为什么60GB内存在MySQL连接器fetchall（）上消失了？

Question

MySQL 5.7.18 Python 2.7.5 熊猫0.17.1 CentOS 7.3

一个MySQL表：

CREATE TABLE test (
  id varchar(12)
) ENGINE=InnoDB;

大小为10GB。

select round(((data_length) / 1024 / 1024 / 1024)) "GB"
from information_schema.tables 
where table_name = "test"

10GB

盒子有250GB内存：

$ free -hm
              total        used        free      shared  buff/cache   available
Mem:           251G         15G        214G        2.3G         21G        232G
Swap:          2.0G        1.2G        839M

选择数据：

import psutil
print '1 ' + str(psutil.phymem_usage())

import os
import sys
import time
import pyodbc 
import mysql.connector
import pandas as pd
from datetime import date
import gc
print '2 ' + str(psutil.phymem_usage())

db = mysql.connector.connect({snip})
c = db.cursor()
print '3 ' + str(psutil.phymem_usage())

c.execute("select id from test")
print '4 ' + str(psutil.phymem_usage())

e=c.fetchall()
print 'getsizeof: ' + str(sys.getsizeof(e))
print '5 ' + str(psutil.phymem_usage())

d=pd.DataFrame(e)
print d.info()
print '6 ' + str(psutil.phymem_usage())

c.close()
print '7 ' + str(psutil.phymem_usage())

db.close()
print '8 ' + str(psutil.phymem_usage())

del c, db, e
print '9 ' + str(psutil.phymem_usage())

gc.collect()
print '10 ' + str(psutil.phymem_usage())

time.sleep(60)
print '11 ' + str(psutil.phymem_usage())

输出：

1 svmem(total=270194331648L, available=249765777408L, percent=7.6, used=39435464704L, free=230758866944L, active=20528222208, inactive=13648789504, buffers=345387008L, cached=18661523456)
2 svmem(total=270194331648L, available=249729019904L, percent=7.6, used=39472222208L, free=230722109440L, active=20563484672, inactive=13648793600, buffers=345387008L, cached=18661523456)
3 svmem(total=270194331648L, available=249729019904L, percent=7.6, used=39472222208L, free=230722109440L, active=20563484672, inactive=13648793600, buffers=345387008L, cached=18661523456)
4 svmem(total=270194331648L, available=249729019904L, percent=7.6, used=39472222208L, free=230722109440L, active=20563484672, inactive=13648793600, buffers=345387008L, cached=18661523456)
getsizeof: 1960771816
5 svmem(total=270194331648L, available=181568315392L, percent=32.8, used=107641655296L, free=162552676352L, active=88588271616, inactive=13656334336, buffers=345395200L, cached=18670243840)
<class 'pandas.core.frame.DataFrame'>
Int64Index: 231246823 entries, 0 to 231246822
Data columns (total 1 columns):
0    object
dtypes: object(1)
memory usage: 3.4+ GB
None
6 svmem(total=270194331648L, available=181571620864L, percent=32.8, used=107638353920L, free=162555977728L, active=88587603968, inactive=13656334336, buffers=345395200L, cached=18670247936)
7 svmem(total=270194331648L, available=181571620864L, percent=32.8, used=107638353920L, free=162555977728L, active=88587603968, inactive=13656334336, buffers=345395200L, cached=18670247936)
8 svmem(total=270194331648L, available=181571620864L, percent=32.8, used=107638353920L, free=162555977728L, active=88587603968, inactive=13656334336, buffers=345395200L, cached=18670247936)
9 svmem(total=270194331648L, available=183428308992L, percent=32.1, used=105781678080L, free=164412653568L, active=86735921152, inactive=13656334336, buffers=345395200L, cached=18670260224)
10 svmem(total=270194331648L, available=183428308992L, percent=32.1, used=105781678080L, free=164412653568L, active=86735921152, inactive=13656334336, buffers=345395200L, cached=18670260224)
11 svmem(total=270194331648L, available=183427203072L, percent=32.1, used=105782812672L, free=164411518976L, active=86736560128, inactive=13656330240, buffers=345395200L, cached=18670288896)

我甚至删除了数据库连接并调用了垃圾收集。

10GB的桌子如何耗尽60GB的内存？

Answer 1

简短的回答：python数据结构内存开销。

你有一个~231M行的表占用~10GB，所以每行大约有4个字节。

fetchall将其转换为这样的元组列表：

[('abcd',), ('1234',), ... ]

你的列表有~231M元素并使用~19GB内存：平均每个元组使用8.48个字节。

$ python
Python 2.7.12 (default, Nov 19 2016, 06:48:10)
[GCC 5.4.0 20160609] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys

一个元组：

>>> a = ('abcd',)
>>> sys.getsizeof(a)
64

一个元组的列表：

>>> al = [('abcd',)]
>>> sys.getsizeof(al)
80

两个元组的列表：

>>> al2 = [('abcd',), ('1234',)]
>>> sys.getsizeof(al2)
88

包含10个元组的列表：

>>> al10 = [ ('abcd',) for x in range(10)]
>>> sys.getsizeof(al10)
200

包含1M元组的列表：

>>> a_realy_long = [ ('abcd',) for x in range(1000000)]
>>> sys.getsizeof(a_realy_long )
8697472

几乎我们的数字：列表中每个元组8.6个字节。

不幸的是，你在这里做的并不多：mysql.connector选择数据结构，dict cursor将使用更多的内存。

如果您需要减少内存使用量，则必须使用具有合适大小参数的fetchmany。

Answer 2

编辑：pd.read_sql只接受SQLAlchemy连接。首先使用SQLAlchemy中的create_engine连接到您的数据库：

from sqlalchemy import create_engine
engine = create_engine('mysql://database')

然后在结果对象上调用.connect()：

connection = engine.connect()

将该连接传递给pd.read_sql：

df = pd.read_sql("select id from test", connection)

这应该会减少您的内存占用。

您是否介意在尝试上述操作后发布内存使用结果？