Failure is inevitable. Disks fail. Software bugs lie dormant waiting for just the right conditions to bite. People make mistakes. Data centers are built on farms of unreliable commodity hardware. If you're running in a cloud environment, then many of these factors are outside of your control. To compound the problem, failure is not predictable and doesn't occur with uniform probability and frequency. The lack of a uniform frequency increases uncertainty and risk in the system. In the face of such inevitable and unpredictable failure, how can you build a reliable service that provides the high level of availability your users can depend on?
A naive approach could attempt to prove the correctness of a system through rigorous analysis. It could model all different types of failures and deduce the proper workings of the system through a simulation or another theoretical framework that emulates or analyzes the real operating environment. Unfortunately, the state of the art of static analysis and testing in the industry hasn't reached those capabilities.4
A different approach could attempt to create exhaustive test suites to simulate all failure modes in a separate test environment. The goal of each test suite would be to maintain the proper functioning of each component, as well as the entire system when individual components fail. Most software systems use this approach in one form or another, with a combination of unit and integration tests. More advanced usage includes measuring the coverage surface of tests to indicate completeness.
While this approach does improve the quality of the system and can prevent a large class of failures, it is insufficient to maintain resilience in a
Yet another approach, advocated in this article, is to induce failures in the system to empirically demonstrate resilience and validate intended behavior. Given that the system was designed with resilience to failures, inducing those
Before going further, let's discuss what is meant by resilience and how to increase it. Resilience is an attribute of a system that enables it to deal with failure in a way that doesn't cause the entire system to fail. It could involve minimizing the blast radius when a failure occurs or changing the user experience to work around a failing component. For example, if a movie recommendation service fails, the user can be presented with a nonpersonalized list of popular titles. A complex system is constantly undergoing varying degrees of failure. Resilience is how it can recover, or be insulated, from failure, both current and future.7
There are two ways of increasing the resilience of a system:
• Build your application with redundancy and fault tolerance. In a
• Reduce uncertainty by regularly inducing failure. Increasing the frequency of failure reduces its uncertainty and the likelihood of an inappropriate or unexpected response. Each unique failure can be induced while observing the application. For each undesirable response to an induced failure, the first approach can be applied to prevent its recurrence. Although in practice it is not feasible to induce every possible failure, the exercise of enumerating possible failures and prioritizing them helps in understanding tolerable operating conditions and classifying failures when they fall outside those bounds.
The first item is well covered in other literature. The remainder of this article will focus on the second.
Once you have accepted the idea of inducing failure regularly, there are a few choices on how to proceed. One option is GameDays,1 a set of scheduled exercises where failure is manually introduced or simulated to mirror
But what if you want a solution that is more scalable and
One way of achieving this is to engineer failure to occur in the live environment. This is how the idea for "monkeys" (autonomous agents really, but monkeys inspire the imagination) came to Netflix to wreak havoc and induce failure. Later the monkeys were grouped together and labeled the Simian Army.5 A description of each
The failure of a virtual instance is the most common type of failure encountered in a typical public cloud environment. It can be caused by a power outage in the hosting rack, a disk failure, or a network partition that cuts off access. Regardless of the cause, the result is the same: the instance becomes unavailable. Inducing such failures helps ensure that services don't rely on any
To address this need, Netflix created its first monkey: Chaos Monkey, which randomly terminates virtual instances in a production
Chaos Monkey starts by looking into a service registry to find all the services that are running. In Netflix's case, this is done through a combination of Asgard6 and Edda.2 Each service can override the default Chaos Monkey configuration to change termination probability or opt out entirely. Each hour, Chaos Monkey wakes up, rolls the dice, and terminates the affected instances using AWS (Amazon Web Services) APIs.
Chaos Monkey can optionally send an
With Chaos Monkey, a system is resilient to individual instance failure, but what if an entire data center were to become unavailable? What would be the impact to users if an entire Amazon AZ (availability zone) went offline? To answer that question and to make sure such an event would have minimal customer impact, Netflix created Chaos Gorilla.
Chaos Gorilla causes an entire AZ to fail. It simulates two failure modes:
• Network partition. The instances in the zone are still running and can communicate with each other but are unable to communicate with any service outside the zone and are not reachable by any other service outside the zone.
• Total zone failure. All instances in the zone are terminated.
Chaos Gorilla causes massive damage and requires a sophisticated control system to rebalance load. For Netflix, that system is still being developed, and as a result, Chaos Gorilla is run manually, similar to the GameDay exercises mentioned previously. With each successive run, Chaos Gorilla becomes more aggressive in the way it executes the
A region is made up of multiple data centers (availability zones) that are meant to be isolated from one another. A robust deployment architecture has AZ redundancy by using multiple AZs. In practice, regionwide failures do occur, which makes
Once Chaos Monkey is running and individual instance failure no longer has any impact, a new class of failures emerges. Dealing with instance failure is relatively easy: just terminate the bad instances and let new healthy instances take their places. Detecting when instances become unhealthy, but are still working, is harder, and having resilience to this failure mode is harder still.
Error rates could become elevated, but the service could occasionally return success. The service could reply with successful responses, but latency could increase, causing timeouts.
What Netflix needed was a way of inducing failure that simulated partially healthy instances. This was the genesis of Latency Monkey, which induces artificial delays in the RESTful
The rest of the Simian Army, including Janitor Monkey, takes care of upkeep and other miscellaneous tasks not directly related to availability. A full reference is available at http://techblog.netflix.
While the Simian Army is a novel concept and may require a shift in perspective, it isn't as hard to implement as it initially appears. Understanding what Netflix went through is illustrative for others interested in following such a path.
Netflix is known for being bold in its rapid pursuit of innovation and high availability, but not to the point of callousness. It is careful to avoid any noticeable impact to customers from these failure- induction exercises. To minimize risk, Netflix takes the following steps when introducing a monkey:
1. With the new monkey in the test environment, engineers observe the user experience. The goal is to have negligible or zero impact on the customer. If the engineers see any adverse results, then they make the necessary code changes to prevent recurrence. This step is repeated as many times as necessary until no adverse user experience is observed.
2. Once no adverse results are observed in the test environment, the new monkey is enabled in the production environment. Initially, the new monkey is run in
3. After many services have opted in, the new monkey graduates to
No discussion of resilience would be complete without highlighting the important role of monitoring. Monitoring here means the ability to observe and, optionally, signal an alarm on the external and internal states of the system and its components. In the context of failure induction and resilience, monitoring is important for two reasons:
• During a real, nonsimulated customer-impacting event, it's important to stabilize the system and eliminate customer impact as quickly as possible. Any automation that causes additional failure must be stopped during this time. Failing to do so can cause Chaos Monkey, Latency Monkey, and the other simians to further weaken an already unhealthy system, causing even greater adverse
• Building resilient systems doesn't happen at a single point in time; it's an ongoing process that involves discovering weaknesses and dealing with them in an iterative learning cycle. Deep visibility into the system is key to understanding how the system operates and how it fails. Few
One of the most important first questions to ask during a
Resilience to failure is a lofty goal. It enables a system to survive and withstand failure. There's an even higher peak to strive for, however: making the system stronger and better with each failure. In Nassim Taleb's parlance, it can become antifragile—growing stronger from each successive stressor, disturbance, and failure.8
Netflix has taken the following steps to create a more antifragile system and organization:
1. Every engineer is an operator of the service. This is sometimes referred to in jest as "no ops," though it's really more "distributed ops." Separating development and operations creates a division of responsibilities that can lead to a number of challenges, including network externalities and misaligned incentives. Network externalities are caused by operators feeling the pain of problems that developers introduce. Misaligned incentives are a result of operators wanting stability while developers want velocity. The DevOps movement was started in response to this divide. Instead of separating development and operations, developers should operate their own services. They deploy their code to production and then they are the ones awakened in the middle of the night if any part of it breaks and impacts customers. By combining development and operations, each engineer can respond to failure by altering the service to be more resilient to and tolerant of future failures.
2. Each failure is an opportunity to learn, generating these questions: "How could the failure have been detected more quickly?" "How can the system be more resilient to this type of failure?" "How can this failure be induced on a regular basis?" The result is that each failure makes the system more robust and resilient, analogous to the experience a warrior gains in each battle to make him stronger and fiercer in the next. The system becomes better the more times and ways it fails.
3. A blameless culture is fostered. As an organization, Netflix optimizes for innovation and velocity, and it accepts that mistakes will sometimes occur, using each one as an opportunity to learn. A commonly overheard saying at Netflix is, "If we're not making any mistakes, it means we're not moving quickly enough." Mistakes aren't a bad thing, unless the same mistakes are made over and over again. The result is that people are less worried about making mistakes, and postmortems can be structured as effective opportunities to learn (see step 2).
The more frequently failure occurs, the more prepared the system and organization become to deal with it in a transparent and predictable manner. Inducing failure is the best way of ensuring both system and organizational resilience. The goal is to maximize availability, insulating users of a service from failure and delivering a consistent and available user experience. Resilience can be improved by increasing the frequency and variety of failure and evolving the system to deal better with each newfound failure, thereby increasing antifragility. Focusing on learning and fostering a blameless culture are essential organizational elements in creating proper feedback in the system.
1. Robbins, J., Krishnan, K., Allspaw, J., Limoncelli, T. 2012. Resilience engineering: learning to embrace failure. Communications of the ACM55(11): 40-47; http://dx.doi.org/10.1145/2366316.2366331.
2. Bennett, C. 2012. Edda - Learn the stories of your cloud deployments. The Netflix Tech Blog; http://techblog.netflix.com/2012/11/edda-learn-stories-of-your-cloud.html.
3. Bennett, C., Tseitlin, A. 2012. Chaos Monkey released into the wild. The Netflix Tech Blog; http://techblog.netflix.com/2012/07/chaos-monkey-released-into-wild.html.
4. Chandra, T. D., Griesemer, R., Redstone, J. 2007. Paxos made live: an engineering perspective. In Proceedings of the 26th Annual ACM Symposium on Principles of Distributed Computing: 398-407; http://labs.google.com/papers/paxos_made_live.pdf.
5. Izrailevsky, Y., Tseitlin, A. 2011. The Netflix simian army. The Netflix Tech Blog; http://techblog.netflix.com/2011/07/netflix-simian-army.html.
6. Sondow, J. 2012. Asgard: Web-based cloud management and deployment. The Netflix Tech Blog; http://techblog.netflix.com/2012/06/asgard-web-based-cloud-management-and.html.
7. Strigini, L. 2009. Fault tolerance and resilience: meanings, measures and assessment. London, U.K.: Centre for Software Reliability, City University London; http://www.csr.city.ac.uk/projects/amber/resilienceFTmeasurementv06.pdf.
8. Taleb, N. 2012. Antifragile: Things That Gain from Disorder. Random House.
LOVE IT, HATE IT? LET US KNOW firstname.lastname@example.org
Ariel Tseitlin is director of cloud solutions at Netflix, where he manages the Netflix Cloud and is responsible for cloud tooling, monitoring, performance and scalability, and cloud operations and reliability engineering. He is also interested in resilience and highly available distributed systems. Prior to joining Netflix, he was most recently VP of technology and products at Sungevity and before that was the founder and CEO of CTOWorks.
© 2013 ACM
Originally published in Queue vol. 11, no. 6—
see this item in the ACM Digital Library
Robert V. Binder, Bruno Legeard, Anne Kramer - Model-based Testing: Where Does It Stand?
MBT has positive effects on efficiency and effectiveness, even if it only partially fulfills high expectations.
Terry Coatta, Michael Donat, Jafar Husain - Automated QA Testing at EA: Driven by Events
A discussion with Michael Donat, Jafar Husain, and Terry Coatta
James Roche - Adopting DevOps Practices in Quality Assurance
Merging the art and science of software development
Neil Mitchell - Leaking Space
Eliminating memory hogs
(newest first)There is a need for help from software testers to test the system in an extreme manner to induce failures, not just based on the methods that are taught in school.
Just as in a disease-control lab in healthcare, viruses and infections are resolved off-line then the immunization is applied post-fix, the "production-ready" elements provide that capability (and in a critical instance can be quickly swapped into the production running system to avoid catastrophe). The latter IT governance approach to have the second "back-up" system to switch to -- allows the system to keep running while components are removed from production to identify and take corrective action. Obviously, these are something IT professionals inherently understand and have been trained in. But, before one balks at what I've written, remember that the strategic drivers and decisions lie in the C-Level offices and are usually made by financial people who may, or may not have, a background that intrinsically understands the depth of these issues (until after the crises occur). So, my point -- you may need to explain importance of the "prepared approach" to proactively setting up the means to resolve these issues.
Also, the Armies are a great example of how to automate "negative QA" -- the ability to test that the system produce a predictable response to unforseen run time circumstances. One new simian that would be wonderful to meet would be the "Configuration Orangutan" -- an automated chaos monkey that misconfigures the system randomly and then runs it again non chaotic workloads. Often the human element is the most unpredictable and pernicious.
Great article. Thanks for codifying the Netflix approach so neatly.