Let's build a simple web crawler in Ruby. For inspiration, I'd like to to revisit Alan Skorkin's How to Write a Simple Web Crawler in Ruby and attempt to achieve something similar with a fresh perspective.

We'll adapt Skork's original goals and provide a few of our own:

  • must be able to crawl just a single domain
  • must be able to limit number of pages to crawl
  • the results should be represented as structured data so we don't have an incomprehensible soup of content
  • the results should be enumerable so we can have flexibility in how they're handled

Breaking it down

For this exercise, we're going to crawl Programmable Web to extract data from their API directory.

Rather than take the naive approach of grabbing all content from any page, we're going to build a webcrawler that emits structured data. Traversing from the first page of the api directory, our crawler will visit web pages like a nodes of a tree, collecting data and additional urls along the way.

Imagine that the results of our web crawl as a nested collection of hashes with meaningful key-value pairs.

# results
    name: "Google Maps",
    api_provider: "https://google.com"
    api_homepage: "https://developers.google.com/maps/"
    categories: ["Mapping", "Viewer"],
    provider_formats: ["JSON", "KML", "XML"]
    name: "Twitter",
    api_provider: "https://twitter.com"
    api_homepage: "https://dev.twitter.com/rest/public"
    categories: ["Social", "Blogging"],
    provider_formats: ["Atom", "JSON", "REST", "RSS", "XML"]

Designing the surface

If you've been following my posts lately, you know that I love Enumerable and you may not be surprised that I'd like to model our structured, website data with an Enumerator. This will provide a familiar, flexible interface that can be adapted for logging, storage, transformation, and a wide range of use cases.

I want to simply ask a spider object for its results and get back an enumerator:

=> #<Enumerator: ...>

We'll be able to do some interesting things, like stream the results lazily into a flexible storage engine, e.g. mongodb or PStore, available from the Ruby standard library:

require "pstore"
store  = PStore.new("api_directory.pstore")

# create `spider`, then ...

spider.results.lazy.take(50).each_with_index do |result, i|
  store.transaction do
    store[result[:name]] = result

Writing the crawler

We're going to write a Spider class to enumerate website data. Our spider implementation borrows heavily from joeyAghion's spidey gem, described as a "loose framework for crawling and scraping websites" and Python's venerable Scrapy project, which allows you to scrape websites "in a fast, simple, yet extensible way." Both resources achieve the goals of being easy-to-use and extensible.

We'll build our web crawler piece-by-piece, but if you want a full preview of the source, check out it on GitHub.

Our Spider will maintain a set of urls to visit, data is collects, and a set of url "handlers" that will describe how each page should be processed. We'll take advantage of one external dependency, mechanize, to handle interaction with the pages we visit - to extract data, resolve urls, follow redirects, etc. Below is the #enqueue method to add urls and their handlers to a running list in our spider.

require "mechanize" # as of this writing, the latest release is 2.7.4

class Spider
  def enqueue(url, method)
    url = agent.resolve(url).to_s
    return if @handlers[url]
    @urls << url
    @handlers[url] ||= { method: method, data: {} }


  def agent
    @agent ||= Mechanize.new

As we process each page we'll need a way to record the structured data we extract from various pages. We'll expose a #record method append a hash of data to the @results array.

class Spider
  def record(data = {})
    @results << data

Since our Spider will only know how to enumerate urls and record data, we'll introduce a collaborator object to contain the implementation for consuming data for a specific site. For now, we'll call this object a "processor". The processor will respond to the messages #root and #handler - the first url and handler method to enqueue for the spider, respectively. We'll also provide options for enforcing limits on the number of pages to crawl and the delay between each request.

class Spider
  MAX_URLS = 1000

  def initialize(processor, attrs = {})
    @processor = processor

    @urls     = []
    @results  = []
    @handlers = {}

    @interval = attrs.fetch(:interval, REQUEST_INTERVAL)
    @max_urls = attrs.fetch(:max_urls, MAX_URLS)

    enqueue(processor.root, processor.handler)

Enumerator Two Ways

Now for the real meat of our young Spider. The #results method is the key public interface: it enumerates the enqueued urls and yields members of the @results collection.

class Spider
  def results
    return enum_for(:results) unless block_given?

    index = @results.length
    enqueued_urls.each do |url, handler|

      # process url
      @processor.send(handler[:method], agent.get(url), handler[:data])

      if block_given? && @results.length > index
        yield @results.last
        index += 1

      # crawl delay
      sleep @interval if @interval > 0


  def enqueued_urls
    Enumerator.new do |y|
      index = 0
      while index < @urls.count && index <= @max_urls
        url = @urls[index]
        index += 1
        next unless url
        y.yield url, @handlers[url]

An interesting thing to note is that the size of our url queue and the collected results may be growing as we crawl more pages. For the #enqueued_urls private method, we're using an Enumerator to wrap the logic for iterating over the list of @urls and maintaining state, like the index. The Enumerator class is well-suited to represent a lazily generated collection.

def enqueued_urls
  Enumerator.new do |y|
    # ...

I find it to be a more expressive way to indicate we're enumerating values "on demand" as opposed to "eagerly", like a typical collection.

Notice we're also returning an enumerator from #results:

def results
  return enum_for(:results) unless block_given?
  # ...

This technique provides the method caller to more flexibility when determining how to handler the results. While you could pass a block to consume the results, e.g., spider.results { |r| puts r.inspect }, this is an eager operation. We'd have to wait for all the pages to be processed before continuing with the block. Returning an enumerator offers the potential to stream results to something like a data store.

Why not include Enumerable in our Spider and implement #each instead? As pointed out in Arkency's Stop including Enumerable, return Enumerator instead, our Spider class doesn't itself represent a collection, so exposing the #results method as an enumerator is more appropriate.

From Soup to Net Results

Our Spider is now functional so we can move onto the details of extracting data from an actual website.

Our processor, ProgrammableWeb will be responsible for wrappin a Spider instance and extracting data from the pages it visits. As mentioned previously, our processor will need to define a root url and initial handler method, for which defaults are provided, and delegate the #results method to a Spider instance:

class ProgrammableWeb
  attr_reader :root, :handler

  def initialize(root: "https://www.programmableweb.com/apis/directory", handler: :process_index, **options)
    @root = root
    @handler = handler
    @options = options

  def results(&block)


  def spider
    @spider ||= Spider.new(self, @options)

ProgrammableWeb will define handler methods that deserialize a web page into additional urls and data to add to our collection of results. Our spider will invoke the handlers (as seen above with @processor.send(method, agent.get(url), data)). Each handler method will have the following signature

def handler_method(page, data = {})
  # enqueue urls and/or record data

... where page is an instance of Mechanize::Page (docs) providing a number of methods for interacting with html content:

The root handler method, ProgrammableWeb#process_index, will extract api names in the index list, enqueue api detail pages and additional, paginated indexes. As data is collected, it may be passed on to handlers further down the tree via Spider#enqueue.

def process_index(page, data = {})
  page.links_with(href: %r{\?page=\d+}).each do |link|
    spider.enqueue(link.href, :process_index)

  page.links_with(href: %r{/api/\w+$}).each do |link|
    spider.enqueue(link.href, :process_api, name: link.text)

To process api detail pages, we'll define a separate handler. Since these pages will represent "leaves" in this exercise, we'll merge the data passed in with that extracted from the page and pass it along to Spider#record.

def process_api(page, data = {})
  fields = page.search("#tabs-content .field").each_with_object({}) do |tag, o|
    key = tag.search("label").text.strip.downcase.gsub(%r{[^\w]+}, ' ').gsub(%r{\s+}, "_").to_sym
    val = tag.search("span").text
    o[key] = val

  categories = page.search("article.node-api .tags").first.text.strip.split(/\s+/)

  spider.record data.merge(fields).merge(categories: categories)

As we saw earlier, recorded data is emitted in the Spider#results method.

Now we can make use of our ProgrammableWeb crawler as intended with simple instantiation and the ability to enumerate results as a stream of data:

spider = ProgrammableWeb.new

spider.results.lazy.take(5).each_with_index do |result, i|
  puts "%-3s: %s" % [i, result.inspect]

# 0 : {:name=>"Facebook", :api_provider=>"http://facebook.com", :api_endpoint=>"http://api.facebook.com/restserver.php", :api_homepage=>"https://developers.facebook.com/", :primary_category=>"Social", :secondary_categories=>"Webhooks", :protocol_formats=>"JSON, REST", :ssl_support=>"Yes", :api_kits=>"http://developers.facebook.com/documentation.php?doc=clients", :api_forum=>"http://forum.developers.facebook.com/", :twitter_url=>"http://twitter.com/fbplatform", :developer_support=>"http://developers.facebook.com/group.php?gid=2205007948", :console_url=>"http://developers.facebook.com/tools/explorer", :authentication_mode=>"API Key, OAuth 2, Username/password", :categories=>["Social", "Webhooks"]}
# 1 : {:name=>"LinkedIn", :api_provider=>"http://www.linkedin.com/", :api_endpoint=>"http://api.linkedin.com/v1/", :api_homepage=>"https://developer.linkedin.com/docs", :primary_category=>"Social", :secondary_categories=>"Enterprise", :protocol_formats=>"JSON, JSONP, REST, XML", :other_options=>"JavaScript", :ssl_support=>"Yes", :api_forum=>"https://developer.linkedin.com/forum", :twitter_url=>"https://twitter.com/linkedindev", :console_url=>"http://developer.linkedinlabs.com/jsapi-console/#examples/login/simple.html&{&quot;framework&quot;:&quot;platform.linkedin.com/in.js&quot;,&quot;frameworkurl&quot;:&quot;&quot;,&quot;apikey&quot;:&quo", :authentication_mode=>"OAuth 2", :categories=>["Social", "Enterprise"]}
# 2 : {:name=>"Skype", :api_provider=>"http://skype.com", :api_endpoint=>"http://skype.com", :api_homepage=>"http://www.skype.com/en/developer/", :primary_category=>"Telephony", :secondary_categories=>"Chat, Messaging, Video, Voice", :protocol_formats=>"Unspecified", :other_options=>"Skype proprietary", :ssl_support=>"Yes", :api_kits=>"https://developer.skype.com/Docs/Web https://developer.skype.com/wiki/Java_API", :api_forum=>"http://forum.skype.com/index.php?showforum=16", :developer_support=>"http://forum.skype.com/index.php?showforum=16", :authentication_mode=>"Unspecified", :categories=>["Telephony", "Chat,", "Messaging,", "Video,", "Voice"]}
# 3 : {:name=>"Twitter", :api_provider=>"http://twitter.com", :api_endpoint=>"http://twitter.com/statuses/", :api_homepage=>"https://dev.twitter.com/rest/public", :primary_category=>"Social", :secondary_categories=>"Blogging", :protocol_formats=>"Atom, JSON, REST, RSS, XML", :ssl_support=>"No", :api_kits=>"ActionScript", :api_forum=>"http://groups.google.com/group/twitter-development-talk", :twitter_url=>"http://twitter.com/twitterapi", :contact_email=>"[email protected]", :console_url=>"https://dev.twitter.com/console", :authentication_mode=>"OAuth 2, HTTP Basic Auth, OAuth", :categories=>["Social", "Blogging"]}
# 4 : {:name=>"YouTube", :api_provider=>"http://www.google.com", :api_endpoint=>"http://gdata.youtube.com/feeds/", :api_homepage=>"https://developers.google.com/youtube/", :primary_category=>"Video", :secondary_categories=>"Media", :protocol_formats=>"Atom, RSS, JSON, XML, GData", :other_options=>"Atom Publishing Protocol (Atom/RSS)", :ssl_support=>"No", :api_kits=>"Java, PHP Python, Ruby, ActionScript", :api_forum=>"http://groups.google.com/group/youtube-api/", :twitter_url=>"https://twitter.com/YouTubeDev/", :developer_support=>"http://code.google.com/support/bin/topic.py?topic=12357", :console_url=>"http://code.google.com/apis/ajax/playground/?exp=youtube#simple_embed", :authentication_mode=>"OAuth2", :categories=>["Video", "Media"]}

Wrapping up

I admit, it's arguable that this implementation is "simple". Skorks provided a straightforward, recursive solution to consume unstructured content. Our approach is iterative and requires some work up front to define which links to consume and how to process them with "handlers". However, we were able to achieve an extensible, flexible tool with a nice separation of concerns and a familiar, enumerable interface.

Modeling results from a multi-level page crawl as a collection may not work for every use case, but, for this exercise, it serves as a nice abstraction. It would now be trivial to take our Spider class and implement a new processor for a site like rubygems.org or craigslist and stream the results to a database like Redis or YAML::Store.

Try it yourself and let me know what you think of this approach (full source).

Did you like this post? Do me a favor: share it on Twitter, follow me - @rossta, and sign up for my newsletter. Thanks!

Part of the Enumerable series. Published on Jan 27, 2016