Python - multiprocessing multiple large size files using pandas

Question

I have a y.csv file. The file size is 10 MB and it contains data from Jan 2020 to May 2020.

I also have a separate file for each month. e.g. data-2020-01.csv. It contains detailed data. The file size of each month file is around 1 GB.

I'm splitting the y.csv by month and then process the data by loading the relevant month file. This process is taking too long when I go for large number of months. e.g. 24 months.

I would like to process the data faster. I have access to AWS m6i.8xlarge instance which has 32 vCPU and 128 GB memory.

I'm new to multiprocessing. So can someone guide me here?

This is my current code.

import pandas as pd

periods = [(2020, 1), (2020, 2), (2020, 3), (2020, 4), (2020, 5)]

y = pd.read_csv("y.csv", index_col=0, parse_dates=True).fillna(0)  # Filesize: ~10 MB


def process(_month_df, _index):
    idx = _month_df.index[_month_df.index.get_loc(_index, method='nearest')]
    for _, value in _month_df.loc[idx:].itertuples():

        up_delta = 200
        down_delta = 200

        up_value = value + up_delta
        down_value = value - down_delta

        if value > up_value:
            y.loc[_index, "result"] = 1
            return

        if value < down_value:
            y.loc[_index, "result"] = 0
            return


for x in periods:
    filename = "data-" + str(x[0]) + "-" + str(x[1]).zfill(2)  # data-2020-01
    filtered_y = y[(y.index.month == x[1]) & (y.index.year == x[0])]  # Only get the current month records
    month_df = pd.read_csv(f'{filename}.csv', index_col=0, parse_dates=True)  # Filesize: ~1 GB (data-2020-01.csv)

    for index, row in filtered_y.iterrows():
        process(month_df, index)

Answer 1

A multithreading pool would be ideal for sharing the y dataframe among threads (obviating the need for using shared memory) but is not so good at running the more CPU-intensive processing in parallel. A multiprocessing pool is great for doing CPU-intensive processing but not so great in sharing data across processes without coming up with a shred memory representation of your y dataframe.

Here I have rearranged your code so that I use a multithreading pool to create filtered_y for each period (which is a CPU-intensive operation, but pandas does release the Global Interpreter Lock for certain operations -- hopefully this one). Then we are only passing one-months worth of data to a multiprocessing pool, rather than the entire y dataframe, to process that month with worker function process_month. But since each pool process does not have access to the y dataframe, it just returns the indices that need to be updated with the values to be replaced.

import pandas as pd
from multiprocessing.pool import Pool, ThreadPool, cpu_count

def process_month(period, filtered_y):
    """
    returns a list of tuples consisting of (index, value) pairs
    """
    filename = "data-" + str(period[0]) + "-" + str(period[1]).zfill(2)  # data-2020-01
    month_df = pd.read_csv(f'{filename}.csv', index_col=0, parse_dates=True)  # Filesize: ~1 GB (data-2020-01.csv)
    results = []
    for index, row in filtered_y.iterrows():   
        idx = month_df.index[month_df.index.get_loc(index, method='nearest')]
        for _, value in month_df.loc[idx:].itertuples():
    
            up_delta = 200
            down_delta = 200
    
            up_value = value + up_delta
            down_value = value - down_delta
    
            if value > up_value:
                results.append((index, 1))
                break
    
            if value < down_value:
                results.append((index, 0))
                break
    return results

def process(period):
    filtered_y = y[(y.index.month == period[1]) & (y.index.year == period[0])]  # Only get the current month records
    for index, value in multiprocessing_pool.apply(process_month, (period, filtered_y)):
        y.loc[index, "result"] = value

def main():
    global y, multiprocessing_pool

    periods = [(2020, 1), (2020, 2), (2020, 3), (2020, 4), (2020, 5)]
    y = pd.read_csv("y.csv", index_col=0, parse_dates=True).fillna(0)  # Filesize: ~10 MB

    MAX_THREAD_POOL_SIZE = 100
    thread_pool_size = min(MAX_THREAD_POOL_SIZE, len(periods))
    multiprocessing_pool_size = min(thread_pool_size, cpu_count())
    with Pool(multiprocessing_pool_size) as multiprocessing_pool, \
    ThreadPool(thread_pool_size) as thread_pool:
        thread_pool.map(process, periods)
        
    # Presumably y gets written out again as a CSV file here?

# Required for Windows:
if __name__ == '__main__':
    main()

Version Using Just a Single Multiprocessing Pool

import pandas as pd
from multiprocessing.pool import Pool, ThreadPool, cpu_count

def process_month(period):
    """
    returns a list of tuples consisting of (index, value) pairs
    """
    y = pd.read_csv("y.csv", index_col=0, parse_dates=True).fillna(0)  # Filesize: ~10 MB
    filtered_y = y[(y.index.month == period[1]) & (y.index.year == period[0])]  # Only get the current month records
    filename = "data-" + str(period[0]) + "-" + str(period[1]).zfill(2)  # data-2020-01
    month_df = pd.read_csv(f'{filename}.csv', index_col=0, parse_dates=True)  # Filesize: ~1 GB (data-2020-01.csv)
    results = []
    for index, row in filtered_y.iterrows():   
        idx = month_df.index[month_df.index.get_loc(index, method='nearest')]
        for _, value in month_df.loc[idx:].itertuples():
    
            up_delta = 200
            down_delta = 200
    
            up_value = value + up_delta
            down_value = value - down_delta
    
            if value > up_value:
                results.append((index, 1))
                break
    
            if value < down_value:
                results.append((index, 0))
                break
    return results

def main():
    periods = [(2020, 1), (2020, 2), (2020, 3), (2020, 4), (2020, 5)]

    multiprocessing_pool_size = min(len(periods), cpu_count())
    with Pool(multiprocessing_pool_size) as multiprocessing_pool:
        results_list = multiprocessing_pool.map(process_month, periods)
    y = pd.read_csv("y.csv", index_col=0, parse_dates=True).fillna(0)  # Filesize: ~10 MB
    for results in results_list:
        for index, value in results:
            y.loc[index, "result"] = value
    # Write out new csv file:
    ...

# Required for Windows:
if __name__ == '__main__':
    main()

And now for a variation of this that uses a bit more memory but allows the main process to overlap its processing with the multiprocessing pool. This could be beneficial if the number of indices needing to be updated is quite large:

...
def main():
    periods = [(2020, 1), (2020, 2), (2020, 3), (2020, 4), (2020, 5)]

    multiprocessing_pool_size = min(len(periods), cpu_count() - 1) # save a core for the main process
    y = pd.read_csv("y.csv", index_col=0, parse_dates=True).fillna(0)  # Filesize: ~10 MB
    with Pool(multiprocessing_pool_size) as multiprocessing_pool:
        # Process values as soon as they are returned:
        for results in multiprocessing_pool.imap_unordered(process_month, periods):
            for index, value in results:
                y.loc[index, "result"] = value
    # Write out new csv file:
    ...

This last version could be superior since it first reads the csv file before submitting tasks to the pool and depending on the platform and how it caches I/O operations it could result in the worker function not having to do any physical I/O to read in its copies of the file. But that is one more 10M file that has been read into memory.

Answer 2

As commented in multiple pandas/threading questions, CSV files being IO bound, you can get some benefit from using a ThreadPoolExecutor.

At the same time, if you are going to perform aggregating operations, consider performing the read_csv also inside of your processor and use ProcessPoolExecutor instead.

If you are going to pass a lot of data between your multiprocesses you will also need a proper memory sharing method.

However I see the use of iterrows and itertuples In general those two instructions make my eyes bleed. Are you sure you cannot process the data in a vectorised mode?

This particular section I am not sure what it is supposed to do, and having M rows will make it very slow.

def process(_month_df, _index):
    idx = _month_df.index[_month_df.index.get_loc(_index, method='nearest')]
    for _, value in _month_df.loc[idx:].itertuples():

        up_delta = 200
        down_delta = 200

        up_value = value + up_delta
        down_value = value - down_delta

        if value > up_value:
            y.loc[_index, "result"] = 1
            return

        if value < down_value:
            y.loc[_index, "result"] = 0
            return

Below a vectorized code to find if it is going up or down, and in what row

df=pd.DataFrame({'vals': np.random.random(int(10))*1000+5000}).astype('int64')
print(df.vals.values)

up_value = 6000
down_value = 3000
valsup = df.vals.values + 200*np.arange(df.shape[0])+200
valsdown = df.vals.values - 200*np.arange(df.shape[0])-200

#! argmax returns 0 if all false
# idx_up = np.argmax(valsup > up_value)
# idx_dwn= np.argmax(valsdown < down_value)

idx_up = np.argwhere(valsup > up_value)
idx_dwn= np.argwhere(valsdown < down_value)
idx_up = idx_up[0][0] if len(idx_up) else -1
idx_dwn = idx_dwn[0][0] if len(idx_dwn) else -1


if idx_up < 0 and idx_dwn<0:
    print(f" Not up nor down")
if idx_up < idx_dwn or idx_dwn<0:
    print(f" Result is positive, in position {idx_up}")
else: 
    print(f" Result is negative, in position {idx_dwn}")

For the sake of completeness, benchmarking itertuples() and the argwhere approach for 1000 elements:

• .itertuples(): 757µs
• arange + argwhere: 60µs