10/12/2026 | Press release | Distributed by Public on 10/12/2026 13:36
(Image/Dall-E)
Picture submitting an online application minutes before a deadline. The page just spins. Your internet is fine - other sites load without trouble. The application website hasn't crashed either. Somewhere between your computer and that one server, the connection has simply broken.
This happens more often than researchers once thought. A user and a website can both be online, yet a failure somewhere in between still blocks the connection, quietly cutting off a transaction, a payment, or access to a service someone was counting on.
"Even in a connected internet, even where all the wires are fine, some [computers] can't talk to each other," said John Heidemann, who has studied internet outages for more than a decade.
That is the subject of"Understanding Partial Reachability in the Internet Core," a 2026 paper by Heidemann, principal scientist at the USC Information Sciences Institute who serves as chief scientist for ISI's Networking and Cybersecurity Division; Guillermo Baltra, a USC Viterbi Ph.D. who now serves as an officer in the Chilean Navy; Yuri Pradkin, an ISI senior research engineer; and Tarang Saluja of Swarthmore College. Heidemann is also a research professor in USC Viterbi and the Thomas Lord Department of Computer Science in the USC Stevens School of Computing and AI.
The researchers' central finding: partial failures are at least as common as complete outages. Because partial failures affect some users while sparing others, they are far harder to detect. Finding them could help internet providers repair broken connections, reroute traffic and pinpoint what caused the failure in the first place.
Most people picture the internet as either up or down. In reality, it is made up of many connected networks. A cellphone and a university computer may reach the same website through different internet companies, cables and equipment.
If one of those routes fails, a website may work for one person but not another. Heidemann calls this divide a partition: The user and the website are still online, but information cannot travel between them.
Researchers found that, at any given moment, roughly 1 in 1,000 monitored groups of internet addresses was completely unreachable; partial outages were just as common, with about 1 in 1,000 reachable from some locations but not others. Large services like Amazon have many network connections, so they can reroute their traffic if one path fails. Smaller websites with fewer backup routes can leave some users unable to reach them.
So why has this gone unnoticed for so long? A site that is completely off the Internet (a "complete outage") is obvious: no one can reach the destination. A partition, where only some people cannot reach the site, is subtler, and the failure hard to pinpoint.
Some partial failures vanish as the internet finds another path. Others last for years because two companies refuse to carry each other's data or a government restricts access. The problems can continue until the companies reach an agreement or officials lift the restrictions.
The paper documents a striking case: Hurricane Electric, a global network operator based in Fremont, California, and Cogent Communications, a major provider headquartered in Washington, D.C. Since 2009, the two companies have refused to exchange traffic directly over the newer version of internet addressing. Their older connections still work, but customers using the newer standard can't reach websites or services hosted on the other company's network.
"This works great until it doesn't," Heidemann said of the business arrangements that knit the internet together. When talks fail, he added, a provider can effectively tell a rival, "You don't get access to my customers."
The work grew from more than 20 years of internet observation and includes data collected from 2017 through 2025. The study received support from the National Science Foundation and the Defense Advanced Research Projects Agency, with earlier work supported by the Department of Homeland Security.
The researchers looked for cases where computers in different locations tried to reach the same destination; some got through, others didn't. That mixed pattern signals a broken route, not a dead site. They tracked it across years of records covering millions of internet addresses, then verified their findings against an independent dataset.
Heidemann's research has already led to improvements in some systems. RIPE, a European networking organization that also monitors the Internet's naming system, saw a surprisingly high number of name system failures. Heidemann's team found that more than half of those reports were false, triggered by connection problems about where RIPE's monitoring devices were placed, not problems with the internet itself. RIPE was aware of this problem, but did not realize how often it occurred. It is now retooling its monitoring to filter out the noise to make real failures more obvious.
Beyond helping RIPE improve its monitoring, the USC researchers have made their findings available through a website maintained by USC's Information Sciences Institute. Daily reports show where connections failed and which locations were affected. Researchers and internet providers can request access to the data to study recurring problems, investigate broken connections and explore ways to route traffic around them.
The database can't tell providers exactly how to fix a given problem. But by surfacing failures that once went unnoticed, it gives them somewhere to start: investigating broken routes, correcting technical errors or resolving the business disputes that keep customers locked out of parts of the internet.
Heidemann hopes the research brings both clarity and pressure. "Some of the sunlight we spread on the problem of business practices that interfere with connectivity," he said, "will perhaps encourage companies to reach an agreement more quickly."
Published on October 12th, 2026
Last updated on October 12th, 2026