"The network is reliable" tops Peter Deutsch's classic list, "Eight fallacies of distributed computing" (https://blogs.oracle.com/jag/resource/Fallacies.html), "all [of which] prove to be false in the long run and all [of which] cause big trouble and painful learning experiences." Accounting for and understanding the implications of network behavior is key to designing robust distributed programs—in fact, six of Deutsch's "fallacies" directly pertain to limitations on networked communications. This should be unsurprising: the ability (and often requirement) to communicate over a shared channel is a defining characteristic of distributed programs, and many of the key results in the field pertain to the possibility and impossibility of performing distributed computations under particular sets of network conditions.
For example, the celebrated FLP impossibility result9 demonstrates the inability to guarantee consensus in an asynchronous network (i.e., one facing indefinite communication partitions between processes) with one faulty process. This means that, in the presence of unreliable (untimely) message delivery, basic operations such as modifying the set of machines in a cluster (i.e., maintaining group membership, as systems such as Zookeeper are tasked with today) are not guaranteed to complete in the event of both network asynchrony and individual server failures. Related results describe the inability to guarantee the progress of serializable transactions,7 linearizable reads/writes,11 and a variety of useful, programmer-friendly guarantees under adverse conditions.3 The implications of these results are not simply academic: these impossibility results have motivated a proliferation of systems and designs offering a range of alternative guarantees in the event of network failures.5 Under a friendlier, more reliable network that guarantees timely message delivery, however, FLP and many of these related results no longer hold:8 by making stronger guarantees about network behavior, we can circumvent the programmability implications of these impossibility proofs.
Therefore, the degree of reliability in deployment environments is critical in robust systems design and directly determines the kinds of operations that systems can reliably perform without waiting. Unfortunately, the degree to which networks are actually reliable in the real world is the subject of considerable and evolving debate. Some people have claimed that networks are reliable (or that partitions are rare enough in practice) and that we are too concerned with designing for theoretical failure modes. Conversely, others attest that partitions do occur in their deployments, and that, as James Hamilton of AWS (Amazon Web Services) neatly summarizes (http://perspectives.mvdirona.com/2010/04/07/StonebrakerOnCAPTheoremAndDatabases.aspx), "Network partitions should be rare but net gear continues to cause more issues than it should." So who's right?
A key challenge in this discussion is the lack of evidence. We have few normalized bases for comparing network and application reliability—and even less data. We can track link availability and estimate packet loss, but understanding the end-to-end effect on applications is more difficult. The scant evidence we have is difficult to generalize: it is often deployment-specific and closely tied to particular vendors, topologies, and application designs. Worse, even when organizations have a clear picture of their network's behavior, they rarely share specifics. Finally, distributed systems are designed to resist failure, which means that noticeable outages often depend on complex interactions of failure modes. Many applications silently degrade when the network fails, and resulting problems may not be understood for some time, if ever.
As a result, much of what we believe about the failure modes of real-world distributed systems is founded on guesswork and rumor. Sysadmins and developers will swap stories over beer, but detailed, public postmortems and comprehensive surveys of network availability are few and far between. In this article, we'd like to informally bring a few of these stories (which, in most cases, are unabashedly anecdotal) together. Our focus is on descriptions of actual network behavior when possible and (more often), when not, on the implications of network failures and asynchrony for real-world systems deployments. We believe this is a first step toward a more open and honest discussion of real-world partition behavior, and, ultimately, toward more robust distributed systems design.
To start off, let's consider evidence from big players in distributed systems: companies running globally distributed infrastructure with hundreds of thousands of servers. These reports perhaps best summarize operations in the large, distilling the experience of operating what are likely the biggest distributed systems ever deployed. These companies' publications (unlike many of the reports we will examine later) often capture aggregate system behavior and large-scale statistical trends, and indicate (often obliquely) that partitions are of concern in their deployments.
A team from the University of Toronto and Microsoft Research studied the behavior of network failures in several of Microsoft's data centers.12 They found an average failure rate of 5.2 devices per day and 40.8 links per day, with a median time to repair of approximately five minutes (and a maximum of one week). While the researchers note that correlating link failures and communication partitions is challenging, they estimate a median packet loss of 59,000 packets per failure. Perhaps of more concern is their finding that network redundancy improves median traffic by only 43 percent—that is, network redundancy does not eliminate common causes of network failure.
A joint study between researchers at the University of California, San Diego, and HP Labs examined the causes and severity of network failures in HP's managed networks by analyzing support-ticket data (http://www.hpl.hp.com/techreports/2012/HPL-2012-101.pdf). "Connectivity"-related tickets accounted for 11.4 percent of support tickets (14 percent of which were of the highest-priority level), with a median incident duration of 2 hours and 45 minutes for the highest-priority tickets and a median duration of 4 hours and 18 minutes for all tickets.
Google's paper (http://research.google.com/archive/chubby-osdi06.pdf) describing the design and operation of Chubby, its distributed lock manager, outlines the root causes of 61 outages over 700 days of operation across several clusters. Of the nine outages that lasted more than 30 seconds, four were caused by network maintenance and two were caused by "suspected network connectivity problems."
In Design Lessons and Advice from Building Large Scale Distributed Systems (http://www.cs.cornell.edu/projects/ladis2009/talks/dean-keynote-ladis2009.pdf), Google Fellow Jeff Dean suggested that a typical first year for a new Google cluster involves:
• Five racks going wonky (40-80 machines seeing 50 percent packet loss).
• Eight network maintenance events (four of which might cause ~30-minute random connectivity losses).
• Three router failures (resulting in the need to pull traffic immediately for an hour).
While Google doesn't tell us much about the application-level consequences of its network partitions, Dean suggested that they were of concern, citing the perennial challenge of creating "easy-to-use abstractions for resolving conflicting updates to multiple versions of a piece of state," useful for "reconciling replicated state in different data centers after repairing a network partition."
Amazon's Dynamo paper (http://www.allthingsdistributed.com/files/amazon-dynamo-sosp2007.pdf) frequently cites the incidence of partitions as a key design consideration. Specifically, the authors note that they rejected designs from "traditional replicated relational database systems" because they "are not capable of handling network partitions."
Yahoo! PNUTS/Sherpa was designed as a distributed database operating in geographically distinct data centers. Originally, PNUTS supported a strongly consistent "timeline consistency" operation, with one master per data item. The developers noted, however, that in the event of network partitioning or server failures, this design decision was too restrictive for many applications:16
The first deployment of Sherpa supported the timeline-consistency model—namely, all replicas of a record apply all updates in the same order—and has API-level features to enable applications to cope with asynchronous replication. Strict adherence leads to difficult situations under network partitioning or server failures. These can be partially addressed with override procedures and local data replication, but in many circumstances, applications need a relaxed approach.
According to the same report, PNUTS now offers weaker consistency alternatives providing availability during partitions.
Data-center networks are subject to power failure, misconfiguration, firmware bugs, topology changes, cable damage, and malicious traffic. Their failure modes are accordingly diverse.
As Microsoft's SIGCOMM paper suggests, redundancy doesn't always prevent link failure. When a power distribution unit failed and took down one of two redundant top-of-rack switches, Fog Creek lost service for a subset of customers on that rack but remained consistent and available for most users. The other switch in that rack, however, also lost power for undetermined reasons. That failure isolated the two neighboring racks from each other, taking down all On Demand services.
During a planned network reconfiguration to improve reliability, Fog Creek Software suddenly lost access to its network.10
A network loop had formed between several switches. The gateways controlling access to the switch management network were isolated from each other, generating a split-brain scenario. Neither was accessible due to a...multi-switch BPDU (bridge protocol data unit) flood, indicating a spanning-tree flap. This is most likely what was changing the loop domain.
According to the BPDU standard, the flood shouldn't have happened. But it did, and this deviation from the system's assumptions resulted in two hours of total service unavailability.
To address high latencies caused by a daisy-chained network topology, Github installed a set of aggregation switches in its data center (https://github.com/blog/1346-network-problems-last-friday). Despite a redundant network, the installation process resulted in bridge loops, and switches disabled links to prevent failure. This problem was quickly resolved, but later investigation revealed that many interfaces were still pegged at 100 percent capacity.
While that problem was under investigation, a misconfigured switch triggered aberrant automatic fault-detection behavior: when one link was disabled, the fault detector disabled all links, leading to 18 minutes of downtime. The problem was traced to a firmware bug preventing switches from updating their MAC (media access control) address caches correctly, forcing them to broadcast most packets to every interface.
In December 2012 (https://github.com/blog/1364-downtime-last-saturday), a planned software update on an aggregation switch caused instability at Github. To collect diagnostic information, the network vendor killed a particular software agent running on one of the aggregation switches.
Github's aggregation switches are clustered in pairs using a feature called MLAG (multi-chassis link aggregation), which presents two physical switches as a single L2 (layer-2) device. The MLAG failure-detection protocol relies on both Ethernet link state and a logical heartbeat message exchanged between nodes. When the switch agent was killed, it was unable to shut down the Ethernet link, preventing the still-healthy aggregation switch from handling link aggregation, spanning-tree, and other L2 protocols. This forced a spanning-tree leader election and reconvergence for all links, blocking all traffic between access switches for 90 seconds.
This 90-second network partition caused file servers using Pacemaker and DRBD (Distributed Replicated Block Device) for HA (high availability) failover to declare each other dead, and to issue STONITH (shoot the other node in the head) messages to one another. The network partition delayed delivery of those messages, causing some file-server pairs to believe they were both active. When the network recovered, both nodes shot each other at the same time. With both nodes dead, files belonging to the pair were unavailable.
To prevent file-system corruption, DRBD requires that administrators ensure the original primary node is still the primary node before resuming replication. For pairs where both nodes were primary, the ops team had to examine log files or bring each node online in isolation to determine its state. Recovering those downed file-server pairs took five hours, during which Github service was significantly degraded.
Large-scale virtualized environments are notorious for transient latency, dropped packets, and full-blown network partitions, often affecting a particular software version or availability zone. Sometimes the failures occur between specific subsections of the provider's data center, revealing planes of cleavage in the underlying hardware topology.
In a comment on Call me maybe: MongoDB (http://aphyr.com/posts/284-call-me-maybe-mongodb), Scott Bessler observed exactly the same failure mode Kyle demonstrated earlier:
[This scenario] happened to us today when EC2 West region had network issues that caused a network partition that separated PRIMARY from its 2 SECONDARIES in a 3 node replset. 2 hours later the old primary rejoined and rolled back everything on the new primary.
This partition caused two hours of write loss. From our conversations with large-scale MongoDB users, we gather that network events causing failover on Amazon's EC2 (Elastic Compute Cloud) are common. Simultaneous primaries accepting writes for multiple days are anecdotally common.
Outages can leave two nodes connected to the Internet but unable to see each other. This type of partition is especially dangerous, as writes to both sides of a partitioned cluster can cause inconsistency and lost data. Paul Mineiro reports exactly this scenario in an Mnesia cluster (http://dukesoferl.blogspot.com/2008/03/network-partition-oops.html?m=1), which diverged overnight. The cluster's state wasn't critical, so the operations team simply nuked one side of the cluster. They conclude: "The experience has convinced us that we need to prioritize up our network partition recovery strategy."
Network disruptions in EC2 can affect only certain groups of nodes. For example, one report of a total partition between the front-end and back-end servers (https://forums.aws.amazon.com/thread.jspa?messageID=454155) states that a site's servers lose their connections to all back-end instances for a few seconds, several times a month. Even though the disruptions were short, they resulted in 30- to 45-minute outages and a corrupted index for ElasticSearch. As problems escalated, the outages occurred "2 to 4 times a day."
On April 21, 2011, AWS suffered unavailability for 12 hours,2 causing hundreds of high-profile Web sites to go offline. As a part of normal AWS scaling activities, Amazon engineers had shifted traffic away from a router in the EBS (Elastic Block Store) network in a single U.S. East AZ (Availability Zone), but, due to incorrect routing policies:
...many EBS nodes in the affected Availability Zone were completely isolated from other EBS nodes in its cluster. Unlike a normal network interruption, this change disconnected both the primary and secondary network simultaneously, leaving the affected nodes completely isolated from one another.
The partition, coupled with aggressive failure-recovery code, caused a mirroring storm that caused network congestion and triggered a previously unknown race condition in EBS. EC2 was unavailable for roughly 12 hours, and EBS was unavailable or degraded for more than 80 hours.
The EBS failure also caused an outage in Amazon's RDS (Relational Database Service). When one AZ fails, RDS is designed to failover to a different AZ; however, 2.5 percent of multi-AZ databases in U.S. East failed to failover because of a bug in the failover protocol.
This correlated failure caused widespread outages for clients relying on AWS. For example, Heroku reported between 16 and 60 hours of unavailability for its users' databases.
On July 18, 2013, Twilio's billing system, which stores account credits in Redis, failed.19 A network partition isolated the Redis primary from all secondaries. Because Twilio did not promote a new secondary, writes to the primary remained consistent. When the primary became visible to the secondaries again, however, all secondaries simultaneously initiated a full resynchronization with the primary, overloading it and causing Redis-dependent services to fail.
The ops team restarted the Redis primary to address the high load. Upon restart, however, the Redis primary reloaded an incorrect configuration file, which caused it to enter read-only mode. With all account balances at zero, and in read-only mode, every Twilio API call caused the billing system to recharge customer credit cards automatically, resulting in 1.1 percent of customers being overbilled over a period of 40 minutes. For example, Appointment Reminder reported that every SMS message and phone call it issued resulted in a $500 charge to its credit card, which stopped accepting charges after $3,500.
Twilio recovered the Redis state from an independent billing system—a relational data store—and after some hiccups, restored proper service, including credits to affected users.
Running your own data center can be cheaper and more reliable than using public cloud infrastructure, but it means you have to be a network and server administrator. What about hosting providers, which rent dedicated or virtualized hardware to users and often take care of the network and hardware setup for you?
Freistil IT hosts its servers with a colocation/managed-hosting provider. Its monitoring system alerted Freistil to a 50-100 percent packet loss localized to a specific data center.15 The network failure, caused by a router firmware bug, returned the next day. Elevated packet loss caused the GlusterFS distributed file system to enter split-brain undetected:
...we became aware of [problems] in the afternoon when a customer called our support hotline because their website failed to deliver certain image files. We found that this was caused by a split-brain situation...and the self-heal algorithm built into the Gluster file system was not able to resolve this inconsistency between the two data sets.
Repairing that inconsistency led to a "brief overload of the Web nodes because of a short surge in network traffic."
Anecdotally, many major managed hosting providers experience network failures. One company running 100-200 nodes on a major hosting provider reported that in a 90-day period the provider's network went through five distinct periods of partitions. Some partitions disabled connectivity between the provider's cloud network and the public Internet, and others separated the cloud network from the provider's internal managed-hosting network.
A post to Linux-HA details a long-running partition between a Heartbeat pair (http://readlist.com/lists/lists.linux-ha.org/linux-ha/6/31964.html), in which two Linode VMs each declared the other dead and claimed a shared IP for themselves. Successive posts suggest further network problems: e-mails failed to dispatch because of DNS (Domain Name System) resolution failure, and nodes reported, "Network unreachable." In this case, the impact appears to have been minimal, in part because the partitioned application was just a proxy.
While we have largely focused on failures over local area networks (or near-local networks), WAN (wide area network) failures are also common, if less frequently documented. These failures are particularly interesting because there are often fewer redundant WAN routes and because systems guaranteeing high availability (and disaster recovery) often require distribution across multiple data centers. Accordingly, graceful degradation under partitions or increased latency is especially important for geographically widespread services.
Researchers at the UCSD analyzed five years of operation in the CENIC (Corporation for Education Network Initiatives in California) WAN,18 which contains more than 200 routers across California. By cross-correlating link failures and additional external BGP (Border Gateway Protocol) and trace-route data, they discovered more than 500 "isolating network partitions" that caused connectivity problems between hosts. Average partition duration ranged from 6 minutes for software-related failures to more than 8.2 hours for hardware-related failures (median 2.7 and 32 minutes; 95th percentile of 19.9 minutes and 3.7 days, respectively).
PagerDuty designed its system to remain available in the face of node, data-center, or even provider failure; its services are replicated between two EC2 regions and a data center hosted by Linode. On April 13, 2013, an AWS peering point in northern California degraded, causing connectivity issues for one of PagerDuty's EC2 nodes. As latencies between AWS Availability Zones rose, the notification dispatch system lost quorum and stopped dispatching messages entirely.
Even though PagerDuty's infrastructure was designed with partition tolerance in mind, correlated failures caused by a shared peering point between two data centers resulted in 18 minutes of unavailability, dropping inbound API requests and delaying queued pages until quorum was reestablished.
Despite the high level of redundancy in Internet systems, some network failures take place on a global scale.
CloudFlare runs 23 data centers with redundant network paths and anycast failover. In response to a DDoS (distributed denial-of-service) attack against one of its customers, the CloudFlare operations team deployed a new firewall rule to drop packets of a specific size.17 Juniper's FlowSpec protocol propagated that rule to all CloudFlare edge routers—but then:
What should have happened is that no packet should have matched that rule because no packet was actually that large. What happened instead is that the routers encountered the rule and then proceeded to consume all their RAM until they crashed.
Recovering from the failure was complicated by routers that failed to reboot automatically and by inaccessible management ports.
Even though some data centers came back online initially, they fell back over again because all the traffic across our entire network hit them and overloaded their resources.
CloudFlare monitors its network carefully, and the operations team had immediate visibility into the failure. Coordinating globally distributed systems is complex, however, and calling on-site engineers to find and reboot routers by hand takes time. Recovery began after 30 minutes and was complete after an hour of unavailability.
A firmware bug introduced as a part of an upgrade in Juniper Networks' routers caused outages in Level 3 Communications' networking backbone in 2011. This subsequently knocked services offline, including Time Warner Cable, RIM BlackBerry, and several UK Internet service providers.
There have been several global Internet outages related to BGP misconfiguration. Notably, in 2008 Pakistan Telecom, responding to a government edict to block YouTube.com, incorrectly advertised its (blocked) route to other providers, which hijacked traffic from the site and briefly rendered it unreachable.
In 2010 a group of Duke University researchers achieved a similar effect by testing an experimental flag in the BGP (http://www.merit.edu/mail.archives/nanog/msg11505.html). Similar incidents occurred in 2006, knocking sites such as Martha Stewart Living and the New York Times offline; in 2005, where a misconfiguration in Turkey attempted a redirect for the entire Internet; and in 1997.
Unreliable networking hardware and/or drivers are implicated in a broad array of partitions.
As a classic example of NIC (network interface controller) unreliability, Marc Donges and Michael Chan (http://www.spinics.net/lists/netdev/msg210485.html) describe how their popular Broadcom BCM5709 chip dropped inbound but not outbound packets. The primary server was unable to service requests, but, because it could still send heartbeats to its hot spare, the spare considered the primary alive and refused to take over. Their service was unavailable for five hours and did not recover without a reboot.
Sven Ulland followed up, reporting the same symptoms with the BCM5709S chipset on Linux 2.6.32-41squeeze2. Despite pulling commits from mainline, which supposedly fixed a similar set of issues with the bnx2 driver, Ulland's team was unable to resolve the issue until version 2.6.38.
As a large number of servers shipped the BCM5709, the larger impact of these firmware bugs was widely observed. For example, the 5709 had a bug in the 802.3x flow control, leading to extraneous PAUSE frames when the chipset crashed or its buffer filled up. This problem was magnified by the BCM56314 and BCM56820 switch-on-a-chip devices (found in many top-of-rack switches), which, by default, sent PAUSE frames to any interface communicating with the offending 5709 NIC. This led to cascading failures on entire switches or networks.
The bnx2 driver could also cause transient or flapping network failures, as described in an ElasticSearch failure report. Meanwhile, the Broadcom 57711 was notorious for causing high latencies under load with jumbo frames, a particularly thorny issue for ESX users with iSCSI-backed storage.
A motherboard manufacturer failed to flash the EEPROM correctly for its Intel 82574-based system. The result was a very-hard-to-diagnose error in which an inbound SIP (Session Initiation Protocol) packet of a particular structure would disable the NIC.14 Only a cold restart would bring the system back to normal.
After a scheduled upgrade, CityCloud noticed unexpected network failures in two distinct GlusterFS pairs, followed by a third.6 Suspecting link aggregation, CityCloud disabled the feature on its switches and allowed self-healing operations to proceed.
Roughly 12 hours later, the network failures returned. CityCloud identified the cause as a driver issue and updated the downed node, returning service. The outage, however, resulted in data inconsistency between GlusterFS pairs and data corruption between virtual machine file systems.
Not all asynchrony originates in the physical network. Sometimes dropped or delayed messages are a consequence of crashes, program errors, operating-system scheduler latency, or overloaded processes. The following studies highlight the fact that communication failures—wherein the system delays or drops messages—can occur at any layer of the software stack, and that designs that expect synchronous communication may behave unexpectedly during periods of asynchrony.
Bonsai.io discovered (http://www.bonsai.io/blog/2013/03/05/outage-post-mortem) high CPU and memory use on an ElasticSearch node combined with difficulty connecting to various cluster components, likely a consequence of an "excessively high number of expensive requests being allowed through to the cluster."
Upon restarting the servers, the cluster split into two independent components. A subsequent restart resolved the split-brain behavior, but customers complained they were unable to delete or create indices. The logs revealed that servers were repeatedly trying to recover unassigned indices, which "poisoned the cluster's attempt to service normal traffic which changes the cluster state." The failure led to 20 minutes of unavailability and six hours of degraded service.
Stop-the-world garbage collection and blocking for disk I/O can cause runtime latencies on the order of seconds to minutes. As Searchbox IO and several other production users (https://github.com/elasticsearch/elasticsearch/issues/2488) have found, GC (garbage collection) pressure in an ElasticSearch cluster can cause secondary nodes to declare a primary dead and to attempt a new election. Because of nonmajority quorum configuration, ElasticSearch elected two different primaries, leading to inconsistency and downtime. Surprisingly, even with majority quorums, due to protocol design, ElasticSearch does not currently prevent simultaneous master election; GC pauses and high IO_WAIT times due to I/O can cause split-brain behavior, write loss, and index corruption.
In 2012, a routine database migration caused unexpectedly high load on the MySQL primary at Github.13 The cluster coordinator, unable to perform health checks against the busy MySQL server, decided the primary was down and promoted a secondary. The secondary had a cold cache and performed poorly, causing failover back to the original primary. The operations team manually halted this automatic failover, and the site appeared to recover.
The next morning, the operations team discovered that the standby MySQL node was no longer replicating changes from the primary. Operations decided to disable the coordinator's maintenance mode and allow the replication manager to fix the problem. Unfortunately, this triggered a segfault in the coordinator, and a conflict between manual configuration and the automated replication tools rendered github.com unavailable.
The partition caused inconsistency in the MySQL database—both between the secondary and primary, and between MySQL and other data stores such as Redis. Because foreign key relationships were not consistent, Github showed private repositories to the wrong users' dashboards and incorrectly routed some newly created repositories.
When a two-node cluster partitions, there are no cases in which a node can reliably declare itself to be the primary. When this happens to a DRBD file system, as one user reported (http://serverfault.com/questions/485545/dual-primary-ocfs2- drbd-encountered-split-brain-is-recovery-always-going-to-be), both nodes can remain online and accept writes, leading to divergent file system-level changes.
Short-lived failures can lead to long outages. In a Usenet post to novell.support.cluster-services, an admin reports that a two-node failover cluster running Novell NetWare experienced transient network outages. The secondary node eventually killed itself, and the primary (though still running) was no longer reachable by other hosts on the network. The post goes on to detail a series of network partition events correlated with backup jobs.
One VoltDB user reports regular network failures causing replica divergence (https://forum.voltdb.com/showthread.php?552-Nodes-stop -talking-to-each-other-and-form-independent-clusters) but also indicates that the network logs included no dropped packets. Because this cluster had not enabled split-brain detection, both nodes ran as isolated primaries, causing significant data loss.
Sometimes, nobody knows why a system partitions. This RabbitMQ failure (http://serverfault.com/questions/497308/rabbitmq-network-partition-error) seems like one of those cases: few retransmits, no large gaps between messages, and no clear loss of connectivity between nodes. Increasing the partition-detection time-out to two minutes reduced the frequency of partitions but didn't prevent them altogether.
Another EC2 split-brain (http://elasticsearch-users.115913.n3.nabble.com/EC2-discovery-leads-to-two-masters-td3239318.html): a two-node cluster failed to converge on "roughly 1 out of 10 startups" when discovery messages took longer than three seconds to exchange. As a result, both nodes would start as primaries with the same cluster name. Since ElasticSearch doesn't demote primaries automatically, split-brain persisted until administrators intervened. Increasing the discovery time-out to 15 seconds resolved the issue.
There are a few scattered reports of Windows Azure partitions, such as one account (http://rabbitmq.1065348.n5.nabble.com/Instable-HA-cluster-td24690.html) of a RabbitMQ cluster that entered split-brain on a weekly basis. There's also a report of an ElasticSearch split-brain (https://groups.google.com/forum/?fromgroups#!topic/elasticsearch/muZtKij3nUw), but since Azure is a relative newcomer compared with EC2, descriptions of its network reliability are limited.
This article is meant as a reference point—to illustrate that, according to a wide range of (often informal) accounts, communication failures occur in many real-world environments. Processes, servers, NICs, switches, and local and wide area networks can all fail, with real economic consequences. Network outages can suddenly occur in systems that have been stable for months at a time, during routine upgrades, or as a result of emergency maintenance. The consequences of these outages range from increased latency and temporary unavailability to inconsistency, corruption, and data loss. Split-brain is not an academic concern: it happens to all kinds of systems—sometimes for days on end. Partitions deserve serious consideration.
On the other hand, some networks really are reliable. Engineers at major financial firms have anecdotally reported that despite putting serious effort into designing systems that gracefully tolerate partitions, their networks rarely, if ever, exhibit partition behavior. Cautious engineering and aggressive network advances (along with lots of money) can prevent outages. Moreover, in this article, we have presented failure scenarios; we acknowledge it's much harder to demonstrate that network failures have not occurred.
Not all organizations, however, can afford the cost or operational complexity of highly reliable networks. From Google and Amazon (which operate commodity and/or low-cost hardware because of sheer scale) to one-person startups built on shoestring budgets, communication-isolating network failures are a real risk, in addition to the variety of other failure modes (including human error) that real-world distributed systems face.
It's important to consider this risk before a partition occurs, because it's much easier to make decisions about partition behavior on a whiteboard than to redesign, reengineer, and upgrade a complex system in a production environment—especially when it's throwing errors at your users. For some applications, failure is an option—but you should characterize and explicitly account for it as a part of your design. Finally, given the additional latency1 and coordination benefits4 of partition-aware designs, you might just find that accounting for these partitions delivers benefits in the average case as well.
We invite you to contribute your own experiences with or without network partitions. Open a pull request on https://github.com/aphyr/partitions-post (which, incidentally, contains all references), leave a comment, write a blog post, or release a postmortem. Data will inform this conversation, future designs, and, ultimately, the availability of the systems on which we all depend.
1. Abadi, D. 2012. Consistency tradeoffs in modern distributed database system design: CAP is only part of the story. Computer 45(2): 37-42; http://dl.acm.org/citation.cfm?id=2360959.
2. Amazon Web Services. 2011. Summary of the Amazon EC2 and Amazon RDS service disruption in the U.S. East region; http://aws.amazon.com/message/65648/.
3. Bailis, P., Davidson, A., Fekete, A., Ghodsi, A., Hellerstein, J.M., Stoica, I. 2014. Highly available transactions: virtues and limitations. In Proceedings of VLDB (to appear); http://www.bailis.org/papers/hat-vldb2014.pdf.
4. Bailis, P., Fekete, A., Franklin, M.J., Ghodsi, A., Hellerstein, J.M., Stoica, I. 2014. Coordination-avoiding database systems; http://arxiv.org/abs/1402.2237.
5. Bailis, P., Ghodsi, A. 2013. Eventual consistency today: limitations, extensions, and beyond. ACM Queue 11(3); http://queue.acm.org/detail.cfm?id=2462076.
6. CityCloud. 2011; https://www.citycloud.eu/cloud-computing/post-mortem/.
7. Davidson, S.B., Garcia-Molina, H., Skeen, D. 1985. Consistency in a partitioned network: a survey. ACM Computing Surveys 17(3): 341-370; http://dl.acm.org/citation.cfm?id=5508.
8. Dwork, C., Lynch, M., Stockmeyer, L. 1988. Consensus in the presence of partial synchrony. Journal of the ACM 35(2): 288-323. http://dl.acm.org/citation.cfm?id=42283.
9. Fischer, M. J., Lynch, N.A., Patterson, M.S. 1985. Impossibility of distributed consensus with one faulty process. Journal of the ACM 32(2): 374-382; http://dl.acm.org/citation.cfm?id=214121.
10. Fog Creek Software. 2012. May 5-6 network maintenance post-mortem; http://status.fogcreek.com/2012/05/may-5-6-network-maintenance-post-mortem.html.
11. Gilbert, S., Lynch, N. 2002. Brewer's conjecture and the feasibility of consistent, available, partition-tolerant Web services. ACM SIGACT News 33(2): 51-59; http://dl.acm.org/citation.cfm?id=564601.
12. Gill, P., Jain, N., Nagappan, N. 2011. Understanding network failures in data centers: measurement, analysis, and implications. In Proceedings of SIGCOMM; http://research.microsoft.com/en-us/um/people/navendu/papers/sigcomm11netwiser.pdf.
13. Github. 2012. Github availability this week; https://github.com/blog/1261-github-availability-this-week.
14. Kielhofner, K. 2013. Packets of death; http://blog.krisk.org/2013/02/packets-of-death.html.
15. Lillich, J. 2013. Post mortem: network issues last week; http://www.freistil.it/2013/02/post-mortem-network-issues-last-week/.
16. Narayan, P.P.S. 2010. Sherpa update; https://developer.yahoo.com/blogs/ydn/sherpa-7992.html#4.
17. Prince, M. 2013. Today's outage post mortem; http://blog.cloudflare.com/todays-outage-post-mortem-82515.
18. Turner, D., Levchenko, K., Snoeren, A., Savage, S. 2010. California fault lines: understanding the causes and impact of network failures. In Proceedings of SIGCOMM ; http://cseweb.ucsd.edu/~snoeren/papers/cenic-sigcomm10.pdf.
19. Twilio. 2013. Billing incident post-mortem: breakdown, analysis and root cause; http://www.twilio.com/blog/2013/07/billing-incident-post-mortem.html.
LOVE IT, HATE IT? LET US KNOW
Peter Bailis is a graduate student of computer science and a member of the AMPLab and BOOM projects at UC Berkeley. He currently studies database and distributed systems, with a particular focus on fast and scalable data serving and transaction processing. He holds an A.B. from Harvard College and is the recipient of the NSF Graduate Research Fellowship and the Berkeley Fellowship for Graduate Study. He blogs regularly at http://bailis.org/blog and tweets as @pbailis.
Kyle Kingsbury is the author of Riemann, Timelike, and a slew of other open-source packages. In his free time he verifies distributed systems' safety claims as a part of the Jepsen project.
© 2014 ACM 1542-7730/14/0700 $10.00
Originally published in Queue vol. 12, no. 7—
see this item in the ACM Digital Library
Neal Cardwell, Yuchung Cheng, C. Stephen Gunn, Soheil Hassas Yeganeh, Van Jacobson - BBR: Congestion-Based Congestion Control
Measuring bottleneck bandwidth and round-trip propagation time
Josh Bailey, Stephen Stuart - Faucet: Deploying SDN in the Enterprise
Using OpenFlow and DevOps for rapid development
Amin Vahdat, David Clark, Jennifer Rexford - A Purpose-built Global Network: Google's Move to SDN
A discussion with Amin Vahdat, David Clark, and Jennifer Rexford
Harlan Stenn - Securing the Network Time Protocol
Crackers discover how to use NTP as a weapon for abuse.
(newest first)Having designed realtime distributed conference calling systems in the early 2000s based on SIP softswitches running in datacenters with engineering attempting to replicate telco-grade reliability, which at the application level attempted but never honestly achieved 4 nines on commodity hardware running on N+N machines and which maintain state in replicated SQL databases across two datatcenters I have experienced a number of these failure scenarios, from the should-never-go-down-dual-loop-fiber-WAN link going down, datacenter-wide power outages due to generator failure and UPS failure combined with datacenter power circuit engineering mistakes, slow system (sometimes) and slow network connectivity (hardware issues where a switch didn't go down but just hung, absorbing traffic without response) triggering failure detection and failover processes leading to dual-primary split brains, DNS hangs/failures affecting one or both sites, NTP failures causing bad dates at one site resulting in unintended consequences, NTP failures and slow clock skew arising across sites, etc. The rules and practices around designing a network, and application, to avoid and deal with these things is a little like the fire code; it accumulates over time based on the painful experience of disaster.
I will make a few observations. 1) If one doesn't personally receive alerts whenever failover triggering happens, one doesn't necessarily intuitively recognize the frequency of these partition events. Resiliency is often a non-event otherwise. If you silently and successfully failover every two weeks and don't know it, your notion of how reliable your system is may cloak real problems. And if one doesn't appropriately log and analyze-over-time failover-trigger logs one doesn't have good perspective. To learn from mistakes one might be wise adopt the telco industry practice of every failure resulting in a Reason For Outage document, the corpus of which can be periodically reviewed. 2) If one doesn't have some post-event integrity checking one doesn't necessarily recognize the consequences of a failover. The more resilient one makes their system, ironically the less obvious it can be that there is some remaining integrity issue once you have mitigated how your application deals with the most critical scenarios. 3) If you haven't taken the time+expense to inject hard failures and hanging behavior in each component of your system you are almost certainly not going to experience reliability , but in practice, often one doesn't have the luxury of testing failures of your network-wide infrastructure like colo-wide or site-wide switches, and for certain devices it is hard to simulate the "hanging" failure case which is different than the "down" failure case. 4) Slow-systems-triggering-failover events due to the database or garbage collection being busy can often be papered over, but not fully resolved, by overprovisioning or tuning and will tend to crop up later in other usage scenarios. 5) Ultimately it's worth trying to rethink one's design to avoid as much network communication as possible.