In large distributed environments such as the Internet, systems designers have typically been willing to live with some degree of cache inconsistency to reduce server hot spots, network traffic, and retrieval latency. In reference [19], we evaluate hierarchical invalidation schemes for Internet caching, based on traces of Internet Web traffic and based on the topology of the current U.S. Internet backbone. In the absence of explicit expiration times and last-modification times, this study found that with a default cache TTL of 5 days, 20%of references are stale. Further, it found that the network cost of hierarchical invalidation exceeded the savings of Web caching for default cache TTLs shorter than five days.
Note that achieving a well-working hierarchical invalidation scheme will not be easy. First, hierarchical invalidation requires support from all data providers and caches. Second, invalidation of widely shared objects will cause bursts of synchronization traffic. Finally, hierarchical invalidation cannot prevent stale references, and would require considerable complexity to deal with machine failures.
At present we do not believe hierarchical invalidation can practically replace TTL based consistency in a wide-area distributed environment. However, part of our reluctance to recommend hierarchical invalidation stems from the current informal nature of Internet information. While most data available on the Internet today cause no problems even if stale copies are retrieved, as the Internet evolves to support more mission critical needs, it may make sense to try to overcome the hurdles of implementing a hybrid hierarchical invalidation mechanism for the applications that demand data coherence.