Crawling with Python, Selenium and Docker
TD;DR: Using selenium inside a docker container to crawl webistes that need javascript or user interaction + a cluster of those using docker swarm.
While simple HTTP requests are good enough 90% to get the data you want from a website I am always looking for better ways to optimize my crawlers specially in websites that require javascript and user interaction, a login or a click in the right place sometimes give you the access you need. I am looking at to you government websites!
Recently I have seen more solutions to some of these problems in python such as Splash from ScrapingHub that is basically a QT browser with an scriptable API. I havent tried it and it definetly looks like a viable option but if I am going to render a webpage I want to do it in a "real" (Chrome) browser.
An easy way to use Chrome (or Firefox or any other popular browser) with an scriptable and multi-language API is using Selenium, which is generally used to test websites, making clicks, fake/real logins and more. To make this process reproducible I used docker to install Selenium, Chrome and the Selenium Chrome driver in a single container, this image can be found at docker-selenium but more and in docker hub also more (and probably) better selenium images can be found in Docker Hub and the result should be the same.
Browser¶
Getting Selenium running its just as easy as having docker and:
$ docker pull danielfrg/selenium
$ docker run -it -p 4444:4444 danielfrg/selenium
And now you'll have a selenium instance running in port 4444. If you use docker-machine get the docker IP from there and point your brower to that port. In my case http://192.168.99.101:4444/wd/hub - this is the same URL you use to point to the Selenium Python API.
Now in Python we can use it for example to crawl the homepage of yelp.com.
Disclaimer: This is just an example. I don't know the terms of service of yelp, if you want to try this in a bigger scale read the TOS.
Code¶
Now into the code: Create a Remote Driver, point it to the Selenium URL. Query for yelp.com and just for fun get a screenshot.
from selenium import webdriver
from selenium.webdriver.common.desired_capabilities import DesiredCapabilities
driver = webdriver.Remote(command_executor='http://192.168.99.101:4444/wd/hub',
desired_capabilities=DesiredCapabilities.CHROME)
driver.get("http://www.yelp.com")
image = driver.get_screenshot_as_base64()
from IPython.display import HTML
HTML("""<img src="data:image/png;base64,{0}">""".format(image))
""".format(image))
Here you can see that the render of the website is correct and in my case it pointed me to the Austin website.
Now we can use the Python API to can query for the content that we want from the website. To help me a little bit while looking at HTML source and searching for the css (or xpath) that I need I love the selector gadget plugin. For this example I want to get the "Best of Yelp: Austin" section that has some javascript and a scroll button with the different categories. The Selection in this case would be: #best-of-yelp-module .navigation li.
best = driver.find_element_by_id('best-of-yelp-module')
navigation = best.find_element_by_class_name('navigation')
sections = navigation.find_elements_by_tag_name('li')
len(sections)
23
So now I know there are 23 links on that section, 21 categories since there are two buttons to scroll.
Now I am going to click on each one of this buttons this is going to change the content of another part of the site: .main-content and after that I am getting the name of the category and the list of businesses .main-content .biz-name. A little note: I am waiting for 1 second after the click() event while the content of the webpage updates, this is to wait for request that might happen and the fade in/out effects.
import time
biz = {}
for section in sections:
section.click()
time.sleep(1)
content = best.find_element_by_class_name('main-content')
sec_name = content.text.split('\n')[0]
biz_names = content.find_elements_by_class_name('biz-name')
biz_names = [name.text for name in biz_names if name.text]
biz[sec_name] = biz_names
After about 30 seconds I have the content on a dictionary and I can take a look at it.
biz
{u'Active Life': [u'Rocket Electrics',
u'Barton Creek Greenbelt',
u'Live Love Paddle',
u'Eastside Yoga',
u'Town Lake Metropolitan Parks'],
u'Arts & Entertainment': [u'Bijou Studio',
u'Broadway in Austin at Texas Performing Arts',
u'Skylark Lounge',
u'Esther\u2019s Follies',
u'Pinballz Arcade'],
u'Automotive': [u'Brakes To Go - Mobile Brake Repair',
u'Terry\u2019s Burnet Road Vehicle Inspections',
u'Juke Auto',
u'Auto Glass Solutions',
u'Above & Beyond Mobile Detailing Services'],
u'Beauty & Spas': [u'Shaman Modifications Tattoo & Body Piercing',
u'Mantis Massage',
u'Beautiful Me',
u'L7 Salon',
u'Blush'],
u'Education': [u'The Stitch Lab',
u'The University of Texas',
u'Austin Moto Academy',
u'The New Movement',
u'Happy Time Defensive Driving'],
u'Event Planning & Services': [u'DJ Gatsby & Dapper Sounds',
u'Korey Howell Photography',
u'Spike Gillespie',
u'Kimber Modern',
u'Tiny Tails To You'],
u'Financial Services': [u'Kirk Ingels - State Farm Insurance Agent',
u'Brad Barta, CPA',
u'Padgett Business Services',
u'Benjamin-Fry Insurance',
u'Vista Lending'],
u'Food': [u'Sunrise Mini Mart',
u'Con\u2019 Olio Oils & Vinegars',
u'Antonelli\u2019s Cheese Shop',
u'Patika Coffee',
u'Texas Coffee Traders'],
u'Health & Medical': [u'Forest Family Dentistry',
u'Beautiful Me',
u'Austin Ashiatsu',
u'Tech Ridge Dental',
u'Eyes of Texas Laser Center'],
u'Home Services': [u'Einstein Moving Company',
u'Peace Frog Carpet & Tile Cleaning',
u'Jackson\u2019s Moving and Delivery',
u'Blue Dragon Plumbing',
u'Green Gorilla Movers'],
u'Hotels & Travel': [u'Austin Eats Food Tours',
u'Kimber Modern',
u'Live Love Paddle',
u'Cynthia\u2019s Manhattan Limousine',
u'PubCrawler of Austin'],
u'Local Flavor': [u'Yelp\u2019s Summer Stunner at @VulcanATX',
u'Yelp\u2019s Whodunnit @ W Austin',
u'Yelp\u2019s Pinup Party at Getaway Motor Club',
u'Yelp Finds Salvation at Salvation Pizza',
u'The Wild Parrots of East Austin'],
u'Local Services': [u'Peace Frog Carpet & Tile Cleaning',
u'1Up Repairs',
u'Expert Appliance Repair',
u'Hart\u2019s Carpet Cleaning',
u'Olga European Fashion'],
u'Mass Media': [u'KOOP Radio',
u'Bumperactive',
u'The Austin Chronicle',
u'The Onion -- America\u2019s Finest News Source',
u'KLRU - PBS Austin'],
u'Nightlife': [u'East End Wines',
u'Skylark Lounge',
u'Esther\u2019s Follies',
u'Santorini Cafe',
u'Patika Wine and Coffee'],
u'Pets': [u'Healthy Pet',
u'South Park Animal Hospital',
u'Lisa\u2019s Loving Pet Care',
u'Classi Diva Pet Spa',
u'Zoom Room Dog Training'],
u'Professional Services': [u'Creation Station Printing',
u'Alexander Law Firm',
u'Super Cheap Signs',
u'Shaolin-do Kung Fu and Tai Chi',
u'PostNet'],
u'Public Services & Government': [u'Texas State Capitol',
u'Harry Ransom Center',
u'The International Office',
u'Household Hazardous Waste Facility',
u'LBJ Presidential Library'],
u'Real Estate': [u'Realty Austin - Central',
u'Eric Bramlett Residential Real Estate',
u'Don Powers - Dwell Realty',
u'Paul Reddam',
u'Reinae Kessler - Austin Home Girls Realty'],
u'Religious Organizations': [u'St. David\u2019s Episcopal Church',
u'Christian Life Church',
u'Saint Mary Cathedral',
u'St Mary\u2019s Cathedral',
u'Vox Veniae'],
u'Restaurants': [u'Franklin Barbecue',
u'Hey!... You Gonna Eat or What?',
u'Beware Coffee',
u'Ng BMT',
u'Via 313 Pizza'],
u'Shopping': [u'Rocket Electrics',
u'Gene Rogers Optical',
u'Urban Mattress South Austin',
u'Urban Mattress',
u'Olga European Fashion']}
After that we can just stop the driver and stop the docker container.
driver.quit()
Scale¶
Ok, that looks pretty cool but how to scale this?
Fortunately with docker now you have a lot of options such as Kubernetes and Docker Swarm. In this case I decided to use Docker Swarm since I was lucky to test Rackspace new container service while I was at PyTexas this last weekend.
The process to start the container is exactly the same since docker-swarm serves the same docker API, you just need to point docker client to the docker-swarm cluster. Once thats done I can just execute the same docker commands to create multiple docker containers in the cluster.
$ docker pull danielfrg/selenium # Pull the container in all the nodes
$ docker run -d -p 4444 danielfrg/selenium # Multiple times to start multiple containers
$ docker run -d -p 4444 danielfrg/selenium
$ docker run -d -p 4444 danielfrg/selenium
$ docker run -d -p 4444 danielfrg/selenium
$ docker ps
CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES
34867f93db5a danielfrg/selenium "sh /opt/selenium/ent" 4 seconds ago Up 2 seconds xxx.xx.x.3:49154->4444/tcp 40291b37-aa07-43fd-a7de-1f61673a89a1-n3/sleepy_rosalind
b45d64827e21 danielfrg/selenium "sh /opt/selenium/ent" 6 seconds ago Up 4 seconds xxx.xx.x.2:49154->4444/tcp 40291b37-aa07-43fd-a7de-1f61673a89a1-n2/focused_brown
8a9da9801ee3 danielfrg/selenium "sh /opt/selenium/ent" 8 seconds ago Up 7 seconds xxx.xx.x.1:49154->4444/tcp 40291b37-aa07-43fd-a7de-1f61673a89a1-n1/kickass_euclid
59de7a4811ae danielfrg/selenium "sh /opt/selenium/ent" 13 seconds ago Up 11 seconds xxx.xx.x.3:49153->4444/tcp 40291b37-aa07-43fd-a7de-1f61673a89a1-n3/elegant_sammet
11b13963a2e0 danielfrg/selenium "sh /opt/selenium/ent" 15 seconds ago Up 13 seconds xxx.xx.x.2:49153->4444/tcp 40291b37-aa07-43fd-a7de-1f61673a89a1-n2/dreamy_curie
7beecb6a9e8a danielfrg/selenium "sh /opt/selenium/ent" 16 seconds ago Up 15 seconds xxx.xxx.x.1:49153->4444/tcp 40291b37-aa07-43fd-a7de-1f61673a89a1-n1/lonely_bell
Now that the containers are running we can use the Python docker API to get the IP and port of the containers.
import os
from docker import Client
from docker.utils import kwargs_from_env
kwargs = kwargs_from_env()
kwargs['tls'].assert_hostname = False
client = Client(**kwargs)
containers = client.containers()
seleniums = [c for c in containers if c['Image'] == 'danielfrg/selenium']
urls = [s['Ports'][0]['IP'] + ':' + str(s['Ports'][0]['PublicPort']) for s in seleniums]
urls
[u'104.130.0.130:49154', u'104.130.0.83:49154', u'104.130.0.80:49154', u'104.130.0.130:49153', u'104.130.0.83:49153', u'104.130.0.80:49153']
Now with this we have a pool of seleniums running that we can use to crawl not only one single page but a larger number of pages.
Final thoughts¶
This was a simple experiment for a more complete crawler that is based on newer technologies and that can target a larger number of websites. As I mention earlier data in the goverment website is very helpful in a lot of cases but is usually in some old system that requires clicks, search forms and stuff like that, this technique can be useful in those cases.
For a complete system maybe a better integration with Scrapy is needed and there are some experiements between scrapy + selenium: