Why does 60GB memory disappear on a MySQL connector fetchall()?

Question

MySQL 5.7.18
Python 2.7.5
Pandas 0.17.1
CentOS 7.3

A MySQL table:

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

The size is 10GB.

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

10GB

The box has 250GB memory:

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

Select the data:

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())

The output:

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)

I even deleted the database connection and called garbage collection.

How could a 10GB table use up 60GB of my memory?

Answer 1

The short answer: python data structures memory overhead.

You have a table with ~231M rows taking ~10GB, so each row has about 4 bytes.

fetchall translate that into a list of tuples like this:

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

Your list has ~231M elements and uses ~19GB of memory: on average each tuple uses 8.48 bytes.

$ 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

A tuple:

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

A list of one tuple:

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

A list of two tuples:

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

A list with 10 tuples:

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

A list with 1M tuples:

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

Almost our number: 8.6 bytes per tuple in the list.

Unfortunately there isn't much you can do here: mysql.connector chooses the data structure and dict cursor would use even more memory.

If you need to reduce memory usage you must use fetchmany with suitable size argument.

Answer 2

Edit: pd.read_sql only accepts SQLAlchemy connections. Start by using create_engine from SQLAlchemy to connect to your database:

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

then call .connect() on the resulting object:

connection = engine.connect()

Pass that connection to pd.read_sql :

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

This should decrease your memory footprint.

Would you mind posting the memory usage results after you try the above?