问题
I set up a system to filter the twitter real time stream sample. Obviously, the database writes are too slow to keep up with anything more complex than a couple of low-volume keywords. I implemented django-rq as a simple queuing system to push the tweets off into a redis based queue as they came in, and that works great. My issue is on the other side. The context to this question is I have a system that's running right now, with 1.5m tweets for analysis, and another 375,000 queued through redis. At current rates of performance, it's going to take me ~3 days to catch up, if I turn off the streams, which I don't want to. If I maintain the streams, then it'll take about a month, on my last estimates.
The database has a couple of million rows across two main tables now, and the writes are very slow. The optimal number of rq-workers seems to be four, and that's averaging out at 1.6 queue tasks per second. (Code of what's being enqueued below). I thought that maybe the issue was the opening of DB connections for every new queue task, so put CONN_MAX_AGE to 60, but that hasn't improved anything.
Having just tested this on localhost, I got in excess of 13 writes/second, on a Macbook 2011, with Chrome, etc etc running, but there are only a few thousand rows in that database, which leads me to believe it's size related. There are a couple of get_or_create commands I'm using (see below), which could be slowing things down, but can't see any other way through using them - I need to check if the user exists, and I need to check if the tweet already exists (I could possibly, I suspect, move the latter to a try/except, on the basis that tweets coming in from the live stream shouldn't already exist, for obvious reasons.) Would I get much performance gain out of that? As this is running still, I'm keen to optimise the code a bit and get some faster/more efficient workers in there so I can catch up! Would running a pre-vetting worker to batch things up work? (i.e. so I can batch create users that don't exist, or something similar?)
I"m running a 4 Core/8Gb Ram droplet on digital ocean, so feel this is some pretty terrible performance, and presumably code related. Where am I going wrong here?
(I've posted this here rather than code-review, as I think this is relevant to the Q&A format for SO, as I'm trying to solve a specific code problem, rather than 'how can I do this generally better?')
Note: I'm working in django 1.6 as this is code that I've had floating around for a while and wasn't confident about upgrading at the time - it's not public facing, so unless there's a compelling reason right now (like this performance issue), I wasn't going to upgrade (for this project).
Stream Listener:
class StdOutListener(tweepy.StreamListener):
def on_data(self, data):
# Twitter returns data in JSON format - we need to decode it first
decoded = json.loads(data)
#print type(decoded), decoded
# Also, we convert UTF-8 to ASCII ignoring all bad characters sent by users
try:
if decoded['lang'] == 'en':
django_rq.enqueue(read_both, decoded)
else:
pass
except KeyError,e:
print "Error on Key", e
except DataError, e:
print "DataError", e
return True
def on_error(self, status):
print status
Read User/Tweet/Both
def read_user(tweet):
from harvester.models import User
from django.core.exceptions import ObjectDoesNotExist, MultipleObjectsReturned
#We might get weird results where user has changed their details"], so first we check the UID.
#print "MULTIPLE USER DEBUG", tweet["user"]["id_str"]
try:
current_user = User.objects.get(id_str=tweet["user"]["id_str"])
created=False
return current_user, created
except ObjectDoesNotExist:
pass
except MultipleObjectsReturned:
current_user = User.objects.filter(id_str=tweet["user"]["id_str"])[0]
return current_user, False
if not tweet["user"]["follow_request_sent"]:
tweet["user"]["follow_request_sent"] = False
if not tweet["user"]["following"]:
tweet["user"]["following"] = False
if not tweet["user"]["description"]:
tweet["user"]["description"] = " "
if not tweet["user"]["notifications"]:
tweet["user"]["notifications"] = False
#If that doesn't work"], then we'll use get_or_create (as a failback rather than save())
from dateutil.parser import parse
if not tweet["user"]["contributors_enabled"]:
current_user, created = User.objects.get_or_create(
follow_request_sent=tweet["user"]["follow_request_sent"],
_json = {},
verified = tweet["user"]["verified"],
followers_count = tweet["user"]["followers_count"],
profile_image_url_https = tweet["user"]["profile_image_url_https"],
id_str = tweet["user"]["id_str"],
listed_count = tweet["user"]["listed_count"],
utc_offset = tweet["user"]["utc_offset"],
statuses_count = tweet["user"]["statuses_count"],
description = tweet["user"]["description"],
friends_count = tweet["user"]["friends_count"],
location = tweet["user"]["location"],
profile_image_url= tweet["user"]["profile_image_url"],
following = tweet["user"]["following"],
geo_enabled = tweet["user"]["geo_enabled"],
profile_background_image_url =tweet["user"]["profile_background_image_url"],
screen_name = tweet["user"]["screen_name"],
lang = tweet["user"]["lang"],
profile_background_tile = tweet["user"]["profile_background_tile"],
favourites_count = tweet["user"]["favourites_count"],
name = tweet["user"]["name"],
notifications = tweet["user"]["notifications"],
url = tweet["user"]["url"],
created_at = parse(tweet["user"]["created_at"]),
contributors_enabled = False,
time_zone = tweet["user"]["time_zone"],
protected = tweet["user"]["protected"],
default_profile = tweet["user"]["default_profile"],
is_translator = tweet["user"]["is_translator"]
)
else:
current_user, created = User.objects.get_or_create(
follow_request_sent=tweet["user"]["follow_request_sent"],
_json = {},
verified = tweet["user"]["verified"],
followers_count = tweet["user"]["followers_count"],
profile_image_url_https = tweet["user"]["profile_image_url_https"],
id_str = tweet["user"]["id_str"],
listed_count = tweet["user"]["listed_count"],
utc_offset = tweet["user"]["utc_offset"],
statuses_count = tweet["user"]["statuses_count"],
description = tweet["user"]["description"],
friends_count = tweet["user"]["friends_count"],
location = tweet["user"]["location"],
profile_image_url= tweet["user"]["profile_image_url"],
following = tweet["user"]["following"],
geo_enabled = tweet["user"]["geo_enabled"],
profile_background_image_url =tweet["user"]["profile_background_image_url"],
screen_name = tweet["user"]["screen_name"],
lang = tweet["user"]["lang"],
profile_background_tile = tweet["user"]["profile_background_tile"],
favourites_count = tweet["user"]["favourites_count"],
name = tweet["user"]["name"],
notifications = tweet["user"]["notifications"],
url = tweet["user"]["url"],
created_at = parse(tweet["user"]["created_at"]),
contributors_enabled = tweet["user"]["contributers_enabled"],
time_zone = tweet["user"]["time_zone"],
protected = tweet["user"]["protected"],
default_profile = tweet["user"]["default_profile"],
is_translator = tweet["user"]["is_translator"]
)
#print "CURRENT USER:""], type(current_user)"], current_user
#current_user"], created = User.objects.get_or_create(current_user)
return current_user, created
def read_tweet(tweet, current_user):
import logging
logger = logging.getLogger('django')
from datetime import date, datetime
#print "Inside read_Tweet"
from harvester.models import Tweet
from django.core.exceptions import ObjectDoesNotExist, MultipleObjectsReturned
from django.db import DataError
#We might get weird results where user has changed their details"], so first we check the UID.
#print tweet_data["created_at"]
from dateutil.parser import parse
tweet["created_at"] = parse(tweet["created_at"])
try:
#print "trying tweet_data["id"
current_tweet =Tweet.objects.get(id_str=tweet["id_str"])
created=False
return current_user, created
except ObjectDoesNotExist:
pass
except MultipleObjectsReturned:
current_tweet =Tweet.objects.filter(id_str=tweet["id_str"])[0]
try:
current_tweet, created = Tweet.objects.get_or_create(
truncated=tweet["truncated"],
text=tweet["text"],
favorite_count=tweet["favorite_count"],
author = current_user,
_json = {},
source=tweet["source"],
retweeted=tweet["retweeted"],
coordinates = tweet["coordinates"],
entities = tweet["entities"],
in_reply_to_screen_name = tweet["in_reply_to_screen_name"],
id_str = tweet["id_str"],
retweet_count = tweet["retweet_count"],
favorited = tweet["favorited"],
user = tweet["user"],
geo = tweet["geo"],
in_reply_to_user_id_str = tweet["in_reply_to_user_id_str"],
lang = tweet["lang"],
created_at = tweet["created_at"],
place = tweet["place"])
print "DEBUG", current_user, current_tweet
return current_tweet, created
except DataError, e:
#Catchall to pick up non-parsed tweets
print "DEBUG ERROR", e, tweet
return None, False
def read_both(tweet):
current_user, created = read_user(tweet)
current_tweet, created = read_tweet(tweet, current_user)
回答1:
I eventually managed to cobble together an answer from some redditors and a couple of other things.
Fundamentally, though I was doing a double lookup on the id_str field, which wasn't indexed. I added indexes db_index=True to that field on both read_tweet and read_user, and moved read tweet to a try/except Tweet.objects.create approach, falling back to the get_or_create if there's a problem, and saw a 50-60x speed improvement, with the workers now being scalable - if I add 10 workers, I get 10x speed.
I currently have one worker that's happily processing 6 or so tweets a second. Next up I'll add a monitoring daemon to check the queue size and add extra workers if it's still increasing.
tl;dr - REMEMBER INDEXING!
来源：https://stackoverflow.com/questions/30969171/slow-django-database-operations-on-large-ish-dataset