简介
这是一篇有关如何将大量的数据放入有限的内存中的简略教程。
与客户工作时,有时会发现他们的数据库实际上只是一个csv或excel文件仓库,你只能将就着用,经常需要在不更新他们的数据仓库的情况下完成工作。大部分情况下,如果将这些文件存储在一个简单的数据库框架中或许更好,但时间可能不允许。这种方法对时间、机器硬件和所处环境都有要求。
下面介绍一个很好的例子:假设有一堆表格(没有使用neo4j、mongodb或其他类型的数据库,仅仅使用csvs、tsvs等格式存储的表格),如果将所有表格组合在一起,得到的数据帧太大,无法放入内存。所以第一个想法是:将其拆分成不同的部分,逐个存储。这个方案看起来不错,但处理起来很慢。除非我们使用多核处理器。
目标
这里的目标是从所有职位中(大约1万个),找出相关的的职位。将这些职位与政府给的职位代码组合起来。接着将组合的结果与对应的州(行政单位)信息组合起来。然后用通过word2vec生成的属性信息在我们的客户的管道中增强已有的属性。
这个任务要求在短时间内完成,谁也不愿意等待。想象一下,这就像在不使用标准的关系型数据库的情况下进行多个表的连接。
数据
示例脚本
下面的是一个示例脚本,展示了如何使用multiprocessing来在有限的内存空间中加速操作过程。脚本的第一部分是和特定任务相关的,可以自由跳过。请着重关注第二部分,这里侧重的是multiprocessing引擎。
#import the necessary packages
import pandas as pd
import us
import numpy as np
from multiprocessing import pool,cpu_count,queue,manager
# the data in one particular column was number in the form that horrible excel version
# of a number where ‘12000’ is ‘12,000’ with that beautiful useless comma in there.
# did i mention i excel bothers me?
# instead of converting the number right away, we only convert them when we need to
def median_maker(column):
return np.median([int(x.replace(‘,’,”)) for x in column])
# dictionary_of_dataframes contains a dataframe with information for each title; e.g title is ‘data scientist’
# related_title_score_df is the dataframe of information for the title; columns = [‘title’,’score’]
### where title is a similar_title and score is how closely the two are related, e.g. ‘data analyst’, 0.871
# code_title_df contains columns [‘code’,’title’]
# oes_data_df is a huge dataframe with all of the bureau of labor statistics(bls) data for a given time period (yay free data, boo bad census data!)
def job_title_location_matcher(title,location):
try:
related_title_score_df = dictionary_of_dataframes[title]
# we limit dataframe1 to only those related_titles that are above
# a previously established threshold
related_title_score_df = related_title_score_df[title_score_df[‘score’]>80]
#we merge the related titles with another table and its codes
codes_reltitles_scores = pd.merge(code_title_df,related_title_score_df)
codes_reltitles_scores = codes_reltitles_scores.drop_duplicates()
# merge the two dataframes by the codes
merged_df = pd.merge(codes_reltitles_scores, oes_data_df)
#limit the bls data to the state we want
all_merged = merged_df[merged_df[‘area_title’]==str(us.states.lookup(location).name)]
#calculate some summary statistics for the time we want
group_med_emp,group_mean,group_pct10,group_pct25,group_median,group_pct75,group_pct90 = all_merged[[‘tot_emp’,’a_mean’,’a_pct10′,’a_pct25′,’a_median’,’a_pct75′,’a_pct90′]].apply(median_maker)
row = [title,location,group_med_emp,group_mean,group_pct10,group_pct25, group_median, group_pct75, group_pct90]
#convert it all to strings so we can combine them all when writing to file
row_string = [str(x) for x in row]
return row_string
except:
# if it doesnt work for a particular title/state just throw it out, there are enough to make this insignificant
‘do nothing’
这里发生了神奇的事情:
#runs the function and puts the answers in the queue
def worker(row, q):
ans = job_title_location_matcher(row[0],row[1])
q.put(ans)
# this writes to the file while there are still things that could be in the queue
# this allows for multiple processes to write to the same file without blocking eachother
def listener(q):
f = open(filename,’wb’)
while 1:
m = q.get()
if m ==’kill’:
break
f.write(‘,’.join(m) + ‘n’)
f.flush()
f.close()
def main():
#load all your data, then throw out all unnecessary tables/columns
filename = ‘skill_test_pool.txt’
#sets up the necessary multiprocessing tasks
manager = manager()
q = manager.queue()
pool = pool(cpu_count() + 2)
watcher = pool.map_async(listener,(q,))
jobs = []
#titles_states is a dataframe of millions of job titles and states they were found in
for i in titles_states.iloc:
job = pool.map_async(worker, (i, q))
jobs.append(job)
for job in jobs:
job.get()
q.put(‘kill’)
pool.close()
pool.join()
if __name__ == “__main__”:
main()
由于每个数据帧的大小都不同(总共约有100gb),所以将所有数据都放入内存是不可能的。通过将最终的数据帧逐行写入内存,但从来不在内存中存储完整的数据帧。我们可以完成所有的计算和组合任务。这里的“标准方法”是,我们可以仅仅在“job_title_location_matcher”的末尾编写一个“write_line”方法,但这样每次只会处理一个实例。根据我们需要处理的职位/州的数量,这大概需要2天的时间。而通过multiprocessing,只需2个小时。
虽然读者可能接触不到本教程处理的任务环境,但通过multiprocessing,可以突破许多计算机硬件的限制。本例的工作环境是c3.8xl ubuntu ec2,硬件为32核60gb内存(虽然这个内存很大,但还是无法一次性放入所有数据)。这里的关键之处是我们在60gb的内存的机器上有效的处理了约100gb的数据,同时速度提升了约25倍。通过multiprocessing在多核机器上自动处理大规模的进程,可以有效提高机器的利用率。也许有些读者已经知道了这个方法,但对于其他人,可以通过multiprocessing能带来非常大的收益。顺便说一句,这部分是skill assets in the job-market这篇博文的延续。