I have a 10gb CSV file that contains some information that I need to use.
As I have limited memory on my PC, I can not read all the file in memory in one single batch. Instead, I would like to iteratively read only some rows of this file.
Say that at the first iteration I want to read the first 100, at the second those going to 101 to 200 and so on.
Is there an efficient way to perform this task in Python?
May Pandas provide something useful to this? Or are there better (in terms of memory and speed) methods?
Here is the short answer.
chunksize = 10 ** 6
for chunk in pd.read_csv(filename, chunksize=chunksize):
process(chunk)
Here is the very long answer.
To get started, you’ll need to import pandas and sqlalchemy. The commands below will do that.
import pandas as pd
from sqlalchemy import create_engine
Next, set up a variable that points to your csv file. This isn’t necessary but it does help in re-usability.
file = '/path/to/csv/file'
With these three lines of code, we are ready to start analyzing our data. Let’s take a look at the ‘head’ of the csv file to see what the contents might look like.
print pd.read_csv(file, nrows=5)
This command uses pandas’ “read_csv” command to read in only 5 rows (nrows=5) and then print those rows to the screen. This lets you understand the structure of the csv file and make sure the data is formatted in a way that makes sense for your work.
Before we can actually work with the data, we need to do something with it so we can begin to filter it to work with subsets of the data. This is usually what I would use pandas’ dataframe for but with large data files, we need to store the data somewhere else. In this case, we’ll set up a local sqllite database, read the csv file in chunks and then write those chunks to sqllite.
To do this, we’ll first need to create the sqllite database using the following command.
csv_database = create_engine('sqlite:///csv_database.db')
Next, we need to iterate through the CSV file in chunks and store the data into sqllite.
chunksize = 100000
i = 0
j = 1
for df in pd.read_csv(file, chunksize=chunksize, iterator=True):
df = df.rename(columns={c: c.replace(' ', '') for c in df.columns})
df.index += j
i+=1
df.to_sql('table', csv_database, if_exists='append')
j = df.index[-1] + 1
With this code, we are setting the chunksize at 100,000 to keep the size of the chunks managable, initializing a couple of iterators (i=0, j=0) and then running a through a for loop. The for loop read a chunk of data from the CSV file, removes space from any of column names, then stores the chunk into the sqllite database (df.to_sql(…)).
This might take a while if your CSV file is sufficiently large, but the time spent waiting is worth it because you can now use pandas ‘sql’ tools to pull data from the database without worrying about memory constraints.
To access the data now, you can run commands like the following:
df = pd.read_sql_query('SELECT * FROM table', csv_database)
Of course, using ‘select *…’ will load all data into memory, which is the problem we are trying to get away from so you should throw from filters into your select statements to filter the data. For example:
df = pd.read_sql_query('SELECT COl1, COL2 FROM table where COL1 = SOMEVALUE', csv_database)
You can use pandas.read_csv() with chuncksize parameter:
http://pandas.pydata.org/pandas-docs/stable/generated/pandas.read_csv.html#pandas.read_csv
for chunck_df in pd.read_csv('yourfile.csv', chunksize=100):
# each chunck_df contains a part of the whole CSV
This code may help you for this task. It navigates trough a large .csv file and does not consume lots of memory so that you can perform this in a standard lap top.
import pandas as pd
import os
The chunksize here orders the number of rows within the csv file you want to read later
chunksize2 = 2000
path = './'
data2 = pd.read_csv('ukb35190.csv',
chunksize=chunksize2,
encoding = "ISO-8859-1")
df2 = data2.get_chunk(chunksize2)
headers = list(df2.keys())
del data2
start_chunk = 0
data2 = pd.read_csv('ukb35190.csv',
chunksize=chunksize2,
encoding = "ISO-8859-1",
skiprows=chunksize2*start_chunk)
headers = []
for i, df2 in enumerate(data2):
try:
print('reading cvs....')
print(df2)
print('header: ', list(df2.keys()))
print('our header: ', headers)
# Access chunks within data
# for chunk in data:
# You can now export all outcomes in new csv files
file_name = 'export_csv_' + str(start_chunk+i) + '.csv'
save_path = os.path.abspath(
os.path.join(
path, file_name
)
)
print('saving ...')
except Exception:
print('reach the end')
break
Method to transfer huge CSV into database is good because we can easily use SQL query.
We have also to take into account two things.
FIRST POINT: SQL also are not a rubber, it will not be able to stretch the memory.
For example converted to bd file:
https://nycopendata.socrata.com/Social-Services/311-Service-Requests-
from-2010-to-Present/erm2-nwe9
For this db file SQL language:
pd.read_sql_query("SELECT * FROM 'table'LIMIT 600000", Mydatabase)
It can read maximum about 0,6 mln records no more with 16 GB RAM memory of PC (time of operation 15,8 second).
It could be malicious to add that downloading directly from a csv file is a bit more efficient:
giga_plik = 'c:/1/311_Service_Requests_from_2010_to_Present.csv'
Abdul = pd.read_csv(giga_plik, nrows=1100000)
(time of operation 16,5 second)
SECOND POINT: To effectively using SQL data series converted from CSV we ought to memory about suitable form of date. So I proposer add to ryguy72's code this:
df['ColumnWithQuasiDate'] = pd.to_datetime(df['Date'])
All code for file 311 as about I pointed:
start_time = time.time()
### sqlalchemy create_engine
plikcsv = 'c:/1/311_Service_Requests_from_2010_to_Present.csv'
WM_csv_datab7 = create_engine('sqlite:///C:/1/WM_csv_db77.db')
#----------------------------------------------------------------------
chunksize = 100000
i = 0
j = 1
## --------------------------------------------------------------------
for df in pd.read_csv(plikcsv, chunksize=chunksize, iterator=True, encoding='utf-8', low_memory=False):
df = df.rename(columns={c: c.replace(' ', '') for c in df.columns})
## -----------------------------------------------------------------------
df['CreatedDate'] = pd.to_datetime(df['CreatedDate']) # to datetimes
df['ClosedDate'] = pd.to_datetime(df['ClosedDate'])
## --------------------------------------------------------------------------
df.index += j
i+=1
df.to_sql('table', WM_csv_datab7, if_exists='append')
j = df.index[-1] + 1
print(time.time() - start_time)
At the end I would like to add: converting a csv file directly from the Internet to db seems to me a bad idea. I propose to download base and convert locally.
Related
I am very new to programming and I'm currently trying to breakdown a CSV file into bite-sized chunks and saving those chunks as individual files. I was able to do it using a super roundabout way but now I want to figure out how to write a function that does this for me.
This is the code I have now:
import pandas as pd
def chunky(file, chunksize, iterator, rounds):
df = pd.read_csv(file, chunksize = chunksize, iterator = iterator)
count = 0
while count < rounds:
count += 1
file = next(df)
file.to_csv("output.csv")
However, when I do this it will print the output as one file but I am looking to save multiple individual files. Any and all help is appreciated! Thank you so much in advance :)
Don't use both chunksize and iterator.
Also make sure you create a new output file for each subfile.
The code below will create a set of new files up to 2000 rows long. The new output files are named after the input file, but with a number with leading zeros (change the 5 accordingly if you need more or less files).
import os
reader = pd.read_csv(filename, chunksize=2000)
for i, df in enumerate(reader):
output = os.path.splitext(filename)[0] + f'-{i+1:05d}.csv'
df.to_csv(output, index=False)
It's my first time creating a code for processing files with a lot of data, so I am kinda stuck here.
What I'm trying to do is to read a list of path, listing all of the csv files that need to be read, retrieve the HEAD and TAIL from each files and put it inside a list.
I have 621 csv files in total, with each files consisted of 5800 rows, and 251 columns
This is the data sample
[LOGGING],RD81DL96_1,3,4,5,2,,,,
LOG01,,,,,,,,,
DATETIME,INDEX,SHORT[DEC.0],SHORT[DEC.0],SHORT[DEC.0],SHORT[DEC.0],SHORT[DEC.0],SHORT[DEC.0],SHORT[DEC.0],SHORT[DEC.0]
TIME,INDEX,FF-1(１A) ,FF-1(１B) ,FF-1(１C) ,FF-1(２A),FF-2(１A) ,FF-2(１B) ,FF-2(１C),FF-2(２A)
47:29.6,1,172,0,139,1258,0,0,400,0
47:34.6,2,172,0,139,1258,0,0,400,0
47:39.6,3,172,0,139,1258,0,0,400,0
47:44.6,4,172,0,139,1263,0,0,400,0
47:49.6,5,172,0,139,1263,0,0,450,0
47:54.6,6,172,0,139,1263,0,0,450,0
The problem is, while it took about 13 seconds to read all the files (still kinda slow honestly)
But when I add a single line of append code, the process took a lot of times to finish, about 4 minutes.
Below is the snipset of the code:
# CsvList: [File Path, Change Date, File size, File Name]
for x, file in enumerate(CsvList):
timeColumn = ['TIME']
df = dd.read_csv(file[0], sep =',', skiprows = 3, encoding= 'CP932', engine='python', usecols=timeColumn)
# The process became long when this code is added
startEndList.append(list(df.head(1)) + list(df.tail(1)))
Why that happened? I'm using dask.dataframe
Currently, your code isn't really leveraging Dask's parallelizing capabilities because:
df.head and df.tail calls will trigger a "compute" (i.e., convert your Dask DataFrame into a pandas DataFrame -- which is what we try to minimize in lazy evaluations with Dask), and
the for-loop is running sequentially because you're creating Dask DataFrames and converting them to pandas DataFrames, all inside the loop.
So, your current example is similar to just using pandas within the for-loop, but with the added Dask-to-pandas-conversion overhead.
Since you need to work on each of your files, I'd suggest checking out Dask Delayed, which might be more elegant+ueful here. The following (pseudo-code) will parallelize the pandas operation on each of your files:
import dask
import pandas as pd
for file in list_of_files:
df = dask.delayed(pd.read_csv)(file)
result.append(df.head(1) + df.tail(1))
dask.compute(*result)
The output of dask.visualize(*result) when I used 4 csv-files confirms parallelism:
If you really want to use Dask DataFrame here, you may try to:
read all files into a single Dask DataFrame,
make sure each Dask "partition" corresponds to one file,
use Dask Dataframe apply to get the head and tail values and append them to a new list
call compute on the new list
A first approach using only Python as starting point:
import pandas as pd
import io
def read_first_and_last_lines(filename):
with open(filename, 'rb') as fp:
# skip first 4 rows (headers)
[next(fp) for _ in range(4)]
# first line
first_line = fp.readline()
# start at -2x length of first line from the end of file
fp.seek(-2 * len(first_line), 2)
# last line
last_line = fp.readlines()[-1]
return first_line + last_line
data = []
for filename in pathlib.Path('data').glob('*.csv'):
data.append(read_first_and_last_lines(filename))
buf = io.BytesIO()
buf.writelines(data)
buf.seek(0)
df = pd.read_csv(buf, header=None, encoding='CP932')
I have a process which is reading from 4 databases with 4 tables each. I am consolidating that data into 1 postgres database with 4 tables total. (Each of the original 4 databases have the same 4 tables which need to be consolidated).
The way I am doing it now works using pandas. I read one table from all 4 databases at a time, concatenate the data into one dataframe, then I use to_sql to save it on my postgres database. I then loop through to the remaining databases doing the same thing for the other tables.
My issue is speed. One of my tables has about 1 - 2mil rows per date so it can take about 5,000 - 6,000 seconds to finish writing the data to postgres. It is much quicker to write it to a .csv file and then use COPY FROM in pgadmin.
Here is my current code. Note that there are some function calls but it is just referring to the table names basically. I also have some basic logging being done but that is not too necessary. I am adding a column for the source database which is required though. I am stripping .0 from fields which are actually strings but pandas sees them as a float too, and I fill blank integers with 0 and make sure the columns are really type int.
def query_database(table, table_name, query_date):
df_list = []
log_list = []
for db in ['NJ', 'NJ2', 'LA', 'NA']:
start_time = time.clock()
query_timestamp = dt.datetime.now(pytz.timezone('UTC')).strftime('%Y-%m-%d %H:%M:%S')
engine_name = '{}{}{}{}'.format(connection_type, server_name, '/', db)
print('Accessing {} from {}'.format((select_database(db)[0][table]), engine_name))
engine = create_engine(engine_name)
df = pd.read_sql_query(query.format(select_database(db)[0][table]), engine, params={query_date})
query_end = time.clock() - start_time
df['source_database'] = db
df['insert_date_utc'] = query_timestamp
df['row_count'] = df.shape[0]
df['column_count'] = df.shape[1]
df['query_time'] = round(query_end, 0)
df['maximum_id'] = df['Id'].max()
df['minimum_id'] = df['Id'].min()
df['source_table'] = table_dict.get(table)
log = df[['insert_date_utc', 'row_date', 'source_database', 'source_table', 'row_count', 'column_count', 'query_time', 'maximum_id', 'minimum_id']].copy()
df.drop(['row_count', 'column_count', 'query_time', 'maximum_id', 'minimum_id', 'source_table'], inplace=True, axis=1)
df_list.append(df)
log_list.append(log)
log = pd.concat(log_list)
log.drop_duplicates(subset=['row_date', 'source_database', 'source_table'], inplace=True, keep='last')
result = pd.concat(df_list)
result.drop_duplicates('Id', inplace=True)
cols = [i.strip() for i in (create_columns(select_database(db)[0][table]))]
result = result[cols]
print('Creating string columns for {}'.format(table_name))
for col in modify_str_cols(select_database(db)[0][table]):
create_string(result, col)
print('Creating integer columns for {}'.format(table_name))
for col in modify_int_cols(select_database(db)[0][table]):
create_int(result, col)
log.to_sql('raw_query_log', cms_dtypes.pg_engine, index=False, if_exists='append', dtype=cms_dtypes.log_dtypes)
print('Inserting {} data into PostgreSQL'.format(table_name))
result.to_sql(create_table(select_database(db)[0][table]), cms_dtypes.pg_engine, index=False, if_exists='append', chunksize=50000, dtype=create_dtypes(select_database(db)[0][table]))
How can I insert a COPY TO and COPY FROM into this to speed it up? Should I just write the .csv files and then loop over those or can I COPY from memory to my postgres?
psycopg2 offers a number of specific copy related apis. If you want to use csv, you have to use copy_expert (which allows you to specify a fully copy statement).
Normally when I have done this, I have used copy_expert() and a file-like object which iterates through a file on disk. That seems to work reasonably well.
This being said, in your case, I think copy_to and copy_from are better matches because it is simply postgres to postgres transfer here. Note these use PostgreSQL's copy output/input syntax and not csv (if you want to use csv, you have to use copy_expert)
Note before you decide how to do things, you will need to note:
copy_to copies to a file-like object (such as StringIO) and copy_from/copy_expert files from a file-like object. If you want to use a panda data frame you are going to have to think about this a little and either create a file-like object or use csv along with StringIO and copy_expert to generate an in-memory csv and load that.
I'm trying to process a large (2gb) csv file on a machine with only 4gb of RAM (don't ask) to produce a different, formatted csv containing a subset of data that needs some processing. I'm reading the file and creating a HDFstore that I query later for the data that I require for output. Everything works except that I cant retrieve data from the store using Term - error message comes back that PLOT is not a column name. Individual variables look fine and the store is what I expect I just can't see where the error is. (nb pandas v14 and numpy1.9.0). Very new to this so apologies for the clunky code.
#wibble wobble -*- coding: utf-8 -*-
# short version
def filesport():
import pandas as pd
import numpy as np
from pandas.io.pytables import Term
Location = r"CL_short.csv"
store = pd.HDFStore('blarg.h5')
maxlines = sum(1 for line in open (Location))
print maxlines
#set chunk small for test file
chunky=4
plotty =pd.DataFrame(columns=['PLOT'])
dfdum=pd.DataFrame(columns=['PLOT', 'mDate', 'D100'])
#read file in chunks to avoid RAM blowing up
bucket = pd.read_csv(Location, iterator=True, chunksize=chunky, usecols= ['PLOT','mDate','D100'])
for chunk in bucket:
store.append('wibble', chunk, format='table', data_columns=['PLOT','mDate','D100'], ignore_index=True)
#retrieve plot numbers and select unique items
plotty = store.select('wibble', "columns = ['PLOT']")
plotty.drop_duplicates(inplace=True)
#iterate through unique plots to retrieve data and put in dataframe for output
for index, row in plotty.iterrows():
dfdum = store.select('wibble', [Term('PLOT', '=', plotty.iloc[index]['PLOT'])])
#process dfdum for output to new csv
print("successful completion")
filesport()
Final listing for those that wish to fight through the tumbleweed to reach here and are similarly bemused by processing large .csv files and the various methods of trying to retrieve/process data. The biggest problem was getting the sytax of the pytables Term right. Despite several examples indicating that it was possible to use 'A >20' etc this never worked for me. I set up a string condition containing the Term query and this worked (it is in the documentation TBF).
Also found it easier to query the HDF to retrieve unique items direct from the store in a list which could then be sorted and iterated through to retrieve data plot by plot. Note that I wanted the final csv file to have plot and then all the D100 data in date order, hence the pivot at the end.
Reading the csv file in chunks meant that each plot retrieved from the store had a header and this got written to the final csv which messed things up. I'm sure there's a more elegant way of only writing one header than the one I've shown here.
It works, takes about 2 hours to process the data and produce the final csv file (initial file 2GB, 30+million lines, data for 100,000+ unique plots, machine has 4GB of RAM but running 32-bit which means that only 2.5GB of RAM was available).
Good luck if you have a similar problem, and I hope you find this useful
#wibble wobble -*- coding: utf-8 -*-
def filesport():
import pandas as pd
import numpy as np
from pandas.io.pytables import Term
print (pd.__version__)
print (np.__version__)
Location = r"conliq_med.csv"
store = pd.HDFStore('blarg.h5')
maxlines = sum(1 for line in open (Location))
print maxlines
chunky=100000
#read file in chunks to avoid RAM blowing up select only needed columns
bucket = pd.read_csv(Location, iterator=True, chunksize=chunky, usecols= ['PLOT','mDate','D100'])
for chunk in bucket:
store.append('wibble', chunk, format='table', data_columns=['PLOT','mDate','D100'], ignore_index=True)
#retrieve unique plots and sort
plotty = store.select_column('wibble', 'PLOT').unique()
plotty.sort()
#set flag for writing file header
i=0
#iterate through unique plots to retrieve data and put in dataframe for output
for item in plotty:
condition = 'PLOT =' + str(item)
dfdum = store.select('wibble', [Term(condition)])
dfdum["mDate"]= pd.to_datetime(dfdum["mDate"], dayfirst=True)
dfdum.sort(columns=["PLOT", "mDate"], inplace=True)
dfdum["mDate"] = dfdum["mDate"].map(lambda x: x.strftime("%Y - %m"))
dfdum=dfdum.pivot("PLOT", "mDate", "D100")
#only print one header to file
if i ==0:
dfdum.to_csv("CL_OP.csv", mode='a')
i=1
else:
dfdum.to_csv("CL_OP.csv", mode='a', header=False)
print("successful completion")
filesport()
I get an error 'TypeError: 'TextFileReader' object does not support item assignment' when I try to add columns and modify header names etc in chunks.
My issue is I am using a slow work laptop to process a pretty large file (10 million rows). I want to add some simple columns (1 or 0 values), concatenate two columns to create a unique ID, change the dtype for other columns, and rename some headers so they match with other files that I will .merge later. I could probably split this csv (maybe select date ranges and make separate files), but I would like to learn how to use chunksize or deal with large files in general without running into memory issues. Is it possible to modify a file in chunks and then concatenate them all together later?
I am doing a raw data clean up which will then be loaded into Tableau for visualization.
Example (reading/modifying 10 million rows):
> rep = pd.read_csv(r'C:\repeats.csv.gz',
> compression = 'gzip', parse_dates = True , usecols =
> ['etc','stuff','others','...'])
> rep.sort()
> rep['Total_Repeats'] = 1
> rep.rename(columns={'X':'Y'}, inplace = True)
> rep.rename(columns={'Z':'A'}, inplace = True)
> rep.rename(columns={'B':'C'}, inplace = True)
> rep['D']= rep['E'] + rep['C']
> rep.rename(columns={'L':'M'}, inplace = True)
> rep.rename(columns={'N':'O'}, inplace = True)
> rep.rename(columns={'S':'T'}, inplace = True)
If you pass chunk_size keyword to pd.read_csv, it returns iterator of csv reader. and you can write processed chunks with to_csv method in append mode. you will be able to process large file, but you can't sort dataframe.
import pandas as pd
reader = pd.read_csv(r'C:\repeats.csv.gz',
compression = 'gzip', parse_dates=True, chunk_size=10000
usecols = ['etc','stuff','others','...'])
output_path = 'output.csv'
for chunk_df in reader:
chunk_result = do_somthing_with(chunk_df)
chunk_result.to_csv(output_path, mode='a', header=False)
Python's usually pretty good with that as long as you ignore the .read() part when looking at large files.
If you just use the iterators, you should be fine:
with open('mybiginputfile.txt', 'rt') as in_file:
with open('mybigoutputfile.txt', 'wt') as out_file:
for row in in_file:
'do something'
out_file.write(row)
Someone who knows more will explain how the memory side of it works, but this works for me on multi GB files without crashing Python.
You might want to chuck the data into a proper DB before killing your laptop with the task of serving up the data AND running Tableau too!