Multiple POST request handling concurrently in Python 2.7 - python

I have a tornado server listening to port 6789 for POST requests on "/train" and "/predict". train method might take upto 3 hrs to complete, while predict might return in 2 minutes. I want them to be handled concurrently. So even when "/train" is running, if a POST request for "/predict" arrives, it can handle that concurrently and return its output without waiting for "/train" to complete.
I have tried using ThreadPool but it still doesn't run concurrently.
My present code is as follows. It functions but if request to train is made and then request to predict is made. It waits for train to complete before handling predict. Assume train and predict functions are present and don't take any parameters.
import logging
import time
import threading
from multiprocessing.pool import ThreadPool
import flask
from tornado import wsgi, httpserver, ioloop
from flask import Flask
from train_script import train
from predict_script import predict
app = Flask(__name__)
#app.route("/train", methods=['POST'])
def train_run():
payload = flask.request.get_json(silent=True)
if payload is not None:
try:
async_result = pool.apply_async(train)
response = async_result.get()
resp = flask.jsonify(response)
resp.status_code = 200
except Exception as ex:
resp = flask.jsonify({"status": "Failure"})
resp.status_code = 500
else:
resp = flask.jsonify({"status": "Failure"})
resp.status_code = 500
return resp
#app.route("/predict", methods=['POST'])
def predict_run():
payload = flask.request.get_json(silent=True)
if payload is not None:
try:
async_result = pool.apply_async(predict)
response = async_result.get()
resp = flask.jsonify(response)
resp.status_code = 200
except Exception as ex:
resp = flask.jsonify({"status": "Failure"})
resp.status_code = 500
else:
resp = flask.jsonify({"status": "Failure"})
resp.status_code = 500
return resp
if __name__ == "__main__":
port = 6789
http_server = httpserver.HTTPServer(wsgi.WSGIContainer(app))
pool = ThreadPool(processes=10)# Expects max concurrent requests to be 10
http_server.listen(port)
logging.info("Tornado server starting on port {}".format(port))
ioloop.IOLoop.instance().start()

Tornado's WSGIContainer does not support any kind of concurrency. Either use Tornado's RequestHandler interfaces without Flask or WSGI, or use Flask with gunicorn or uwsgi. You gain almost nothing and lose a lot by combining Tornado with WSGI frameworks, so this is only useful in certain specialized situations.

Related

Downloading images with gevent

My task is to download 1M+ images from a given list of urls. What is the recommended way to do so?
After having read Greenlet Vs. Threads I looked into gevent, but I fail to get it reliably to run. I played around with a test set of 100 urls and sometimes it finishes in 1.5s but sometimes it takes over 30s which is strange as the timeout* per request is 0.1, so it should never take more than 10s.
*see below in code
I also looked into grequests but they seem to have issues with exception handling.
My 'requirements' are that I can
inspect the errors raised while downloading (timeouts, corrupt images...),
monitor the progress of the number of processed images and
be as fast as possible.
from gevent import monkey; monkey.patch_all()
from time import time
import requests
from PIL import Image
import cStringIO
import gevent.hub
POOL_SIZE = 300
def download_image_wrapper(task):
return download_image(task[0], task[1])
def download_image(image_url, download_path):
raw_binary_request = requests.get(image_url, timeout=0.1).content
image = Image.open(cStringIO.StringIO(raw_binary_request))
image.save(download_path)
def download_images_gevent_spawn(list_of_image_urls, base_folder):
download_paths = ['/'.join([base_folder, url.split('/')[-1]])
for url in list_of_image_urls]
parameters = [[image_url, download_path] for image_url, download_path in
zip(list_of_image_urls, download_paths)]
tasks = [gevent.spawn(download_image_wrapper, parameter_tuple) for parameter_tuple in parameters]
for task in tasks:
try:
task.get()
except Exception:
print 'x',
continue
print '.',
test_urls = # list of 100 urls
t1 = time()
download_images_gevent_spawn(test_urls, 'download_temp')
print time() - t1
I think it will be better to stick with urllib2, by example of https://github.com/gevent/gevent/blob/master/examples/concurrent_download.py#L1
Try this code, I suppose it is what you're asking.
import gevent
from gevent import monkey
# patches stdlib (including socket and ssl modules) to cooperate with other greenlets
monkey.patch_all()
import sys
urls = sorted(chloya_files)
if sys.version_info[0] == 3:
from urllib.request import urlopen
else:
from urllib2 import urlopen
def download_file(url):
data = urlopen(url).read()
img_name = url.split('/')[-1]
with open('c:/temp/img/'+img_name, 'wb') as f:
f.write(data)
return True
from time import time
t1 = time()
tasks = [gevent.spawn(download_file, url) for url in urls]
gevent.joinall(tasks, timeout = 12.0)
print "Sucessful: %s from %s" % (sum(1 if task.value else 0 for task in tasks), len(tasks))
print time() - t1
There's a simple solution using gevent and Requests simple-requests
Use Requests Session for HTTP persistent connection. Since gevent makes Requests asynchronous, I think there's no need for timeout in HTTP requests.
By default, requests.Session caches TCP connections (pool_connections) for 10 hosts and limits 10 concurrent HTTP requests per cached TCP connections (pool_maxsize). The default configuration should be tweaked to suit the need by explicitly creating an http adapter.
session = requests.Session()
http_adapter = requests.adapters.HTTPAdapter(pool_connections=100, pool_maxsize=100)
session.mount('http://', http_adapter)
Break the tasks as producer-consumer. Image downloading is producer task and Image processing is consumer task.
If the image processing library PIL is not asynchronous, it may block producer coroutines. If so, consumer pool can be a gevent.threadpool.ThreadPool. f.e.
from gevent.threadpool import ThreadPool
consumer = ThreadPool(POOL_SIZE)
This is an overview of how it can be done. I didn't test the code.
from gevent import monkey; monkey.patch_all()
from time import time
import requests
from PIL import Image
from io import BytesIO
import os
from urlparse import urlparse
from gevent.pool import Pool
def download(url):
try:
response = session.get(url)
except Exception as e:
print(e)
else:
if response.status_code == requests.codes.ok:
file_name = urlparse(url).path.rsplit('/',1)[-1]
return (response.content,file_name)
response.raise_for_status()
def process(img):
if img is None:
return None
img, name = img
img = Image.open(BytesIO(img))
path = os.path.join(base_folder, name)
try:
img.save(path)
except Exception as e:
print(e)
else:
return True
def run(urls):
consumer.map(process, producer.imap_unordered(download, urls))
if __name__ == '__main__':
POOL_SIZE = 300
producer = Pool(POOL_SIZE)
consumer = Pool(POOL_SIZE)
session = requests.Session()
http_adapter = requests.adapters.HTTPAdapter(pool_connections=100, pool_maxsize=100)
session.mount('http://', http_adapter)
test_urls = # list of 100 urls
base_folder = 'download_temp'
t1 = time()
run(test_urls)
print time() - t1
I will suggest to pay attention to Grablib http://grablib.org/
It is an asynchronic parser based on pycurl and multicurl.
Also it tryes to automatically solve network error (like try again if timeout, etc).
I believe the Grab:Spider module will solve your problems for 99%.
http://docs.grablib.org/en/latest/index.html#spider-toc

Make Post Requests with Files Simultaneously

I write a simple server and it runs well.
So I want to write some codes which will make many post requests to my server simultaneously to simulate a pressure test. I use python.
Suppose the url of my server is http://myserver.com.
file1.jpg and file2.jpg are the files needed to be uploaded to the server.
Here is my testing code. I use threading and urllib2.
async_posts.py
from Queue import Queue
from threading import Thread
from poster.encode import multipart_encode
from poster.streaminghttp import register_openers
import urllib2, sys
num_thread = 4
queue = Queue(2*num_thread)
def make_post(url):
register_openers()
data = {"file1": open("path/to/file1.jpg"), "file2": open("path/to/file2.jpg")}
datagen, headers = multipart_encode(data)
request = urllib2.Request(url, datagen, headers)
start = time.time()
res = urllib2.urlopen(request)
end = time.time()
return res.code, end - start # Return the status code and duration of this request.
def deamon():
while True:
url = queue.get()
status, duration = make_post(url)
print status, duration
queue.task_done()
for _ in range(num_thread):
thd = Thread(target = daemon)
thd.daemon = True
thd.start()
try:
urls = ["http://myserver.com"] * num_thread
for url in urls:
queue.put(url)
queue.join()
except KeyboardInterrupt:
sys.exit(1)
When num_thread is small (ex: 4), my code runs smoothly. But as I switch num_thread to slightly larger number, say 10, all the threading things break down and keep throwing httplib.BadStatusLine error.
I don't know why my code goes wrong or maybe there is better way to do this?
A a reference, my server is written in python using flask and gunicorn.
Thanks in advance.

How to robustly debug gevent 'This operation would block forever' exception?

I have some concurrent code using gevent:
gevent.monkey.patch_all()
....
jobs = [gevent.spawn(self.generate_resource_cache, idx, resource, id_fields_map[resource])
for idx, resource in enumerate(available_resources)]
gevent.joinall(jobs)
My generate_resource_cache method is using requests library which means I have to mankeypatch gevent.
With requests, I'm using custom session:
def get_session(self):
s = Settings.Instance()
if not self.session:
retry = Retry(total=s.TOTAL_RETRIES, backoff_factor=s.BACKOFF_FACTOR,
status_forcelist=(range(400, 421) + range(500, 505)))
size = s.CONCURRENT_SIZE
adapter = requests.adapters.HTTPAdapter(pool_connections=size, pool_maxsize=size, pool_block=True, max_retries=retry)
self.session = requests_cache.CachedSession(s.CACHE_NAME, backend='sqlite',
fast_save=s.FAST_SAVE, allowable_methods=('GET', 'POST')) if s.CACHING else requests.Session()
if s.PROXIES:
self.session.proxies = s.PROXIES
self.session.headers.update({'X_HTTP_METHOD_OVERRIDE' : 'get'})
self.session.mount('http://', adapter)
self.session.mount('https://', adapter)
return self.session
After few hours of running this script it prints to stdout:
WARNING:root:epoll module not found; using select()
And to stderr:
Exception caught: This operation would block forever
I'm wondering what can I do to better understand why I'm getting this exception in the first place and how can I prevent it from happening?

Parallel/Concurrent http request sending from urllib2 in python

I was trying to send out http post requests in parallel to a web service. To be specific, I'd like to do load test the web service under concurrent requests environment in terms of response time. So I plan to use thread and urllib2 to achieve the job - each thread carries out a http request by urllib2.
Here is how I did it:
import urllib2 as u2
import time
from threading import Thread
def run_job():
try:
req = u2.Request(
url = **the web service url**,
data = **data send to the web service**,
headers = **http header web service requires**,
)
opener = u2.build_opener()
u2.install_opener(opener)
start_time = time.time()
response = u2.urlopen(req, timeout = 60)
end_time = time.time()
html = response.read()
code = response.code
response.close()
if code == 200:
print end_time - start_time
else:
print -1
except Exception, e:
print -2
if __name__ == "__main__":
N = 1
if len(sys.argv) > 1:
N = int(sys.argv[1])
threads = []
for i in range(1, N):
t = Thread(target=run_job, args=())
threads.append(t)
[x.start() for x in threads]
[x.join() for x in threads]
In the meantime, I use fidller2 to capture the requests sending out. Fiddler is a tool to compose http request and send it out, and it can also capture the http requests through the host.
When I looked at the fidller2, those requests are sending out one by one instead of sending out all together at a time, which is what I expect to happen. If what represented in fiddler is right (requests are sending out one by one), to my knowledge, I think there must be some queue that the requests are waiting in. Could someone shed some light what happened behind this? And if it possible, how to do the real parallel requests?
Also, I have put two time stamps before and after the request takes place, if requests are waiting in queue after urllib2.urlopen is executed, then the delta of two time stamps includes the time spending in waiting queue. Is it possible to be more precise - to measure the time between request sending out and received?
Many Thanks,

How to make this Twisted Python Proxy faster?

The code below is an HTTP proxy for content filtering. It uses GET to send the URL of the current site to the server, where it processes it and responds. It runs VERY, VERY, VERY slow. Any ideas on how to make it faster?
Here is the code:
from twisted.internet import reactor
from twisted.web import http
from twisted.web.proxy import Proxy, ProxyRequest
from Tkinter import *
#import win32api
import urllib2
import urllib
import os
import webbrowser
cwd = os.path.abspath(sys.argv[0])[0]
proxies = {}
user = "zachb"
class BlockingProxyRequest(ProxyRequest):
def process(self):
params = {}
params['Location']= self.uri
params['User'] = user
params = urllib.urlencode(params)
req = urllib.urlopen("http://weblock.zbrowntechnology.info/ProgFiles/stats.php?%s" % params, proxies=proxies)
resp = req.read()
req.close()
if resp == "allow":
pass
else:
self.transport.write('''BLOCKED BY ADMIN!''')
self.transport.loseConnection()
ProxyRequest.process(self)
class BlockingProxy(Proxy):
requestFactory = BlockingProxyRequest
factory = http.HTTPFactory()
factory.protocol = BlockingProxy
reactor.listenTCP(8000, factory)
reactor.run()
Anyone have any ideas on how to make this run faster? Or even a better way to write it?
The main cause of slowness in this proxy is probably these three lines:
req = urllib.urlopen("http://weblock.zbrowntechnology.info/ProgFiles/stats.php?%s" % params, proxies=proxies)
resp = req.read()
req.close()
A normal Twisted-based application is single threaded. You have to go out of your way to get threads involved. That means that whenever a request comes in, you are blocking the one and only processing thread on this HTTP request. No further requests are processed until this HTTP request completes.
Try using one of the APIs in twisted.web.client, (eg Agent or getPage). These APIs don't block, so your server will handle concurrent requests concurrently. This should translate into much smaller response times.

Categories