|
|
|
|
|
|
|
|
Here's how spiders search the Web collecting information for you.
by Mike Thomas
Web-crawling robots, or spiders, have a certain mystique among Internet users. We all use search engines like Lycos and Infoseek to find resources on the Internet, and these engines use spiders to gather the information they present to us. Very few of us, however, actually use a spider program directly.
Spiders are network applications which traverse the Web, accumulating statistics about the content found. So how does a web spider work? The algorithm is straightforward:
Listing 1 is a program, spider.pl, which implements the above algorithm in Perl. This program should run on any Linux system with Perl version 4 or higher installed. Note that all code mentioned in this article assumes Perl is installed in /usr/bin/Perl. These scripts are available for download on my web page at http://www.javanet.com/~thomas/.
To run the spider at the shell prompt use the command:
spider.pl <starting-URL<search-phrase>The spider will commence the search. The starting URL must be fully specified, or it may not parse correctly. The spider searches the initial page and all its descendant pages for the given search phrase. The URL of any page with a match is printed. To print a list of URLs from the SSC site containing the phrase ``Linux Journal'', type:
spider.pl http://www.ssc.com/ "Linux Journal"The Perl variable $DEBUG, defined in the first few lines of spider.pl, is used to control the amount of output the spider produces. $DEBUG can range from 0 (matching URLs are printed) to 2 (status of the program and dumps of internal data structures are output).
The most interesting thing about the spider program is the fact that it is a network program. The subroutine get_http() encapsulates all the network programming required to implement a spider; it does the ``fetch'' alluded to in step 2 of the above algorithm. This subroutine opens a socket to a server and uses the HTTP protocol to retrieve a page. If the server has a port number appended to it, this port is used to establish the connection; otherwise, the well-known port 80 is used.
Once a connection to the remote machine has been established, get_http() sends a string such as:
GET /index.html HTTP/1.0This string is followed by two newline characters. This is a snippet of the Hypertext Transport Protocol (HTTP), the protocol on which the Web is based. This request asks the web server to which we are connected to send the contents of the file /index.html to us. get_http() then reads the socket until an end of file is encountered. Since HTTP is a connectionless protocol, this is the extent of the conversation. We submit a request, the web server sends a response and the connection is terminated.
The response from the web server consists of a header, as specified by the HTTP standard, and the HTML-tagged text making up the page. These two parts of the response are separated by a blank line. Running the spider at debug level 2 will display the HTTP headers for you as a page is fetched. The following is a typical response from a web server.
HTTP/1.0 200 OK Date: Tue, 11 Feb 1997 21:54:05 GMT Server: Apache/1.0.5 Content-type: text/html Content-length: 79 Last-modified: Fri, 22 Nov 1996 10:11:48 GMT <HTML><TITLE>My Web Page</TITLE> <BODY> This is my web page. </BODY> </HTML>The spider program checks the Content-type field in the HTTP header as it arrives. If the content is of any MIME type other than text/html or text/plain, the download is aborted. This avoids the time-consuming download of things like .Z and .tar.gz files, which we don't wish to search. While most sites use the FTP protocol to transfer this type of file, more and more sites are using HTTP.
There is a hardware dependency in get_http() that you should be aware of if you are running Linux on a SPARC or Alpha. When building the network addresses for the socket, the Perl pack() routine is used to encode integer data. The line:
$sockaddr="S n a4 x8";is suitable only for 32-bit CPUs. To get around this, see Mike Mull's article ``Perl and Sockets'' in LJ Issue 35.
Once the spider has downloaded the HTML source for a web page, we can scan it for text matching the search phrase and notify the user if we find a match.
We can also find any hypertext links embedded in the page and use them as a starting point for a further search. This is exactly what the spider program does; it scans the HTML content for anchor tags of the form <A HREF="url"> and adds any links it finds to its queue of URLs.
A hyperlink in an HTML page can be in one of several forms. Some of these must be combined with the URL of the page in which they're embedded to get a complete URL. This is done by the fqURL() function. It combines the URL of the current page and the URL of a hyperlink found in that page to produce a complete URL for the hyperlink.
For example, here are some links which might be found in a fictitious web page at http://www.ddd.com/clients/index.html, together with the resulting URL produced by fqURL().
| URL in Anchor Tag | Resulting URL |
| http://www.eee.org/index.html | http://www.eee.org/index.html |
| att.html | http://www.ddd.com/clients/att.html |
| /att.html | http://www.ddd.com/att.html |
As these examples show, the spider can handle both a fully-specified URL and a URL with only a document name. When only a document name is given, it can be either a fully qualified path or a relative path. In addition, the spider can handle URLs with port numbers embedded, e.g., http://www.ddd.com:1234/index.html.
One function not implemented in fqURL() is the stripping of back-references (../) from a URL. Ideally, the URL /test/.../index.html is translated to /index.html, and we know that both point to the same document.
Once we have a fully-specified URL for a hyperlink, we can add it to our queue of URLs to be scanned. One concern that crops up is how to limit our search to a given subset of the Internet. An unrestricted search would end up downloading a good portion of the world-wide Internet content--not something we want to do to our compadres with whom we share network bandwidth. The approach spider.pl takes is to discard any URL that does not have the same host name as the beginning URL; thus, the spider is limited to a single host. We could also extend the program to specify a set of legal hosts, allowing a small group of servers to be searched for content.
Another issue that arises when handling the links we've found is how to prevent the spider from going in circles. Circular hyperlinks are very common on the Web. For example, page A has a link to page B, and page B has a link back to page A. If we point our spider at page A, it finds the link to B and checks it out. On B it finds a link to A and checks it out. This loop continues indefinitely. The easiest way to avoid getting trapped in a loop is to keep track of where the spider has been and ensure that it doesn't return. Step 2 in the algorithm shown at the beginning of this article suggests that we ``pull a URL out of our queue'' and visit it. The spider program doesn't remove the URL from the queue. Instead, it marks that URL as having been scanned. If the spider later finds a hyperlink to this URL, it can ignore it, knowing it has already visited the page. Our URL queue holds both visited and unvisited URLs.
The set of pages the spider has visited will grow steadily, and the set of pages it has yet to visit can grow and and shrink quickly, depending on the number of hyperlinks found in each page. If a large site is to be traversed you may need to store the URL queue in a database, rather than in memory as we've done here. The associative array that holds the URL queue, %URLqueue, could easily be linked to a GDBM database with the Perl 4 functions dbmopen() and dbmclose() or Perl 5 functions tie() and untie().
Note that you should not unleash this beast on the Internet at large, not only because of the bandwidth it consumes, but also because of Internet conventions. The document request the spider sends is a one line GET request. To strictly follow the HTTP protocol, it should also include User-Agent and From fields, giving the remote server the opportunity to deny our request and/or collect statistics.
This program also ignores the ``robots.txt'' convention that is used by administrators to deny access to robots. The file /robots.txt should be checked before any further scanning of a host. This file indicates if scanning from a robot is welcome and declares any subdirectories that are off-limits. A robots.txt file that excludes scanning of only 2 directories looks like this:
Useuagent: * Disallow: /tmp/ Disallow: /cgi-bin/A file that prohibits all scanning on a particular web server looks like this:
User-agent: * Disallow: /Robots like our spider can place a heavy load on a web server, and we don't wish to use it on servers that have been declared off-limits to robots by their administrators
How might we use the spider program, other than as a curiosity? One use for the program would be as a replacement for one of the web site index and query programs like Harvest (http://harvest.cs.colorado.edu/Harvest/) or Excite for Web Servers (http://www.excite.com/navigate/prodinfo.html). These programs are large and complicated. They often provide the functionality of the Perl spider program, a means of archiving the text retrieved and a CGI query engine to run against the resulting database. Ongoing maintenance is required, since the query engine runs against the database rather than against the actual site content; therefore, the database must be regenerated whenever a change is made to the content of the site.
Some search engines, such as Excite for Web Servers, cannot index the content at a remote site. These engines build their database from the files which make up the web site, rather than from data retrieved across a network. If you had two web sites whose content was to appear in a single search application, these tools would not be appropriate. Furthermore, the Linux version of Excite for Web Servers is still in the ``coming soon'' stage.
Listing 2 and Listing 3 show a simple CGI search engine that is implemented using the spider.pl program. Listing 2 is an HTML form which calls spiderfind.cgi to process its input. Listing 3 is spiderfind.cgi. It first uses Brigitte Jellinek's library to move the data entered in the form into an associative array. It then calls the spider.pl program using the Perl system() function and passes the form data as parameters. Finally, it converts the output from spider.pl into a series of HTML links. The user's browser will display a list of hyperlinked URLs in which the search text was found. Note that the name of the host to search is specified by a hidden field in the HTML document. There are better and more security-conscious ways for two Perl programs to interact than through a Perl system() call, but I wanted to use an unmodified copy of spider.pl for this demonstration.
This script doesn't provide the complete functionality of the packages mentioned above, and it won't perform as well. Since we're doing the search against web server documents across the Net, we don't have the advantage of index files; therefore, the search will be slower and more processor-intensive. However, this script is easy to install and easier to maintain than those engines.
Another application that could be built using the spider.pl program is a broken link scanner for the Web. The HTTP response we showed previously began with the line ``HTTP/1.0 200 OK'', indicating the request could be fulfilled. If we tried to hit a URL with a non-existent document, we would get the line ``HTTP/1.0 404 Not found'' instead. We could use this as an indication that the document does not exist and print the URL which referenced this page.
The modifications to the spider program needed to accomplish this are minor. Every time a hyperlink's URL is added to the URL queue, we also record the URL of the document in which we found the hyperlink. Then, when the spider checks out the hyperlink and receives a ``404 Not found'' response, it outputs the URL of the referring page.
Mike Thomas is an Internet application developer working for a consulting firm in Saskatchewan, Canada. Mike lives in Massachusetts and uses two Linux systems to telecommute 2000 miles to his job and to Graduate School at the University of Regina. He can be reached by e-mail at thomas@javanet.com.
