Editing Cascading failure (section)

== In computer networks ==

Cascading failures can also occur in [[computer network]]s (such as the [[Internet]]) in which [[Network traffic control|network traffic]] is severely impaired or halted to or between larger sections of the network, caused by failing or disconnected hardware or software. In this context, the cascading failure is known by the term '''cascade failure'''. A cascade failure can affect large groups of people and systems.

The cause of a cascade failure is usually the overloading of a single, crucial [[Router (computing)|router]] or node, which causes the node to go down, even briefly. It can also be caused by taking a node down for maintenance or upgrades. In either case, traffic is [[routing|routed]] to or through another (alternative) path. This alternative path, as a result, becomes overloaded, causing it to go down, and so on. It will also affect systems which depend on the node for regular operation.

=== Symptoms ===

The symptoms of a cascade failure include: [[packet loss]] and high [[network latency]], not just to single systems, but to whole sections of a network or the internet. The high latency and packet loss is caused by the nodes that fail to operate due to [[congestion collapse]], which causes them to still be present in the network but without much or any useful communication going through them. As a result, routes can still be considered valid, without them actually providing communication.

If enough routes go down because of a cascade failure, a complete section of the network or internet can become unreachable. Although undesired, this can help speed up the recovery from this failure as connections will time out, and other nodes will give up trying to establish connections to the section(s) that have become cut off, decreasing load on the involved nodes.

A common occurrence during a cascade failure is a walking failure, where sections go down, causing the next section to fail, after which the first section comes back up. This ripple can make several passes through the same sections or connecting nodes before stability is restored.

=== History ===

Cascade failures are a relatively recent development, with the massive increase in traffic and the high interconnectivity between systems and networks. The term was first applied in this context in the late 1990s by a Dutch IT professional and has slowly become a relatively common term for this kind of large-scale failure.{{Citation needed|date=January 2009}}

=== Example ===

Network failures typically start when a single network node fails. Initially, the traffic that would normally go through the node is stopped. Systems and users get errors about not being able to reach hosts. Usually, the redundant systems of an ISP respond very quickly, choosing another path through a different backbone. The routing path through this alternative route is longer, with more [[Hop (telecommunications)|hops]] and subsequently going through more systems that normally do not process the amount of traffic suddenly offered.

This can cause one or more systems along the alternative route to go down, creating similar problems of their own.

Related systems are also affected in this case. As an example, [[Domain name system|DNS]] resolution might fail and what would normally cause systems to be interconnected, might break connections that are not even directly involved in the actual systems that went down. This, in turn, may cause seemingly unrelated nodes to develop problems, that can cause another cascade failure all on its own.

In December 2012, a partial loss (40%) of [[Gmail]] service occurred globally, for 18 minutes. This loss of service was caused by a routine update of load balancing software which contained faulty logic—in this case, the error was caused by logic using an inappropriate 'all' instead of the more appropriate 'some'.<ref>{{Cite web|url=https://arstechnica.com/information-technology/2012/12/why-gmail-went-down-google-misconfigured-chromes-sync-server/|title = Why Gmail went down: Google misconfigured load balancing servers (Updated)|date = 11 December 2012}}</ref> The cascading error was fixed by fully updating a single node in the network instead of partially updating all nodes at one time.