How to concatenate multiple pandas.DataFrames without running into MemoryError

The problem is, like viewed in the others answers, a problem of memory. And a solution is to store data on disk, then to build an unique dataframe.

With such huge data, performance is an issue.

csv solutions are very slow, since conversion in text mode occurs. HDF5 solutions are shorter, more elegant and faster since using binary mode. I propose a third way in binary mode, with pickle, which seems to be even faster, but more technical and needing some more room. And a fourth, by hand.

Here the code:

import numpy as np
import pandas as pd
import os
import pickle

# a DataFrame factory:
dfs=[]
for i in range(10):
    dfs.append(pd.DataFrame(np.empty((10**5,4)),columns=range(4)))
    
# a csv solution
def bycsv(dfs):
    md,hd='w',True
    for df in dfs:
        df.to_csv('df_all.csv',mode=md,header=hd,index=None)
        md,hd='a',False
    #del dfs
    df_all=pd.read_csv('df_all.csv',index_col=None)
    os.remove('df_all.csv') 
    return df_all    

    

Better solutions :

def byHDF(dfs):
    store=pd.HDFStore('df_all.h5')
    for df in dfs:
        store.append('df',df,data_columns=list('0123'))
    #del dfs
    df=store.select('df')
    store.close()
    os.remove('df_all.h5')
    return df

def bypickle(dfs):
    c=[]
    with open('df_all.pkl','ab') as f:
        for df in dfs:
            pickle.dump(df,f)
            c.append(len(df))    
    #del dfs
    with open('df_all.pkl','rb') as f:
        df_all=pickle.load(f)
        offset=len(df_all)
        df_all=df_all.append(pd.DataFrame(np.empty(sum(c[1:])*4).reshape(-1,4)))
        
        for size in c[1:]:
            df=pickle.load(f)
            df_all.iloc[offset:offset+size]=df.values 
            offset+=size
    os.remove('df_all.pkl')
    return df_all
    

For homogeneous dataframes, we can do even better :

def byhand(dfs):
    mtot=0
    with open('df_all.bin','wb') as f:
        for df in dfs:
            m,n =df.shape
            mtot += m
            f.write(df.values.tobytes())
            typ=df.values.dtype                
    #del dfs
    with open('df_all.bin','rb') as f:
        buffer=f.read()
        data=np.frombuffer(buffer,dtype=typ).reshape(mtot,n)
        df_all=pd.DataFrame(data=data,columns=list(range(n))) 
    os.remove('df_all.bin')
    return df_all

And some tests on (little, 32 Mb) data to compare performance. you have to multiply by about 128 for 4 Gb.

In [92]: %time w=bycsv(dfs)
Wall time: 8.06 s

In [93]: %time x=byHDF(dfs)
Wall time: 547 ms

In [94]: %time v=bypickle(dfs)
Wall time: 219 ms

In [95]: %time y=byhand(dfs)
Wall time: 109 ms

A check :

In [195]: (x.values==w.values).all()
Out[195]: True

In [196]: (x.values==v.values).all()
Out[196]: True

In [197]: (x.values==y.values).all()
Out[196]: True


            

Of course all of that must be improved and tuned to fit your problem.

For exemple df3 can be split in chuncks of size 'total_memory_size - df_total_size' to be able to run bypickle.

I can edit it if you give more information on your data structure and size if you want. Beautiful question !

I advice you to put your dataframes into single csv file by concatenation. Then to read your csv file.

Execute that:

# write df1 content in file.csv
df1.to_csv('file.csv', index=False)
# append df2 content to file.csv
df2.to_csv('file.csv', mode='a', index=False)
# append df3 content to file.csv
df3.to_csv('file.csv', mode='a', index=False)

# free memory
del df1, df2, df3

# read all df1, df2, df3 contents
df = pd.read_csv('file.csv')

If this solution isn't enougth performante, to concat larger files than usually. Do:

df1.to_csv('file.csv', index=False)
df2.to_csv('file1.csv', index=False)
df3.to_csv('file2.csv', index=False)

del df1, df2, df3

Then run bash command:

cat file1.csv >> file.csv
cat file2.csv >> file.csv
cat file3.csv >> file.csv

Or concat csv files in python :

def concat(file1, file2):
    with open(file2, 'r') as filename2:
        data = file2.read()
    with open(file1, 'a') as filename1:
        file.write(data)

concat('file.csv', 'file1.csv')
concat('file.csv', 'file2.csv')
concat('file.csv', 'file3.csv')

After read:

df = pd.read_csv('file.csv')

Kinda taking a guess here, but maybe:

df1 = pd.concat([df1,df2])
del df2
df1 = pd.concat([df1,df3])
del df3

Obviously, you could do that more as a loop but the key is you want to delete df2, df3, etc. as you go. As you are doing it in the question, you never clear out the old dataframes so you are using about twice as much memory as you need to.

More generally, if you are reading and concatentating, I'd do it something like this (if you had 3 CSVs: foo0, foo1, foo2):

concat_df = pd.DataFrame()
for i in range(3):
    temp_df = pd.read_csv('foo'+str(i)+'.csv')
    concat_df = pd.concat( [concat_df, temp_df] )

In other words, as you are reading in files, you only keep the small dataframes in memory temporarily, until you concatenate them into the combined df, concat_df. As you currently do it, you are keeping around all the smaller dataframes, even after concatenating them.

Similar to what @glegoux suggests, also pd.DataFrame.to_csv can write in append mode, so you can do something like:

df1.to_csv(filename)
df2.to_csv(filename, mode='a')
df3.to_csv(filename, mode='a')

del df1, df2, df3
df_concat = pd.read_csv(filename)

Dask might be good option to try for handling large dataframes - Go through Dask Docs

You can store your individual dataframes in a HDF Store, and then call the store just like one big dataframe.

# name of store
fname = 'my_store'

with pd.get_store(fname) as store:

    # save individual dfs to store
    for df in [df1, df2, df3, df_foo]:
        store.append('df',df,data_columns=['FOO','BAR','ETC']) # data_columns = identify the column in the dfs you are appending

    # access the store as a single df
    df = store.select('df', where = ['A>2'])  # change where condition as required (see documentation for examples)
    # Do other stuff with df #

# close the store when you're done
os.remove(fname)

I've had a similar performance issues while trying to concatenate a large number of DataFrames to a 'growing' DataFrame.

My workaround was appending all sub DataFrames to a list, and then concatenating the list of DataFrames once processing of the sub DataFrames has been completed. This will bring the runtime to almost half.

Another option:

1) Write df1 to .csv file: df1.to_csv('Big file.csv')

2) Open .csv file, then append df2:

with open('Big File.csv','a') as f:
    df2.to_csv(f, header=False)

3) Repeat Step 2 with df3

with open('Big File.csv','a') as f:
    df3.to_csv(f, header=False)

While writing to hard disk, df.to_csv throws an error for columns=False.

The below solutions works fine:

# write df1 to hard disk as file.csv
train1.to_csv('file.csv', index=False)
# append df2 to file.csv
train2.to_csv('file.csv', mode='a', header=False, index=False)
# read the appended csv as df
train = pd.read_csv('file.csv')

This function converts columns' dtype to a minimum possible to fit any value in that column. It can reduce the size even more than 10x in some cases. After this a loop to append all of dataframes to a list, a pd.concat for one big file and at the end delete the dataframe list to free memory.

# Function to reduce memory usage
def reduce_size(df):
# Define lists of dtypes
cat_cols = df.select_dtypes(exclude=[np.number, np.datetime64]).columns 
num_cols = df.select_dtypes(include=[np.number]).columns

# Convert non_numericals to category
df[cat_cols] = df[cat_cols].astype('category')

# Convert numericals to the lowest dtype
for col in num_cols:
    df[col] = pd.to_numeric(df[col], downcast='float')
    # Try to make them integer
    df[col] = pd.to_numeric(df[col], downcast='integer')  
return df

Inspired from @BM's answer, the CSV solution does the trick and was much faster. However, any custom index was being lost through the CSV method. This is my updated solution.

df_file_path = "/path/to/save/df.csv"
def merge_dfs(dfs, keep_index=True, sync_columns=False, dtype=None):
    """
    Merge multiple dataframes into one as an alternative to `pd.concat`.

    This expects all dataframes to have their columns in
    the same order but you can set `sync_columns` to True to ensure that.
    `pd.concat` automatically handles this.

    Some data might be read back in improperly as pandas tries to
    guess the appropriate data type. Use `dtype` for this.
    """
    md, hd = 'w', True
    if sync_columns:
        cols = dfs[0].columns.tolist()

    for df in dfs:
        if sync_columns:
            df = df[cols]
        df.to_csv(df_file_path, mode=md, header=hd, index=keep_index)
        md, hd = 'a', False
    df_concat = pd.read_csv(df_file_path, index_col=(keep_index and 0), dtype=dtype)
    os.remove(df_file_path)
    return df_concat


# Instead of:
# concat_df = pd.concat([df1, df2, df3])
concat_df = merge_dfs([df1, df2, df3])

This method cut down time from 45 minutes to 2 minutes for one of my dataset and memory usage from 140GB to only 17GB! None of the other solutions (HDF, pickle, hand as of 10/2024) worked however, likely from incompatibilities from package version updates.