我想用python尽可能快地下载文件。这是我的代码
import pandas as pd
import requests
from requests_futures.sessions import FuturesSession
import os
import pathlib
from timeit import default_timer as timer
class AsyncDownloader:
"""Download files asynchronously"""
__urls = set()
__dest_path = None
__user_agent = 'Mozilla/5.0 (Windows NT 6.1; Win64; x64; rv:58.0) Gecko/20100101 Firefox/58.0'
__read_timeout = 60
__connection_timeout = 30
__download_count = 0 # unlimited
# http://www.browserscope.org/?category=network
__worker_count = 17 # No of threads to spawn
__chunk_size = 1024
__download_time = -1
__errors = []
# TODO Fetch only content of a specific type from a csv
# TODO Improve code structure so that it can be used as a commandline tool
def set_source_csv(self, source_path, column_name):
self.source_path = source_path
self.column_name = column_name
try:
my_csv = pd.read_csv(source_path, usecols=[self.column_name], chunksize=10)
except ValueError:
print("The column name doesn't exist")
return
else:
# No exception whatsoever
for chunk in my_csv:
AsyncDownloader.__urls.update(set(getattr(chunk, self.column_name)))
def set_destination_path(self, dest_path):
if dest_path.endswith('/'):
dest_path = dest_path[:-1]
self.dest_path = dest_path
# TODO Add exception in case we can't create the directory
pathlib.Path(self.dest_path).mkdir(parents=True, exist_ok=True)
if os.access(self.dest_path, os.W_OK):
AsyncDownloader.__dest_path = pathlib.Path(self.dest_path).resolve()
def set_user_agent(self, useragent):
self.useragent = useragent
AsyncDownloader.__user_agent = self.useragent
def set_connection_timeout(self, ctimeout_secs):
self.timeout_secs = ctimeout_secs
if self.timeout_secs >= 0:
AsyncDownloader.__connection_timeout = self.timeout_secs
def set_read_timeout(self, rtimeout_secs):
self.timeout_secs = rtimeout_secs
if self.timeout_secs >= 0:
AsyncDownloader.__read_timeout = self.timeout_secs
def set_download_count(self, file_count):
self.file_count = file_count
if self.file_count > 0:
AsyncDownloader.__download_count = self.file_count
def set_worker_count(self, worker_count):
self.worker_count = worker_count
if self.worker_count > 0:
AsyncDownloader.__worker_count = self.worker_count
def set_chunk_size(self, chunk_size):
self.chunk_size = chunk_size
if self.chunk_size > 0:
AsyncDownloader.__chunk_size = self.chunk_size
def print_urls(self):
print(AsyncDownloader.__urls)
def get_download_time(self):
return AsyncDownloader.__download_time
def get_errors(self):
return AsyncDownloader.__errors
def download(self):
start = timer()
try:
session = FuturesSession(max_workers=AsyncDownloader.__worker_count)
session.headers.update({'user-agent': AsyncDownloader.__user_agent})
session.request(AsyncDownloader.__connection_timeout,
AsyncDownloader.__connection_timeout, stream=True)
results = []
# Give an accurate file count even if we don't have to download it as it a;ready exist
file_count = 0
for url in AsyncDownloader.__urls:
filename = os.path.basename(url)
# check if we need only a limited number of files
if AsyncDownloader.__download_count != 0:
# No need to download file if it already exist
if pathlib.Path(AsyncDownloader.__dest_path / filename).is_file():
file_count += 1
continue
else:
if file_count < AsyncDownloader.__download_count:
file_count += 1
results.append(session.get(url))
else:
if not pathlib.Path(AsyncDownloader.__dest_path / filename).is_file():
results.append(session.get(url))
for result in results:
# wait for the response to complete, if it hasn't already
response = result.result()
filename = os.path.basename(response.url)
if response.status_code == 200:
with open(pathlib.Path(AsyncDownloader.__dest_path / filename).resolve(), 'wb') as fd:
for chunk in response.iter_content(chunk_size=AsyncDownloader.__chunk_size):
if chunk: # filter out keep-alive new chunks
fd.write(chunk)
end = timer()
AsyncDownloader.__download_time = end - start
except requests.exceptions.HTTPError as errh:
AsyncDownloader.__errors.append("Http Error:" + errh)
# print("Http Error:", errh)
except requests.exceptions.ConnectionError as errc:
AsyncDownloader.__errors.append("Error Connecting:" + errc)
# print("Error Connecting:", errc)
except requests.exceptions.Timeout as errt:
AsyncDownloader.__errors.append("Timeout Error:" + errt)
# print("Timeout Error:", errt)
except requests.exceptions.RequestException as err:
AsyncDownloader.__errors.append("OOps: Something Else" + err)
else:
return
下面的代码做了一个非常糟糕的假设。事实上，我假设第一个网址将首先完成，当然这是不正确的。
# wait for the response to complete, if it hasn't already
response = result.result()
如何确保只处理已完成的请求，而不是以有效的方式采取上述假设？
我将不胜感激任何有关如何提高性能的建议。
亲切的问候
即使连接按顺序完成，您仍然按顺序处理文件。第二个文件必须等待第一个文件写入，依此类推。因此，你可以做的最好的事情是并行处理所有内容（尽管有GIL，这可以完成，因为像写入磁盘和从网络读取等操作会释放它）。基本上，使用常规requests库（而不是requests-futures）并为每个请求创建未来/线程+文件处理。
还有更多的方法可以让它更快，比如在写入时继续下载块（即两个线程，一个用于请求，一个用于文件处理）。通过制作multi-part请求并行读取块，这是“下载加速器”的领域，您可能不希望代码中出现这种复杂性。
编辑：此外，分块下载是懒惰的，这意味着您只是并行地进行初始请求，但实际的分块文件下载是按顺序进行的，因为它是在主线程中完成的。因此，您当前的方法并不比完全同步的方法好多少。上述建议仍然有效。
为了处理你的代码，我创建了一个.csv文件，其中包含几个网站中几个robots.txt文件的链接：GitHub，UDemy，YouTube。
调试后，第一个结果
response = result.result()
是（按此顺序）：UDemy，YouTube，GitHub。为了记录，每个robots.txt的大小增加的顺序与我得到的结果相同。这意味着开始没有问题，尽管我按特定顺序设置了.csv文件，结果按照文件首次下载的顺序排列。
我将不胜感激任何有关如何提高性能的建议。
至于性能，您可以通过创建用于将响应写入文件的线程或使用异步IO库（例如Tinche/aiofiles）来提高速度。
如果你想更进一步，可以尝试通过使用Python的替代实现来改进程序本身的性能，例如PyPy
最简单的方法是不需要任何线程或特殊的异步代码：只需使用常规的requests库及其内置的流媒体选项。你说response = session.get(url, stream=True)然后使用response.iter_content(chunk_size=1024)（例如）一次访问一个块的下载信息。这是一个功能性的例子：
import requests
import os
def stream_multiple(urls):
responses = {url: requests.get(url, stream=True) for url in urls)
streams = {url: responses[url].iter_content(chunk_size=1024)
for url in urls}
handles = {url: open(os.path.basename(url), 'wb') for url in urls}
while streams:
for url in list(streams.keys()):
try:
chunk = next(streams[url])
print("Received {} bytes for {}".format(len(chunk), url))
handles[url].write(chunk)
except StopIteration: # no more contenet
handles[url].close()
streams.pop(url)
样本输出：
rat@pandion:~/tmp$ python smu.py
Received 1296 bytes for http://www.gutenberg.org/files/9490/9490-0.txt
Received 1882 bytes for http://www.gutenberg.org/ebooks/21497.txt.utf-8
Received 1524 bytes for http://www.gutenberg.org/files/1729/1729-0.txt
Received 1508 bytes for http://www.gutenberg.org/ebooks/21790.txt.utf-8
Received 1826 bytes for http://www.gutenberg.org/files/9490/9490-0.txt
Received 2349 bytes for http://www.gutenberg.org/ebooks/21497.txt.utf-8
Received 1834 bytes for http://www.gutenberg.org/files/1729/1729-0.txt
Received 1838 bytes for http://www.gutenberg.org/ebooks/21790.txt.utf-8
Received 2009 bytes for http://www.gutenberg.org/files/9490/9490-0.txt
...
使用线程或多处理可能会略微提高性能，但我怀疑这会更好。几乎在所有情况下，将数据写入磁盘的速度远远快于从网络接收数据。
如果你不担心“猴子补丁”你可以使用gevent
import gevent.monkey
import requests
CONNECTIONS = 10
gevent.monkey.patch_all() # debug in PyCharm: https://blog.jetbrains.com/pycharm/2012/08/gevent-debug-support/
import gevent.pool
def your_request_without_any_changes(url):
return requests.get(url)
pool = gevent.pool.Pool(CONNECTIONS)
for response in pool.imap_unordered(your_request_without_any_changes, ['http://www.google.com'] * 100):
print(response.status_code)
当我们等待响应时，gevent使用“事件循环”和补丁请求库（实际上它发生在较低级别）来切换到另一个任务。