Today, I read a paper titled Lessons from Giant-Scale Services. This paper is written by Eric Brewer, the guy behind CAP theorem. It is an old paper published in 2001. The paper helps the reader build mental model on how to think about availability of large scale distributed systems.
Today, I read paper How Complex System Fail by Richard Cook. This paper was written in 2000. This is a short and easy to read paper written by a doctor. He writes about his learnings in the context of patient care, but most of them are applicable to software development as well.
In this post, I go over the points that I liked from the paper and write how I think they are applicable to software development.
The first point raised in the paper is
All complex systems are intrinsically hazardous systems.
With respect to complex software that most of us build, I don’t think hazardous will be the right word. Luckily, no one will die if the software we write fails. I see it as all complex systems are intrinsically important systems for an organisation. Most of the complex software has business and financial importance. So if the software fails, organization will end up loosing money and reputation. These can be catastrophic for the organization and might bring an organization to its knees. It is the business importance of these systems that drives the creation of defences against failure.
This takes us to the next point in the paper:
Complex systems are heavily and successfully defended against failure.
I don’t think from the day one most complex systems have successful defensive mechanisms placed. But, I agree that over time mechanisms will be added to successful operate the application. Few of the mechanisms that we use in software are data backup, redundancy, monitoring, periodic health checks, training people, creating change approval process, and many others. With time, an organization builds a series of shields that normally divert operations away from failures. Another key point relevant to software is that when we try to rebuild a system we should expect that the new system will become stable over time. Many times you deploy a new system and it fails because the mechanisms used in old system might not be sufficient to run the new system. Over time, an organization will put in place mechanisms to successfully operate the new application.
The next point mentioned in this paper is:
Catastrophe requires multiple failures – single point failures are not enough.
I disagree with author that single point of failures are not enough for catastrophic failures. I think it depends on the way the complex software is built. If your complex software systems is built in such a way that fault in one service brings cause failure in the service using the first one. Then, a single failure in the least important component can cause the entire application to break. So, if you are building complex distributed applications then you should use circuit breakers, service discovery, retries, etc to build resilient applications.
The next point raised by author is not obvious to most of us:
Complex systems contain changing mixtures of failures latent within them.
Author says that it is impossible for us to run complex systems without flaws. I have read a similar point somewhere that Amazon tries for five-nine availability i.e. 99.999%. Five-nines or 99.999% availability means 5 minutes, 15 seconds or less of downtime in a year. They don’t try to go further because it becomes too costly after this time and customers can’t notice the difference between their internet not working and service becoming unavailable.
A corollary of the above point is:
Complex systems run in degraded mode.
The system continues to function because it contains so many redundancies and because people can make it function, despite the presence of many flaws.
The next point in the paper is:
Catastrophe is always just around the corner.
Another way to put the above point is Failure is the norm. Deal with it. If you are building distributed systems then you must be clear in your mind that different components will fail. They might lead to complete failure of the system. So, we should try to build systems that can handle individual component failure. Author writes, it is impossible to eliminate the potential for such catastrophic failure; the potential for such failure is always present by the system’s own nature.
Then, author goes on to write:
Post-accident attribution accident to a ‘root cause’ is fundamentally wrong.
I don’t agree entirely with this point. The author is saying that there is no single cause for failure. According to him, there are multiple contributors to the failure. I agree that multiple reasons would have let to failure. But, I don’t think root cause analysis is fundamentally wrong. I think root cause analysis is the starting point to understand why things go wrong. In software world, blameless postmortem is done to understand failure with the intention of not finding which team or person is responsible but to learn from failure and avoid such failure in future.
The next point brings human biases into picture:
Hindsight biases post-accident assessments of human performance.
It always look obvious in the hindsight that we should have avoided the failure given all the events that happened. This is the most difficult point to follow in practice. I don’t know how we can overcome biases. Author writes, Hindsight bias remains the primary obstacle to accident investigation, especially when expert human performance is involved.
The next point mentioned in the paper is:
All practitioner actions are gambles.
This is an interesting point. If you have ever tried to bring a failure system back to state, you will agree with the author. In those uncertain times, we try different things to bring the system up but at best they are our guesses based on our previous learnings. Author writes, practitioner actions are gambles appears clear after accidents; in general, post hoc analysis regards these gambles as poor ones. But the converse: that successful outcomes are also the result of gambles; is not widely appreciated.
I like the way author makes a point about human expertise:
Human expertise in complex systems is constantly changing.
I think most software organisations want to use the latest and greatest technology to build complex systems but they don’t invent enough in developing expertise. Humans are the most critical part of any complex system so we need to plan for their training skill refinement as a function of system itself. Failure to do this will lead to software failures.
As they say, change is the only constant. The next point talks:
Change introduces new forms of failure.
I will quote directly from the paper as author puts it eloquently
The low rate of overt accidents in reliable systems may encourage changes, especially the use of new technology, to decrease the number of low consequence but high frequency failures. These changes maybe actually create opportunities for new, low frequency but high consequence failures. When new technologies are used to eliminate well understood system failures or to gain high precision performance they often introduce new pathways to large scale, catastrophic failures.
The last point mentioned in the paper sums it all:
Failure free operations require experience with failure.
As they say, failure is the world greater teacher. When you encounter performance issues or difficult to predict situations then we push the envelope of our understanding of the system. These are the situations that help us discover new mechanisms that we need to put in our system to handle future failures.
I thoroughly enjoyed reading this paper. It has many learnings for software developers. If we become conscious about them I think we can write better resilient software.
Today evening, I decided to read paper Simple Testing Can Prevent Most Failures: An analysis of Production Failures in Distributed Data-intensive systems.
This paper asks an important question
Why widely used distributed data systems designed for high availability like Cassandra, Redis, Hadoop, HBase, and HDFS. experience failures and what can be done to increase their resiliency?
We have to answer this question keeping in mind that these systems are developed by some of the best software developers in the world following good software development practices and are intensely tested.
These days most of us are building distributed systems. We can apply the findings shared in this post to build systems that are more resilient to failure.
The paper shares:
Most of the catastrophic system failures are result of incorrect handling of non-fatal errors explicitly signalled in the software
This falls into 1) empty error handling blocks, or error blocks with just log statement 2) the error handing aborts the clusters on an overly-general exception 3) the error handling code contains expressions like “FIXME” or “TODO” in the comments.
Most of the developers are guilty of doing all the three above mentioned. Developers are good at finding that something will go wrong but they don’t know what to do when something goes wrong. I looked at the error handling code in one of my projects and I found the same behaviour. I had written TODO comments or caught general exceptions. These are considered to be bad practices but still most of us end up doing.
Overall, we found that the developers are good at anticipating possible errors. In all but one case, the errors were checked by the developers. The only case where developers did not check the error was an unchecked error system call return in Redis.
Another important point mentioned in the paper is
We found that 74% of the failures are deterministic in that they are guaranteed to manifest with an appropriate input sequence, that almost all failures are guaranteed to manifest on no more than three nodes, and that 77% of the failures can be reproduced by a unit test.
Most popular open source projects use unit testing so it could be surprising that the existing tests were not good enough to catch these bugs. Part of this has to do with the fact that these bugs or failure situations happens when a sequence of events happen. The good part is that sequence is deterministic. As a software developer, I could relate to the fact that most of us are not good at thinking through all the permutation and combinations. So, even though we write unit tests they do not cover all scenarios. I think code coverage tools and mutation testing can help here.
It is now universally agreed that unit testing helps reduce bugs in software. Last few years, I have worked with few big enterprises and I can attest most of their code didn’t had unit tests and even if parts of the code had unit tests those tests were useless. So, even though open source projects that we use are getting better through unit testing most of the code that an average developer writes has a long way to go. One thing that we can learn from this paper is to start write high quality tests.
The paper mentions specific events where most of the bugs happen. Some of these events are:
- Starting up services
- Unreachable nodes
- Configuration changes
- Adding a node
If you are building distributed application, then you can try to test your application for these events. If you are building applications that uses Microservices based architecture then these are interesting events for your application as well. For example, if you call a service that is not available how your system behaves.
As per the paper, these mature open-source systems has mature logging.
76% of the failures print explicit failure related messages.
Paper mentions three reasons why that is the case:
- First, since distributed systems are more complex, and harder to debug, developers likely pay more attention to logging.
- Second, the horizontal scalability of these systems makes the performance overhead of outputing log message less critical.
- Third, communicating through message-passing provides natural points to log messages; for example, if two nodes cannot communicate with each other because of a network problem, both have the opportunity to log the error.
Authors of the paper built a static analysis tool called Aspirator for locating these bug patterns.
If Aspirator had been used and the captured bugs fixed, 33% of the Cassandra, HBase, HDFS, and MapReduce’s catastrophic failures we studied could have been prevented.
Overall, I enjoyed reading this paper. I found it easy to read and applicable to all software developers.