plex URIs that use GET requests to pass on state^a, thus obscuring the identity of the actual resources.

URIs that carry state are used heavily in Web applications but are, to date, largely unanalyzed. For example, in a June 2007 talk, Udi Manber, Google’s VP of engineering, addressed the issue of why Web search is so difficult, ²⁵ explaining that on an average day, 20%–25% of the searches seen by Google have never been submitted before and that each of these searches generates a unique identifier (using server-specific encoding information). So a Web-graph model would represent only the requesting document (whether a user request or a request generated by, for example, a dynamic advertisement content request) linked to the www.google.com node. However if, as is widely reported, Google receives more than 100 million queries per day, and if 20% of them are unique, then more than 20 million links, represented as new URIs that encode the search term(s), should show up in the Web graph every day, or around 200 per second. Do these links follow the same power laws? Do the same growth models explain these behaviors? We simply don’t know.

Analyzing the Web solely as a graph also ignores many of its dynamics ( especially at short timescales). Many phenomena known to Web users (such as denial-of-service attacks caused by flooding a server and the need to click the same link multiple times before getting a response) cannot be explained by the Web-graph model and often can’t be expressed in terms amenable to such graph-based analysis. Representing them at the networking level, ignoring protocols and how they work, also misses key aspects of the Web, as well as a number of behaviors that emerge from the interactions of millions of requests hitting many thousands of servers every second. Web dynamics were analyzed more than a decade ago, ²⁰ but the combination of (i) the exponential growth in the amount of Web content, (ii) the change in the number, power, and diversity of Web servers and appli-

a. These characters, including ?.#, =, and &, followed by keywords, may follow the last “slash” in the URI, thus making for the long URIs often generated by dynamic content servers.

today’s interactive
applications are
very early social
machines, limited
by the fact that they
are largely isolated
one from another.

cations, and (iii) the increasing number of diverse users from everywhere in the world makes a similar analysis impossible today without creating and validating new models of the Web’s dynamics. Such models must also pay special attention to the details of the Web’s architecture, as well as to the complexity of the interactions actually taking place there.

Additionally, modern, sophisticated Web sites provide powerful user-interface functionality by running large script systems within the browser. These applications access the underlying remote data model through Web APIs. This application architecture allows users and entrepreneurs to quickly build many new forms of global systems using the processing power of users’ machines and the storage capacity of a mass of conventional Web servers. Like the basic Web, each such system is interesting mainly for its emergent macro-scale properties, of which we have little understanding. Are such systems stable? Are they fair? Do they effectively create a new form of currency? And if they do should it be regulated?

Similarly, many user-generated content sites now store personal information yet have rather simplistic systems to restrict access to a person’s “friends.” This information is not available to wide-scale analysis. Some other sites must be allowed to access the sites by posing as the user or as a friend; a number of three-party authentication protocols are being deployed to allow this. A complex system is thus being built piece by piece, with no invariants (such as “my employer will never see this picture”) assured for the user.

The purpose of this discussion is not to go into the detail of Web protocols or the relative merits of Web-modeling approaches but to stress that they are critical to the current and continued working of the Web. Understanding the protocols and issues is important to understanding the Web as a technical construct and to analyzing and modeling its dynamic nature. Our ability to engineer Web systems with desirable properties at scale requires that we understand these dynamics. This analysis and modeling are thus an important challenge to computer scientists if they are to be able to understand