Web Crawling and Metadata with Python Author:Andrew Montalenti Date: 2012-10-26 Web Crawling and Metadata with Python

Meta Information Me: I've been using Python for 10 years. I use Python full-time, and have for the last 3 years. Startup: I'm co-founder/CTO of Parse.ly ❏, a tech startup in the digital media space. E-mail me: [email protected] ❏ Follow me on Twitter: @amontalenti ❏ Connect on LinkedIn: http://linkedin.com/in/andrewmontalenti ❏

Parse.ly What do we do? How do we do it?

Meta Slide http://bit.ly/crawling-slides ❏

Crawler "A computer program that browses the World Wide Web in a methodical, automated manner or in an orderly fashion." Open source examples: • Apache Nutch: built by Doug Cutting, creator of Lucene/Hadoop • Heritrix: built by the Internet Archive

Web Data 40 billion pages on the Web today (Google) Growing: size was "just" 15 billion in October 2010 "Deep Web" means it's even bigger

Crawling, Spidering, Scraping These terms are almost synonymous, but sometimes have different meaning and connotations. My take: • Crawling: downloads and processes web content at one or more URLs • Spidering: walks links found in web content, either for a single domain or across the web • Scraping: uses knowledge of HTML pages to convert them into structured data

My first experience with crawlers Parse.ly Reader: personalized news reader built in mid 2009 Crawled 500K web sources for content personalized to individual interests First crawler was really dumb: based on RSS/Atom detection and feed fetching Seeded from domains appearing on top aggregators like Google News Technology: multiprocessing, Postgres, Solr

My current experience Parse.ly shifted into web publisher analytics and APIs in 2010/2011 Upgrades: Scrapy, MongoDB, Redis, Solr, Celery

Parse.ly Network Stats >200 top publishing domains (Quantcast top-10,000 sites) >3B pageviews/month across network >10M unique URLs in our index >1TB of hot production data running in memory

Parse.ly Crawl Infrastructure Have written 125 custom Scrapy crawlers with >10K of custom crawler code (Not proud of this fact; more on this later) Production environment in Rackspace Cloud; several worker nodes Implementation and QA runs in Scrapy Cloud Caching and retry strategy implemented atop Redis Eventual storage in MongoDB, Solr Implemented as Scrapy Components and Pipelines

Our Business We're not in the crawling business. We're in the analytics and APIs business.

Our Strategy We aim to be the #1 technology partner for large-scale publishers. Crawling: means to an end. URLs => Structured Metadata. Metadata Soup: Schema.org, rNews, OpenGraph, hNews, HTML5, ...

Parse.ly Demo Time! Yay!

Reflections on Scaling Crawlers You don't want to write your own crawler infrastructure from scratch. TRUST ME. Lots of hidden problems -- Abstractions: asynchronous network I/O (Twisted), data processing pipelines HTTP/web: retries, throttling, backoff, concurrency, cookie/form handling Infrastructure: crawling queues, health monitoring

Don't use Nutch, Heritrix Didier and I tried to understand, and even customize, Nutch in the early days. We love Lucene/Solr, so we figured it'd be a good fit. But no -- it's a WORLD OF PAIN. (They are for building search engines and archives -- not structured metadata.)

Use Scrapy It's really Pythonic. It's built on proven tools, like Twisted, w3lib, and lxml. It's getting better and better. Just trust me: use Scrapy.

Scrapy Overview $ git clone git://github.com/scrapy/dirbot.git $ cd dirbot $ mkvirtualenv dirbot $ pip install scrapy $ pip install ipython $ scrapy list dmoz $ scrapy crawl dmoz [scrapy] INFO: Scrapy 0.16.0 started (bot: dirbot) ...

Example Output [dmoz] DEBUG: Crawled (200) [dmoz] DEBUG: Crawled (200) [dmoz] DEBUG: Scraped from <200 http://dmoz.org/Comp.../Python/Resources/> Website: name=[u'Top'] url=[u'/'] [dmoz] DEBUG: Scraped from <200 http://dmoz.org/Comp.../Python/Resources/> Website: name=[u'Computers'] url=[u'/Computers/'] [dmoz] DEBUG: Scraped from <200 http://dmoz.org/Comp.../Python/Resources/> Website: name=[u'Programming'] url=[u'/Computers/Programming/'] ... [dmoz] DEBUG: Scraped from <200 http://dmoz.org/.../Python/Books/> Website: name=[u'Text Processing in Python'] url=[u'http://gnosis.cx/TPiP/'] [dmoz] INFO: Spider closed (finished) Links: • http://www.dmoz.org/Computers/Programming/Languages/Python/Resources/ ❏ • http://www.dmoz.org/Computers/Programming/Languages/Python/Books/ ❏

Spider Example class DmozSpider(BaseSpider): name = "dmoz" allowed_domains = ["dmoz.org"] start_urls = [ "http://www.dmoz.org/Computers/Programming/Languages/Python/Books/", "http://www.dmoz.org/Computers/Programming/Languages/Python/Resources/", ] def parse(self, response): hxs = HtmlXPathSelector(response) sites = hxs.select('//ul/li') items = [] for site in sites: item = Website() item['name'] = site.select('a/text()').extract() item['url'] = site.select('a/@href').extract() item['description'] = site.select('text()').extract() items.append(item) return items

Live Demos! Examples: DailyCaller.com, ArsTechnica.com • https://www.stypi.com/pixelmonkey/dailycaller.py ❏ • https://www.stypi.com/pixelmonkey/arstechnica.py ❏

Item from scrapy.item import Item, Field class DbItem(Item): title = Field() link = Field()

DailyCaller: imperative style from scrapy.spider import BaseSpider from scrapy.selector import HtmlXPathSelector from dirbot.items import DbItem class DailycallerSpider(BaseSpider): name = "dailycaller.com" allowed_domains = ["dailycaller.com"] start_urls = ["http://dailycaller.com"] def parse(self, response): hxs = HtmlXPathSelector(response) item = DbItem() item["title"] = hxs.select("//h1/text()").extract()[0] item["link"] = hxs.select("//link[@rel='canonical']/@href").extract()[0] return item

ArsTechnica: declarative style from scrapy.spider import BaseSpider from scrapy.selector import HtmlXPathSelector from dirbot.items import DbItem from scrapy.contrib.loader import XPathItemLoader from scrapy.contrib.loader.processor import TakeFirst class ArstechnicaSpider(BaseSpider): name = "arstechnica.com" allowed_domains = ["arstechnica.com"] start_urls = ["http://arstechnica.com"] def parse(self, response): loader = XPathItemLoader(item=DbItem(), response=response) loader.add_xpath("title", "//meta/[@property='og:title']/@content") loader.add_xpath("link", "//link[@rel='canonical']/@href") item = loader.load_item() item["title"] = loader.get_value(item["title"], TakeFirst(), unicode.title) item["link"] = loader.get_value(item["link"], TakeFirst()) return item

Live Spider Shell >>> fetch("http://dailycaller.com/2012...-most-rallies/") [dailycaller.com] INFO: Spider opened [dailycaller.com] DEBUG: Crawled (200) [s] Available Scrapy objects: [s] hxs [s] item Website: name=None url=None [s] request [s] response <200 http://dailycaller.com/...es/> [s] settings [s] spider [s] Useful shortcuts: [s] shelp() Shell help [s] fetch(req_or_url) Fetch request (or URL) and update local objects [s] view(response) View response in a browser >>> hxs.select("//title/text()")

Scrapy Cloud Demo How we host, test, and QA our spiders across millions of pages.

Schemato Overview Domo arigato, Mr. Schemato!

Schemato Distilling from distillers import Distill, Distiller class NewsDistiller(Distiller): title = Distill("s:headline", "og:title") image_url = Distill("s:associatedMedia.ImageObject/url", "og:image") pub_date = Distill("s:datePublished") author = Distill("s:author", "s:creator.Person/name") section = Distill("s:articleSection") description = Distill("s:description", "og:description") link = Distill("s:url", "og:url") site = Distill("og:site_id") id = Distill("s:identifier")

Schemato Distilling in Action >>> from distillery import NewsDistiller >>> from schemato import Schemato >>> lnk = "http://www.cnn.com/2012/10/26/world/europe/italy-berlusconi-convicted/index.html" >>> cnn = Schemato(lnk) >>> distiller = NewsDistiller(cnn) >>> distiller.distill() {'author': "Ben Wedeman", 'id': None, 'image_url': 'http://i2.cdn.turner.com/cnn/...-video-tease.jpg', 'link': 'http://www.cnn.com/2012/10/26/world/europe/italy-berlusconi-convicted/index.html', 'pub_date': '2012-10-26T14:36:35Z', 'section': 'world', 'title': 'Ex-Italian PM Berlusconi handed 4-year prison term for tax fraud', 'site': 'CNN', 'description': 'Flamboyant former Italian Prime Minister...'}

Schemato: Bridging Gaps Between Standards Facebook OpenGraph provided image_url and link. Schema.org NewsArticle provided the rest. >>> distiller.sources {'author': 's:author', 'id': None, 'image_url': 'og:image', 'link': 'og:url', 'pub_date': 's:datePublished', 'section': 's:articleSection', 'title': 's:headline', 'site': 'og:site_id', 'description': 's:description'}

Data sets to get started Are you interested in tackling some of these web crawling problems on your own project? If so, you may want some data to get started. I currently sell a few news data sets that help with this: • 30M news headlines and 500K web sources, 30gb of JSON data ($300) • 15K news domains that are the most popular in US market ($100) You could use either of these to build your own Google News, for example. Interested? Find me after or tweet me: @amontalenti ❏

Schemato: A Call to Action The time is ripe for the semantic web. Want to build the ultimate web metadata validator, distiller, and extractor? Want to work on getting Schemato to run across millions of URLs? Want your contributions open source on Github? Find me at the sprints on Sunday!

Baby Turtles Use your powers wisely, and always remember...

Magic Turtles! It's turtles all the way down!

Tweet and Meet What did you think? Tweet @amontalenti ❏ with #pydata hash tag! Rate this talk! http://bit.ly/rate-andrew ❏ Connect on LinkedIn: http://linkedin.com/in/andrewmontalenti ❏