHERE " + where
    sql=sql+";"
    getId = itemgetter(*[arguments.index(x) for x in ids])
    elements = dict()

    cursor.execute(sql)
    for record in cursor:
        elements[getId(record)] = Class(*record)
    return elements

# attach the objects in dict2 to dict1, given a 1-many relation between both
def merge(dict1, fieldname, dict2, ids):
    idExtractor = attrgetter(*ids)
    for d in dict1: setattr(dict1[d], fieldname, list())
    for d in dict2:
        dd = dict2[d]
        getattr(dict1[idExtractor(dd)], fieldname).append(dd)

# attach dict2 objects to dict1 objects, given a 1-1 relation
def attach(dict1, fieldname, dict2, ids):
    idExtractor = attrgetter(*ids)
    for d in dict1: dd=dict1[d]; setattr(dd, fieldname, dict2[idExtractor(dd)])

它帮助我加快了查询速度，但是我非常高兴听到专家们对这种方法可能的改进。

Answer 2

下面是您的MySQL脚本的SQLAlchemy版本，与MySQLdb的三个版本相比，它在四秒钟内执行：

from sqlalchemy import Integer, Column, create_engine, MetaData, Table
import datetime

metadata = MetaData()

foo = Table(
    'foo', metadata,
    Column('id', Integer, primary_key=True),
    Column('a', Integer(), nullable=False),
    Column('b', Integer(), nullable=False),
    Column('c', Integer(), nullable=False),
)


class Foo(object):
    def __init__(self, a, b, c):
        self.a = a
        self.b = b
        self.c = c

engine = create_engine('mysql+mysqldb://scott:tiger@localhost/test', echo=True)
start = datetime.datetime.now()

with engine.connect() as conn:
    foos = [
        Foo(row['a'], row['b'], row['c'])
        for row in
        conn.execute(foo.select().limit(1000000)).fetchall()
    ]


print "total time: ", datetime.datetime.now() - start

运行时：

total time:  0:00:04.706010

下面是一个使用ORM完全加载对象行的脚本；通过避免使用yield per一次创建包含所有1M对象的固定列表，这将在SQLAlchemy master（rel 0.9为18秒）中运行13秒：

import time
from sqlalchemy import Integer, Column, create_engine, Table
from sqlalchemy.orm import Session
from sqlalchemy.ext.declarative import declarative_base

Base = declarative_base()


class Foo(Base):
    __table__ = Table(
        'foo', Base.metadata,
        Column('id', Integer, primary_key=True),
        Column('a', Integer(), nullable=False),
        Column('b', Integer(), nullable=False),
        Column('c', Integer(), nullable=False),
    )


engine = create_engine('mysql+mysqldb://scott:tiger@localhost/test', echo=True)

sess = Session(engine)

now = time.time()

# avoid using all() so that we don't have the overhead of building
# a large list of full objects in memory
for obj in sess.query(Foo).yield_per(100).limit(1000000):
    pass

print("Total time: %d" % (time.time() - now))

然后，我们可以分割这两种方法之间的差异，并使用ORM只加载单个列：

for obj in sess.query(Foo.id, Foo.a, Foo.b, Foo.c).yield_per(100).limit(1000000):
    pass

上述操作将在4秒内再次运行。

与原始MySQLdb游标相比，SQLAlchemy Core的比较更为贴切。如果对单个列使用ORM but查询，在最新版本中大约需要4秒。

在ORM级别，速度问题是因为在Python中创建对象很慢，SQLAlchemy ORM在获取对象时对这些对象应用大量的簿记，这是它履行其使用契约所必需的，包括工作单元、标识映射、紧急加载、集合等

要显著加快查询速度，请获取单个列而不是完整对象。请参见 http://docs.sqlalchemy.org/en/latest/faq/performance.html#result-fetching-slowness-orm描述了这一点。

与PeeWee相比，PW是一个简单得多的系统，功能更少，包括它与身份映射没有任何关系。即使使用PeeWee，只要是一个简单的ORM，它仍然需要15秒，这就证明了与使用纯C的原始MySQLdb fetch相比，cPython确实非常慢

与Java相比，Java VM要比cPython快得多。Hibernate异常复杂，但是Java VM由于JIT而非常快，甚至所有的复杂性都以更快的速度结束。如果要将Python与Java进行比较，请使用Pypy。

Answer 3

炼金术很复杂。它必须处理将类型转换为底层数据库本机不支持的Python、具有继承的表、连接、缓存对象、保持一致性、转换的行、部分结果等等。看看sqlalchemy/orm/loading.py:instance_processor——这太疯狂了。

解决方案是拼凑并编译Python代码来处理特定查询的结果，就像Jinja2对模板所做的那样。到目前为止，还没有人做过这项工作，可能是因为常见的情况是两行（这种优化可能会比较麻烦），而需要处理大量数据的人会像您一样手工完成这项工作。