Memento: Help Us Route URI Lookups to the Right Archives

More Memento-enabled web archives are coming online every day, enabling aggregating services such as Time Travel and OldWeb. However, as the number of web archives grows, we must be able to better route URI lookups to the archives that are likely to have the requested URIs. We need assistance from IIPC members to help us better model both what archives contain as well as what people are looking for.

In our TPDL 2015 paper we found that less than 5% of the queried URIs have mementos in any individual archive that is not the Internet Archive. We created four different sample sets of one million URIs each and compared them against three different archives. The table below shows the percentage of the sample URIs found in various archives.

Sample (1M URIs Each) In Archive-It In UKWA In Stanford Union of {AIT, UK, SU}
DMOZ 4.097% 3.594% 0.034% 7.575%
Memento Proxy Logs 4.182% 0.408% 0.046% 4.527%
IA Wayback Logs 3.716% 0.519% 0.039% 4.165%
UKWA Wayback Logs 0.108% 0.034% 0.002% 0.134%

However, these small archives, when aggregated together prove to be much more useful and complete than they are individually. We found that the intersection between these archives is small, so the union of them is large (see the last column in the table above). The figure below shows the overlap among three archives for the sample of one million URIs from DMOZ.

stanford-ukwa-archive-it

We are working on an IIPC funded Archive Profiling project in which we are trying to create a high level summary of the holdings of each archive. Apart from the many other use cases, this will help us route the Memento Aggregator queries to only archives that are likely to return good results for a given URI.

We learned in the recent surge of oldweb.today (that uses MemGator to aggregate mementos from various archives) that some upstream archives had issues handling the sudden increase in the traffic and had to be removed from the list of aggregated archives. Another issue when aggregating large number of archives is that the aggregators follow the buffalo theory where the slowest upstream archive affects the roundtrip time of the aggregator. A single malfunctioning (or down) upstream archive may delay each aggregator response for the set timeout period. There are ways to solve the latter issue such as detecting continuously failing archives at runtime and temporarily disabling them from being aggregated. However, building Archive Profiles and predicting the probability of finding any Mementos in each archive to route the requests solves both the problems. Individual archives only get requests when they are likely to return good results, hence the routing saves their network and computing resources. Additionally, aggregators benefit in terms of the improved response time, because only a small subset of all the known archives is queried for any given URI.

We appreciate Andy Jackson of the UK Web Archive for providing the anonymised Wayback access logs that we used for sampling one of the URI sets. We would like to extend this study on other archives’ access logs to learn what people are looking for when they visit these archives. This will help us build sampling based profiling for archives that may not be able to share CDX files or generate/update full-coverage archive profiles.

We encourage all IIPC member archives to share their access logs just enough to generate at least one million unique URIs that people looked for in their archives. We are only interested in the log entries that have a URI-R in it (e.g., /wayback/14-digit-datetime/{URI}). We can handle all the cleanup and parsing tasks, or you can remove the requesting IP address from the logs (we don’t need it) if you would prefer. The logs can be continuous or consist of many sparse logs. We promise not to publish those logs in the raw form anywhere on the Web. Please feel free to discuss further details with me at salam@cs.odu.edu. Also contact me if you are interested in testing the software for profiling your archive.

by Sawood Alam
Department of Computer Science, Old Dominion University

Advertisements