How dynamic is the web?^{1}
Brian E. Brewington and George Cybenko
Thayer School of Engineering
Dartmouth College
Hanover, New Hampshire 037558000
brian.e.brewington@Dartmouth.edu
george.cybenko@Dartmouth.edu
Abstract:
Recent experiments and analysis suggest that there are about 800 million publiclyindexable web pages. However, unlike books in a traditional library, web pages continue to change even after they are initially published by their authors and indexed by search engines. This paper describes preliminary data on and statistical analysis of the frequency and nature of web page modifications. Using empirical models and a novel analytic metric of "uptodateness", we estimate the rate at which web search engines must reindex the web to remain current.
Keywords:web dynamics, monitoring, document management
1. Introduction
Since its inception scarcely a decade ago, the World Wide Web has become a popular vehicle for disseminating scientific, commercial and personal information. The web consists of individual pages linked to and from other pages through Hyper Text Markup Language (HTML) constructs. The web is patently decentralized. Web pages are created, maintained and modified at random times by thousands, perhaps millions, of users around the world.
Search engines are an index of the web, playing the role of traditional library catalogs. However, a book or magazine does not change once it is published, whereas web pages typically do. Therefore, web search engines must occasionally revisit pages and reindex them to stay current. This is a constant challenge considering that recent empirical studies by Lawrence and Giles [LG99] have estimated the size of the publiclyindexable web to be at least 800 million pages (and climbing). The size of the web is only one factor in the reindexing problem; the rate at which pages change is equally important.
This paper starts with a description of our observational data on the rates of change for a large sample of web pages. Based on this data, we develop an exponential probabilistic model for the times between individual web page changes. We further develop a model for the distribution of the change rates defining those exponential distributions. These two estimates can be combined to answer questions about how fast a search engine must reindex the web to remain "current" with respect to a novel definition of currency. We introduce the concept of currency which defines our notion of being uptodate by using a probability, , that a search engine is current, relative to a grace period, , for a randomly selected web page.
Our observational data is based on statistics gathered from over two million web pages specified by over 25,000 users of a web clipping service [Inf95]. We have observed pages at a rate of about 100,000 pages per day, for a period of over seven months, recording how and when these pages have changed. The data indicate that the time between modifications of a typical web page can be modeled by an exponential distribution, which is parameterized by the rate of changes for the page. Our data further indicate that the reciprocal of that parameter, which is the expected time between changes, is wellmodeled by a Weibull distribution across pages.
As a measure of how uptodate a search engine is, we develop the precise concept of ()currency of a search engine with respect to a changing collection of web pages. Loosely speaking, the search engine data for a given web page is said to be current if the page has not changed between the last time it was indexed and time units ago. In this context, is the "grace period" for allowing unobserved changes to a web page. A search engine for a collection of pages is then said to be ()current if a randomly (according to some specified probability distribution) chosen page in the collection has a search engine entry that is current with probability at least .
To get an intuitive feeling for this concept, we might say that a daily newspaper is (0.90, 1 day)current when it is printed, meaning that the newspaper has at least 0.9 probability of containing 1 day current information on topics of interest to its readers (this reader interest is the specified probability distribution). Here 1 day current means that events that have happened within the last day, namely the grace period, are not expected to be reported and we "forgive" the newspaper for not reporting them. Similarly, hourly television news would be (0.95, 1 hour)current and so on. The idea is that we are willing to "forgive" an index or source if it is not completely uptodate with respect to the grace period, but we have a high expectation that it is uptodate with respect to that time.
Our empirical analysis of web page changes is combined with existing estimates of the web's size to estimate how many pages a search engine must reindex daily to maintain currency of the entire indexable web. Using 800 million documents [LG99] as the size of the web, we show that a (0.95, 1 week)current search engine must download and index at least 45 million pages a day, which would require a bandwidth of around 50 megabits/second (using an average page size of approximately 12 kilobytes and assuming uniform processing). A (0.95, 1 day)current search engine must reindex at the rate of at least 94 million pages daily, or 104 megabits/second. Our results allow estimation of reindexing rates in order to maintain general ()currency of a web index. Previous work on web page change rates has addressed the effect changing pages have on cache consistency [DFKM97]. The metrics used there focus on the effect of dynamics on web caching, rather than on the web page change dynamics themselves. For example, [DFKM97] uses a web page "change ratio," defined as the number of accesses to a changed page divided by the total number of accesses.
Our work also concerns the performance of a search engine in maintaining a web index. In [CLW97], a formal proof is given for the optimal sample period for monitoring a collection of pages that change memorylessly, under certain sampling conditions. Optimality is measured by a sum of total time outofdate for pages in the index, where each term is weighted by expected time between page changes. Our measures are similar in spirit, but introduce a temporal and probabilistic relaxation of what it means to be uptodate, namely the concept of currency.
2. Collecting web page change data
Since early 1996, we have maintained a web clipping service called "The Informant"^{2} that downloads and processes on the order of 100,000 web pages daily. The service monitors specific URLs for changes, and also runs standing user queries against one of four search engines^{3} at specified intervals. Any of three events trigger a notification of a user by email. The user is notified by email if (1) a monitored URL changes, (2) new results appear in the top results returned by a search engine in response to a standing query, or (3) any of the current top search results shows a change. A change, for our purposes, is any alteration of the web page, no matter how minor.
Beginning in March 1999, we started archiving HTML page summary information for all downloads. As of this writing, this has involved the download and processing of over 200 gigabytes of HTML data. The archived information includes the lastmodified time stamp (if given), the time of observation (using the remote server's time stamp if possible), and stylistic information (number of images, tables, links and similar data). The Informant selects and monitors web pages in a very specific way, so conclusions from the data must be interpreted only after knowing our sampling methods.
Since the Informant makes repeated observations of only those pages ranked high by search engines, this biases against those pages which are not relevant to our users' standing queries. Our sample is also biased towards the individual userselected URLs which have been deemed worth monitoring. While neither of these is crippling, they do color our results by being slanted towards those pages that our users wish to monitor. We do not claim that this bias is a popularity bias, since our users' queries are not necessarily the same as those which are of general interest.
Another important consideration is the sample rate. Standing queries are run no more often than once every three days for any single user, and some users' queries are run once every seven days or more. Therefore, the only way a page is observed more than once every three days is if it is needed by a different user on each of those days. A number of popular sites (news sites, shareware distributors, proficient "keyword spammers") fall into this category. Moreover, to keep our service from annoying providers of popular content, we cache pages (and delete the cache prior to gathering each day's results), so no more than one observation is made of a single page per day. In addition, since we run our queries periodically and only at night, sample times for any given page are correlated.
Many monitored sites exhibit a partial overlap between users, resulting in observations being made at irregular intervals. For extremely fastchanging pages, it is quite possible that many changes will occur between observations, making direct observation of all such changes impossible. When LASTMODIFIED information is given in the HTTP header, we can work around this by estimating change rates from ages. This will be discussed in greater detail in later sections.
While LASTMODIFIED information is available for around 65% of our observations, the absence of such information does seem to indicate a more volatile resource. Specifically, not having this timestamp makes an observation of any given resource about twice as likely to show a modification. Therefore, estimates of change rates based solely on pages that provide a timestamp are lower bounds (slowest estimate).
Figure 1: [Histogram: Lastmodified times (GMT), mod 24 • 7 hours.] Peaks in modification frequency are clearly visible during US working hours, and diminish on weekends. Assumpitons of stationarity in page alteration probability will break down at this scale.
Timestamps also show, indirectly, that most webpages are modified during the span of US working hours (between around 8 AM and 8 PM, Eastern time). This is shown in Figure 1. This is where any assumption of stationarity in change probability will break down; modifications are less likely during the low times on this plot.
Figure 2: [Stylistic clues to webpage age.] On the left, we show two distributions of contentlength, or the number of bytes in a webpage. One is for pages dated between 6/94 and 6/95, and the other is for pages last modified between 6/98 and 6/99. Widespread use of spaceintensive scripting languages and stylistic elements <FONT> tags, precise table and image sizing, and so forth) has driven the content length upwards. On the right, a similar trend is seen in the number of images, often used in more recently modified pages to make a more visually appealing presentation. Much of this reflects the shift from an academiccentric web to a commercialcentric one.
Not surprisingly, there is a correlation between the style of a webpage and its age. For example, in Figure 2, we show how the distribution of contentlengths and number of images depends upon age. Each plot shows two distributions, one using data from pages last modified between 6/94 and 6/95, and the other using pages between 6/98 and 6/99, to show how newer pages are frequently longer and have more images. Both distributions in the figure argue for the importance of spacesaving technology (such as compression techniques written into the HTTP1.1 standard, cascading style sheets (CSS), and use of Extended Markup Language (XML) where appropriate). Similar trends, sometimes much more pronounced, are seen in the usage of secondgeneration tags, such as the <TABLE> and <FORM> tags. While it might be feasible to use stylistic cues to estimate ages for pages which do not provide a timestamp, a far better solution is for content providers to include one along with an estimated expiration time. This potentially has many benefits, including better cache performance and fewer wasted observations by search engines (if honesty in expiration estimation is enforced).
Figure 3: [Cumulative distribution of change ratios.] The "change ratio" for a page is defined as the number of changes observed divided by the number of repeat accesses. We have plotted the cumulative distribution of this statistic for pages which have been observed six times or more. This shows that no more than 4% of these pages are totally dynamic, while we have never observed any sort of change for 56% of pages. These values are very dependent upon the sampling scheme and are therefore not comprable to numbers taken from web cachebased studies.
A popular question regarding our data is, "What about dynamicallygenerated pages?" We can determine an upper bound on what percentage of pages are dynamic by looking at how many pages change on every repeat observation. Following [DFKM97 ], we can plot a cumulative distribution function of "change ratios" as in Figure 3. As mentioned in the introduction, a change ratio is defined by the number of changes observed, divided by the number of repeat accesses made. Obviously, this statistic depends heavily upon the sample rate, but it does give a feeling for the distribution of change rates. We have plotted change ratios corresponding to pages which had been observed six times or more. A unit ratio indicates a resource that always changes faster than the sample rate, meaning it may be totally dynamic, although it may just change very quickly. The plot shows that 4% of pages changed on every repeat observation (70% of these pages did not give a timestamp), while no change was observed for 56% of pages. The average page is observed 12 times over an average of 37 days, so this portion of pages that did not change would be much smaller if the monitoring was over a longer timespan.
The difference between a downloaded page's lastmodified timestamp and the time at downloading is defined as the page's age. Recording the ages of the pages in the Informant database allows us to make several inferences about how those ages are distributed.
Estimates of the cumulative distribution function (CDF) and the probability density function (PDF) of page age are shown in Figure 4. A few observations about these plots give insight into the distribution of document ages. About one page in five is younger than eleven days. The median age is around 100 days, so about half of the web's content is younger than three months. The older half has a very long tail: about one page in four is older than one year and sometimes much older than that. In a few rare cases, server clocks are set incorrectly, making the timestamp inaccurate. The oldest pages that appear to have correct timestamps are from around 1992, some of which are "archaeologically" interesting^{4}. Our data on page age is similar to that found in an earlier study [DFKM97]; when the histograms in Figure 4 are altered so that the bins have the same size as in [DFKM97], our distribution matches their data for "infrequentlyaccessed" HTML pages.
Typical age observations are shown in Figures 5 and 6. Since pages are only observed for as long as they remain in any user's search results, many single pages are only monitored for a limited time. As such, no alterations are ever observed on about 56% of the pages we have monitored^{5}. This type of behavior is often appears like the examples shown in Figure 5. When web pages are more dynamic, their age samples look more like the examples in Figure 6, where the pages have progressed through many changes and we have observed the ages over that time span. This usually produces distributions close to an exponential PDF. Some rapidly changing pages appear to be periodic, though the period is rarely larger than one day. Periodicity can be inferred from age distributions that appear to be approximately uniform. Still other pages are entirely dynamic, generated anew with each access, but these are not more than 4% of our collection.
Figure 4: [Estimated distribution of web page ages.] Here we show estimates of the probability density function (PDF) and cumulative distribution function (CDF) of web page age. On the left, we estimate the PDF using a rescaled histogram of web page ages, using only one age observation per page. On the right, the corresponding CDF is formed by integrating the estimate of the PDF.
Figure 5: [Example age observations for relatively static pages.] Many of the pages we monitor do not change during the time they are observed, like the examples shown here. The upper plots are histograms, and the lower plots show the raw data. These examples show that many of the pages are quite old, and for some of them, the only change they will ever experience is their eventual disappearance.
Figure 6: [Example age observations for changing web pages.] For some of our pages, we have observed a number of changes over a long timespan. The distribution of ages over this time is often approximately exponential, as can be seen in the histograms. The raw data is shown in the lower plots.
3. Modeling the changes in a single page
To make further analysis possible, we model the changes in a single web page as a renewal process [Pap84]. A good example and analogy is a system of replacement parts. Imagine a light fixture into which we place a lightbulb. Whenever that bulb burns out, it is replaced immediately. We speak of the time between lightbulb failures as the "lifetime" of a bulb. At a specific instant, we define the time since the present lifetime began to be the "age" of the bulb. The analogy to web page changes is that a page's lifetime is the time between changes (where change is arbitrarily but unambiguously defined). The age is the time between a given instant and the most recent change prior to that instant. We diagram these concepts in Figure 7.
Figure 7: [Lifetimes vs. ages.] A single lifetime is represented by the time separating changes, which are denoted by $\times$'s in the graph. For each lifetime, the age (shown as a dashed line) increases linearly from 0 to the lifetime, then resets to 0 as the next lifetime begins.
In this initial study, we assume that individual lifetimes are independent and identically distributed, and that the lifetime distribution of a particular page does not change over time (the distribution is stationary). Not surprisingly, the
lifetime probability density, f (t), is closely related to the age probability density, g (t).
The act of observing "the age is t units" is the same as knowing "the lifetime is no smaller than t units." Intuitively, this indicates that the PDF g (t) should be proportional to the probability 1F (t) of a given lifetime exceeding t units, where F(t) is the CDF corresponding to f (t). To make g(t) a proper probability distribution, the constant of proportionality is chosen so that g (t) is normalized. This intuition proves correct and formal methods [Pap84] show that
Figure 8: [Relationship between lifetime and age distributions.] On the left, we show three hypothetical lifetime distributions; a Gaussian (dotted), Weibull (solid), and an exponential (dashed). On the right, we show the corresponding age distributions. For the Gaussian, the age distribution is a renormalized and shifted complementary error function (erfc). For the exponential, the age and lifetime distributions are identical. The age distribution for the Weibull has a more general shape. Note that periodic lifetimes imply uniform age distributions.
Some examples of this relationship are shown in Figure 8.
Establishing the relationship of age to lifetime is useful, since it is difficult to sample the distribution f (t) directly. Rather, it can be easier to estimate change rates using samples from the age distribution g (t) and then use (1 ) to estimate F(t) and then f(t). Aliasing of f (t) may happen when a page change is observed, since an observer can only conclude that one or more changes have occurred since the previous observation. In observing ages, there is no such difficulty. Avoiding the aliasing problem is not magic; we are merely making proper use of the fact that the filesystems on which the pages reside have sampled much faster than we can. Clearly, observation of a web page age requires the availability of the LASTMODIFIED information, which restricts our analysis to a smaller sample.
The simplest possible page lifetime model, and a good one to use for this initial investigation, is one in which pages change memorylessly. Intuitively, this means that the probability of a page being altered in some short time interval is independent of how much time has elapsed since the last change was made. This is a common model used in queuing systems and statistical reliability theory [Pap84]. For such pages, f (t) is an exponential distribution with parameter .
This distribution is a good choice, since much of our data on page changes show behavior like that shown in Figure 6. As for the more slowlychanging content, like the examples shown in Figure 5, it is certainly possible that these pages are not at all dynamic or that they change at a very low rate. We proceed with the assumption that all pages are dynamic, even if the only change they will ever experience is their disappearance. For these longer lifetimes, the best we can do is to obtain several (dependent) samples of the age distribution. Pages for which f (t) is an exponential distribution also have exponentially distributed ages g (t), since
1F_{c}(t)  =  
= 
implies
=  
=  (2) 
This means we can estimate a page's lifetime PDF, assuming an exponential distribution, using only page age observations which we easily obtain from the data.
4. Dealing with a growing web
It is clear from the empirical page age distribution shown in Figure 4 that the majority of web pages are young. What is less clear is why. Different explanations can give rise to the same observed age distribution. One the one hand, a fixed population of pages whose change times are governed by identical exponential PDF's will produce an exponential age distribution when sampled collectively, as in (2). At the other extreme, an exponentially growing population of web pages in which changes are rare or even nonexistent will be skewed towards youth as well  there will be exponentially more pages in one generation relative to the previous generation.
The middle ground is an exponentially growing web in which each page changes at time intervals determined by an exponential. Such a model will also yield an exponential distribution of page ages when sampled.
Consider two very different models for the web. First, an exponentiallygrowing population of completely static web pages will produce an exponential distribution of observed page ages. To see this, note that the population at time t is given by an expression of the form where P_{0} is the initial population and is the exponential growth rate parameter. An age distribution at time can be formed by reversing the sense of time, and normalizing by the population size:
(3) 
This distribution will approach an exponential density with parameter as gets large. The HTML version of this paper includes an animation of how an exponential age distribution evolves in time, when conditioned upon an initial age observation. We also animate this evolution for nonexponential age distributions.
But an exponential distribution of page ages can arise for completely different reasons. Consider a fixedsize group of identical pages, each of which changes at time intervals governed by an exponential distribution. Each page undergoes many changes, with each change returning that page to age zero. Such a population also gives rise to essentially an exponential age distribution (see 2). In particular, the age distribution for such a population is
(4) 
As the time since the population's birth, , becomes large, the distribution of observed page ages will also approach an exponential distribution and will be hard to distinguish from that of a growing population of unchanging web pages. The hybrid model we use in this paper represents the middle ground  the web is growing and pages change according to exponential time distributions. These are reasonable working assumptions.
We now combine the effects of web growth and page change dynamics. The web has been growing for several years so that the time since creation of web pages is distributed approximately exponentially:
(5) 
where is the growth rate and t_{c} is the time since creation of a page. We emphasize that t _{c } is not to be confused with our definition of the page's age, since age refers to the time since the last modification.
For an exponentiallygrowing population of dynamic pages, each of which has an exponential age distribution as described by (4), the aggregate age distribution will be a weighted average over time since creation, weighted by the number of pages created at the same time. Specifically,
=  (6)  
=  (7)  
=  
=  
=  (8) 
This means that the age distribution of an exponentially growing population of objects with (identical) exponential age distributions remains exponential, with parameter given by the sum of the population growth and page change rate constants.
The age distribution for the entire population (namely the whole web) is yet another mixture, in which we take expectation of (8) with respect to a joint distribution of growth rate and change rate . For simplicity we use the same growth rate for all change rates. Using a distribution over the inverse rate , with this uniform growth rate , we express the mixture as
(9) 
The only factor remaining before this distribution can be matched to the data is the shape of the distribution w (x) of inverse change rates. In our initial development, we use a generalized exponential (Weibull) distribution over the inverse change rate (which is also the mean change time), such that
(10) 
where is a scale parameter and is a shape parameter. See [MR94] for a discussion of Weibull distributions, as well as a more general discussion of this family of exponential distributions in [Fel71]. The shape parameter can be varied to change the shape from a very sharplypeaked distribution (for ) to an exponential (for ), to a unimodal distribution with maximum at some positive t (for ). The scale parameter adjusts the mean of the distribution.
To determine what values of , , and best model the observations, we numerically evaluate (9) at a number of ages t. This is used to estimate the cumulative age distribution G (t) at N points t_{i} . These estimates, , are compared with samples from the empirical distribution G (t) (as diagrammed in the left half of Figure 4) at points t _{i }. A sum of the squared error over all sample times t _{i} provides a scalar error function of the vector . This error function can be minimized:
(11) 
Figure 9: [Bestfit age CDF.] These plots show the distribution which results from a numerical optimization of (11). The top plot uses a log scale to show the deviations in the fit for small age. The minimization was carried out using linearlyspaced points.
When this minimization is carried out numerically, the optimal values are found to be ,, and . The fitted age distribution is shown in Figure 9. These parameters imply a steeperthanexponential age distribution (since ) and a growth rate that implies a doubling time of around 390 days. This is not unreasonable, as [LG98] estimated a lower bound size of 320 million pages in December 1997, which increased in [LG99] to 800 million pages by February 1999. This would imply a growth constant over the 14 months of , or a doubling time of 318 days. The difference in these estimates tells us to proceed with caution, understanding that estimates based on these results are somewhat uncertain. Moreover, the assumption of exponential growth in the number of documents is based on assertions of exponential growth in the number of web hosts (as in [Gra97 ] and [ISC99], for example). Growth rates have slowed appreciably, especially in the last year; other estimation methods prove more reliable.
5. Estimating the change rate distribution using lifetimes
As mentioned previously, inferring change rates from observed lifetimes is somewhat tricky, since an observed change may only be the most recent of many changes that took place since the last observation. Moreover, changes that take a long time to happen are inherently more difficult to catch. For example, if one were to watch a calendar for three consecutive days, waiting for the month to change, there is a good chance that this event will not be observed. However, as the timespan gets longer it becomes more probable that a change will be seen. In the same way, it is necessary to account for the probability of observing a change, given the timespan of observation.
For a page which changes exponentially at rate , the probability that at least one change will be observed within a timespan is
(12) 
The pages in our collection are observed over many different timespans . Therefore, to determine the probability of observing changes for pages having change rate , we assume that change rate and timespan are independent and weight (12) with respect to the probability of all possible observation timespans (discretized):
(13) 
Figure 10: [Observation time distribution and induced finite time span bias.] The top plot shows the distribution of observation time spans, or the time difference between the first and last observation timestamps for individual pages. The spikes appear in this graph because we only run our checks at night, so timespans tend to cluster around 24hour intervals. These timespans translate into the probability of any mean change time being represented among our observed web page changes.
Possible timespans are distributed as shown in Figure 10. Combining this data with (13) allows us to compute Z _{bias} weighting each mean lifetime's probability of being among the observed data. The distribution of change rates sampled in our experiment is not the true rate distribution, but rather one that is weighted by (13). If the actual density of mean lifetimes is f_{mean}(t), then the observeddensity of mean lifetimes is
(14) 
These mean lifetimes are only seen through a mixture of exponential distributions, so the observed lifetimes should approximate the probability density
(15) 
As with the agebased estimates, we can form a mean squarederror function like (11) and fit the CDF corresponding to (15) to the observed lifetime distribution. We show the distribution of observed lifetimes in Figure 11.
Figure 11: [PDF and CDF of observed lifetimes.] On the left, a rescaled histogram approximates the PDF of observed lifetimes, or differences in successive modification timestamps. On the right, we show the corresponding CDF.
Using F(t) as the cumulative lifetime distribution, and as the estimator, the error function is
(16) 
Figure 12: [Intensity plot of mean squarederror.] The minimum of the error function in the space of shape parameters and scale parameters is marked by a white ``X'' in the center of the dark patch. The error function appears to be unimodal.
Figure 13: [Overlay of model lifetime CDF and observed CDF.] The minimum error distribution (marked ``Trial'' above) found by minimizing the fiterror function is shown along with the observed lifetime distribution (marked ``Reference''). Errors in the fit below around 8 days are due to aliasing, where multiple changes are masked and treated as a single, larger lifetime. Our estimates are only extrapolations in this region and may be inaccurate. The region above 8 days is an extremely precise fit; the two curves are nearly identical.
Figure 14: [Absolute error $\hat{F}F$ vs. lifetime.] These are the errors in the fit to the lifetime CDF; we have used a linear scale and just show the leftmost region. Note the large errors below around 8 days due to aliasing. The effect of diurnal and weekly trends, as plotted in Figure \ref{fig:lastmodmod}, is clearly visible in the long and short period ripples above 8 days. Slight improvements in our estimates could be had if we restricted the fit to samples above 8 days.
As before, we use a Weibull density (10) for the distribution of inverse rates (mean times) . This results in an error surface having a minimum at . An intensity plot of (16) is shown in Figure 12. The CDF and its estimator are overlaid in Figure 13, and the error in this fit is magnified in Figure 14. Using our estimates, the mean lifetime PDF and CDF are shown in Figure 15.
Figure 15: [Mean lifetime estimated PDF and CDF.] Our lifetimebased population parameter estimation implies these distributions of mean lifetimes for the documents observed by the Informant. Note that these mean values are to be distinguished from the distribution of observed lifetimes. The average is around 138 days, the most likely value is 62 days, and the median is 117 days.
The lifetimebased estimates differ substantially from the agebased estimates, but are also more trustworthy, as can been seen by comparing the quality of the fit in Figures 13 and 9. There are two reasons for the difference. First, the assumption of exponential growth used for the agebased estimatation is probably a poor one, as true growth is much slower. Forcing exponential growth on a more slowly growing population forces the dynamics to be underrepresented, driving our estimates away from their true value. The lifetimebased estimation is not perfect either, as change rates may not be independent of observation timespan. A change in a page might very well push it into or out of a user's set of search results. We count on the fact that in observing faster than the search engines, we can observe changes before these force a result from the top of the list. It is difficult to justify an assumption of any particular dependence, since this relationship is controlled by many unknown factors (reindexing time for search engines used and result ranking strategy, for example).
6. How fast do search engines need to work
We now interpret our model of the constantly changing web in terms of web search engine performance. Our measure of performance is based on the intuitive concept of currency that we define below. Our web model and this new performance measure will allow us to estimate the speed at which pages must be reindexed in order to maintain a given level of currency.
Recall from the introduction that a web page's index entry in a search engine is current if the web page has not changed since the last time the page was reindexed and time units ago. We are willing to forgive changes that have occurred within time of the present. The grace period, , relaxes the temporal aspect of what it means to be current. The smaller is, the more "current" our information about the page is. See Figure 16 for a graphical depiction of the concept.
Figure 16: [Definition of ``betacurrent''.] This diagram shows what is meant when we say that an index entry is current with respect to a grace period, beta. In order to be betacurrent, no modification can go unobserved up to beta time units before the present.
To determine whether or not an index entry for a web page is current, we need to know the most recent time at which the page changed. Assume that the page was last observed at time t_{o}. With this notation, the index entry corresponding to a page is current at time t_{n} if the page did not change between the last observation (at time t_{o}) and units before the present, or time (assuming ). For , the entry is by definition current because the most recent unobserved page change can occur either within the grace period or before we observed the page at t_{o} , but this includes all past time.
Combining these two cases, the probability that the search engine entry for a page is current at time is
(17) 
where these probabilities are understood to be for a fixed, given web page. We now compute the probability that the search engine index entry for a randomly (according to some probability distribution) selected web page is current.
The above expression (17) for a single web page is stated in terms of a conditional probability. Given a prior distribution on the variables t _{o} and t _{n}, we can use Bayes' Theorem or the total probability theorem to eliminate them.
In our model, each web page has a change rate and an associated distribution of reindexing times T (a periodic reindexing system will have a single constant T _{0}). These parameters determine density functions which, together with the grace period , specify the probability of being current. First, define the probability to be the probability of a single index entry being current given , T, , and the time t_{n} at which the index is examined. Second, define the density to be the joint probability density for . We assume that is independent of the time t_{n} , which is distributed according to a density x(t_{n}). Using these densities and Bayes' Theorem, the probability that the system is current is
=  
=  (18) 
The integral is restricted to the first octant since no negative times or rates are allowed. In some settings, it is reasonable to assume a dependence between T and , since different revisitation periods are desirable for sources with different change rates.
We will now evaluate (18) for a single, memorylesslychanging page. As before, this page has a change rate , and is observed periodically (every T time units). The probability that the next page change occurs in the time interval [t_{1},t_{2}], where the last observation or change (whichever occurred most recently) was at time , is
If t_{1}=t_{o}, this reduces to so that the probability that a page change did not occur in the interval [t_{o},t_{2}] is the complement, .
To evaluate (18) we need to specify the function as well as the distribution of times x(t_{n}) over which we average the currency of the index. First, we consider the limits on the inner integral over t_{n}. Assuming as we have that all the web pages change memorylessly, it is sufficient to evaluate the inner integral in (18) over a single observation period T, since adding additional periods would only replicate the integral over one period.
For convenience, we choose an interval starting at t_{o}=0, at which time an observation was last made, and extends until the time T at which the next observation occurs. Using this interval, the probability that the page does not change between t_{o}=0 and , and is therefore , is
(19) 
(20) 
Combining these, the expected probability of a single page being current over all values of the observation time t_{n}, using a uniform density x(t_{n})=1/T, is just an average value of the piecewisedefined on the interval . This gives
=  (21)  
=  (22) 
In the first integral of (21), the probability of being current is one when , since this would force any change to be within units of the present. We can clean up (22) by expressing as a fraction of T (that is, ) and setting . With these changes, (22) becomes a function of the dimensionless relative rate, z, and the ratio of the grace period to the observation period, . When z>1, a source is expected to change once or more prior to T, whereas z<1 suggests fewer than one change expected before T. What fraction of these changes fall within the grace period is loosely described by the parameter ; some curves are shown for different choices of in Figure 17.
Figure 17: [Probability of betacurrency vs. relative rate.] Expected value of the probability of being betacurrent, expressed as a function of relative rate (relating reobservation period and mean change time), and the grace period (expressed as a percentage of the reobservation time).
We note in passing some properties of the curves in Figure 17 that verify our intuition. First, note that the probability of being current goes to as the relative rate approaches infinity. High relative rate implies a web page which is observed much too slowly; the page changes many times between observations. As such, in the high rate limit, simply represents the percentage of these changes that occur during the grace period. For the case of low relative rate, where pages are sampled much faster than they change, the probability of a page being current approaches one, regardless of the grace period fraction .
Choosing a random web page to which we apply (22) is equivalent to selecting a value for . In our collections, as discussed earlier, we have observed that the mean time between changes roughly follows a Weibull distribution, (10), which is given by
(23) 
The change rate is the inverse of the mean time between changes, so we can replace in the integral with the change rate .
Using (23), along with the parameter values that resulted from our numerical optimization, we can determine the expected value of (22) over for our collection. This calculation for other collections or other demand distributions depends only on finding the distribution of mean change times for those collections. Our analysis uses a simple periodic, roundrobin reindexing schedule, where the revistation time T is the same for all sources. Since we propose visiting each page every T time units, an accurate model for a real engine would need to account for the growth of the collection over time.
For this preliminary analysis, we assume a constant web size to avoid this difficulty. Using the Weibull distribution for inverse change rates, the expected probability that a uniformly randomly selected page will be current in the search engine index is
(24) 
The integral (24) can only be evaluated in closed form when the Weibull shape parameter is 1; otherwise, numerical evaluation is required. The integral gives an for every pair , defining a search engine "performance surface." This surface can be interpreted in a number of ways. For example, we can choose a probability and determine all pairs that give that probability. Using our parameter choices from the lifetimebased optimization of (16), we have evaluated the integral and plotted it in Figures 18 and 19, which show the level set for . It is important to note that the revisitation times which result from this analysis are upper bounds since our analysis is based on the less volatile pages that provide timestamps.
From that plot, we can see that in order to maintain (0.95, 1day)current search engine, a reindexing period of 8.5 days is necessary. For (0.95, 1week)currency, a reindexing period of 18 days is necessary. Notice that these figures do not depend upon the number of documents in an index, so a reindexing period defines a set of pairs , regardless of changes in the size of the index. Alternatively, we can estimate effective bandwidth requirements to maintain a given level of currency for a uniform index of a given size. By "uniform" we mean that no documents are given any sort of preference; all are reindexed at the same rate. The effective bandwidth is not to be confused with the link bandwidth, it simply describes the overall processing rate, including download and analysis.
For example, an (0.95, 1day) index of the entire web, using the estimate of 800 million pages from [LG99], would require a total effective bandwidth of (approximately)
A more modest index, slightly closer to those actually in use, might have 150 million documents at (0.95, 1week) currency, requiring an effective bandwidth of around
Clearly, other reindexing schemes exist where T is not constant but is a function of ; see [CLW97] for some good discussion on possible schemes. When T is a function of , the integral (24) is modified by substituting in the function and evaluating along the appropriate line in the plane. Additional modifications to this development might include the addition of a noise term to the observation period and choosing the grace period as a function of the change rate .
Figure 18: [Probability alpha as a function of grace period and observation period] Here, we plot the probability surface alpha as a function of the grace period fraction and fixed reindexing period. This surface results from using the more accurate lifetimebased population parameters, although this surface could be constructed for any population. The plane at alpha=0.95 intersects the surface in a level set.
Figure 19: [Relating grace period and observation period: alpha=95% level set.] Here we have plotted two level sets of pairs which yield a probability alpha=0.95 of being betacurrent. The two curves are derived from two different estimation methods, minimizing agebased error or lifebased error. The lifetimebased estimates are much more accurate. Regardless of the size of the collection, this data can be used to estimate how current an engine is when the indexing period takes on a value (in days) along the horizontal axis. As this period becomes large, the relative check rate is too slow, and beta approaches 95% of the period.
7. Summary
This paper describes our efforts at estimating how fast the web is changing, using a combination of empirical data and analytic modeling. From here, we can begin to consider the "dynamics" of information, and how best to deal with observation of changing information sources over limitedbandwidth channels. Much work remains to be done. With a reasonable model of how the web is growing and how fast pages change, we can start to formulate scheduling problems for search engines. These scheduling problems will depend on what objective we are trying to optimize. This work has used a simple, deterministic periodic revisiting strategy. By allowing different revisit intervals for different pages, we can formulate a variety of scheduling problems, holding two of , and the communication resources (that is, server bandwidth) fixed for example. We have not gone into any detail about which changes are "important" and which changes are not, nor have we delved into the reliability and popularity of the web pages in question. These clearly bear heavily on a user's perception of how good a search engine performs. While we have such data available to us in our empirical database, we have not yet addressed this. How can we estimate the currency, in our formal terms of ()currency, of commercial search engines that only allow external probes? How do the different search engines compare in this sense? Indeed, the fastchanging and fastgrowing web may soon force increased reliance on specialty search engines for the most volatile information sources.
Bibliography
 CLW97
 E. G. Coffman, Z. Liu, and R. R. Weber.
Optimal robot scheduling for web search engines.
Journal of Scheduling, 1997.
Available at http://www.inria.fr/mistral/personnel/Zhen.Liu/  DFKM97
 Fred Douglis, Anja Feldmann, Balachander Krishnamurthy, and Jeffrey Mogul.
Rate of change and other metrics: A live study of the world wide web.
In Proceedings of the USENIX Symposium on Internetworking Technologies and Systems, December 1997.
Available from http://www.research.att.com/~anja/feldmann/papers.html.  Fel71
 William Feller.
An Introduction to Probability Theory and Its Applications, volume 2.
Wiley, 2nd edition, 1971.  Gra97
 Matthew Gray.
Internet growth summary.
http://www.mit.edu/people/mkgray/net/internetgrowthrawdata.html, 1997.  Inf95
 Informant, 1995.
http://informant.dartmouth.edu/.  ISC99
 ISC, 1999.
Internet Software Consortium; http://www.isc.org/.  LG98
 S. Lawrence and C. L. Giles.
Searching the World Wide Web.
Science, 28:98100, April 1998.
Available by request at http://www.neci.nj.nec.com/homepages/lawrence/.  LG99
 S. Lawrence and C. L. Giles.
Accessibility of information on the web.
Nature, 1999.  MR94
 Douglas C. Montgomery and George C. Runger.
Applied Statistics and Probability for Engineers.
John Wiley and Sons, Inc., 1994.  Pap84
 Athanasios Papoulis.
Probability, Random Variables and Stochastic Processes.
McGrawHill, 2nd edition, 1984.
Vitae
Brian Brewington received a B.S. in Engineering and Applied Science from the California Institute of Technology in 1995. He began his doctoral research at the Thayer School of Engineering, Dartmouth College, with Professor George Cybenko in the fall of 1995. He will complete the program by late spring 2000. His academic interests include distributed information retrieval and signal processing, and he enjoys time away from work hiking and playing ultimate frisbee. 

George Cybenko (right, above) is the Dorothy and Walter Gramm Professor of Engineering at Dartmouth College. He has done pioneering work on several topics including load balancing for distributed computing, function approximation by neural networks and advanced algorithms for statistical signal processing. Cybenko's current areas of research include distributed information and computing systems, signal processing and mobile computing. In addition to serving on advisory and review boards at Argonne National Laboratory, the Minnesota Supercomputer Institute and the Institute for Mathematics and its Applications, he is the founding EditorinChief of Computing in Science and Engineering, jointly published by the IEEE Computer Society and the American Institute of Physics. Cybenko has B.Sc. (University of Toronto, 1974) and Ph.D. (Princeton, 1978) degrees in Mathematics. Prior to joining Dartmouth, Cybenko was Professor of Electrical and Computer Engineering and Computer Science at the University of Illinois at UrbanaChampaign and Associate Director of the Center for Supercomputer Research and Development. In 1996, he was the Kloosterman Distinguished Visiting Professor at Leiden University, the Netherlands. Cybenko is a Fellow of the IEEE. 
About this document ...
How dynamic is the web?^{1}
This document was generated using the LaTeX 2HTML translator Version 98.1p1 release (March 2nd, 1998)
Copyright © 1993, 1994, 1995, 1996, 1997, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
A great deal of effort was necessary to combine nodes and put this in the standard WWW9 submission format. For wider distribution, this document is also available as PostScript (compressed) or as PDF.
Footnotes
 ... web?^{1}
 This research was partially supported by AFOSR grant F496209710382, DARPA grant F306029820107 and NSF grant CCR9813744. Any opinions, findings, and conclusions are those of the authors and do not necessarily reflect the views of the above agencies.
 ... Informant"^{2}
 http://informant.dartmouth.edu.
 ... engines^{3}
 AltaVista, Excite, Infoseek, and Lycos.
 ... interesting^{4}
 These may not be around for long; before they disappear, see this listing of web servers from 1992 or this web guide from 1993.
 ... monitored^{5}
 This statistic obviously depends upon the length of time we monitor a web page.