Talks

Summarized using AI

Too Big to Fail

Chris Maddox • April 22, 2014 • Chicago, IL

The video "Too Big to Fail" presented by Chris Maddox at RailsConf 2014 discusses strategies for managing failure in software processes, specifically within a payroll system at ZenPayroll. Maddox emphasizes the importance of handling failure, predicting issues, and recovering from failures in a complex, high-stakes environment where user data is critical and where financial transactions are substantial. The presentation focuses on the following key areas:

  • Understanding Failure: Maddox highlights the inevitability of failure in software systems and stresses that instead of trying to become fault-tolerant, teams should focus on how they respond to failures when they happen.
  • Predicting and Avoiding Failure: The talk discusses how to proactively prevent errors through mechanisms like database validations. Running validations for every model in the database every few hours helped identify potential issues before they escalated into user-facing problems.
  • Embracing Failure: Accepting that failures occur and using them as learning opportunities is a key theme. Maddox argues that mistakes in development can lead to improvements and growth in understanding how systems operate.
  • Recovery Mechanisms: The presentation describes the development of a library named "Ultramarathon" to streamline running long processes while accounting for potential failures and enabling easier recovery. The approach includes breaking down tasks and managing state effectively, thereby preventing single points of failure during complex operations.
  • Philosophical Takeaway: Maddox concludes that adopting a mindset that tolerates failure can enhance team morale and system robustness. The approach involves prioritizing user experience, safeguarding sensitive user information, and maintaining operational integrity even when errors occur.

Overall, the presentation serves as a call to action for software teams to rethink their strategies about failure, illustrating that rather than fearing errors, recognizing and learning from them is essential for sustainable growth in software development.

Too Big to Fail
Chris Maddox • April 22, 2014 • Chicago, IL

It's 5am and a multi-million dollar process fails halfway through. Hours of nightmarish, manual brain surgery later, enough is enough.

What happens when background jobs grow as bloated as the MonoRail™ that begot them?

Rather than reach for the latest fad off of HackerNews, we'll user Ruby and Rails to automate error-recovery, concurrent processing, and catch corrupt data before it brings everything down.

Typist, Philosopher at ZenPayroll. Humanist with a penchant for dystopian novels, St. George gin enthusiast, and wearer of colorful pants.

Help us caption & translate this video!

http://amara.org/v/FG1u/

RailsConf 2014

00:00:16.560 Um, let’s see. My name is Chris Maddox, and I work at ZenPayroll, where we handle payroll.
00:00:22.720 Failure is something that we think about a lot, especially since we transfer hundreds of millions of dollars every year, soon to be billions.
00:00:28.960 Our growth team gets really excited about this, saying, 'Oh, billions of dollars!' But honestly, I just want to focus on doing my job well.
00:00:34.480 Today, we're going to talk about failure and some ways we've preemptively addressed it. We have an agenda like a business meeting, which I think is great.
00:00:45.760 We’ll discuss predicting failures, approaches we’ve implemented to avoid mistakes, particularly regarding database validations and user models. Then, we’ll cover recovering from failure, which is critical for us.
00:00:57.280 While preventing failure is important, making sure that if something goes wrong, we don’t lose sensitive user information is our top priority.
00:01:14.240 Finally, we’ll talk about embracing failure, which might sound odd, but I'm a philosophy major—so I'm excited to dive into this topic.
00:01:19.759 Chapter one: changes. DHH, I’m a writer now, so we’re organizing this into chapters. My predictions, however, have been notably inaccurate.
00:01:37.439 Robert Caro, a biographer, has been writing a four-part series on Lyndon B. Johnson for 30 years. I’m only 23, so I can’t fathom that length of time, but I wish him the best.
00:01:56.479 I’ve encountered issues on the internet several times where errors aren't communicated well.
00:02:02.719 For example, you might receive a message saying, 'There were some problems with your form; please fix the errors before submitting again,' yet the system fails to show what’s actually wrong.
00:02:20.959 I'm certain many of you have faced situations where you try to submit your information, only to be confronted with vague error messages that lead to frustration.
00:02:38.080 In the case of a payroll platform, incorrect validations can lead users to believe they’re on track, while the issues lie elsewhere.
00:02:43.120 Unfortunately, this leads to confusion and frustration, not just for the users but also for the support team who may be unable to provide clear answers.
00:02:50.400 This issue worsens as your application grows in complexity. Payroll is intricate, with a lot of information to process.
00:03:02.959 For instance, I mistakenly misconfigured Mr. Vonnegut’s bank account, which means he won’t get paid the way he expects.
00:03:28.159 Yet, the user is unaware of this mistake because they are editing a different section. They will face validations that don’t make sense to them.
00:03:39.599 This ultimately frustrates users when they need support, leading to a snowballing effect of problems.
00:03:52.080 As a result, users are prevented from fixing or even perceiving the real issue, which can lead to cascading errors across processes.
00:04:09.920 Two main culprits behind these failures include using update_column and update_attributes with validate: false, which, while often necessary, can be detrimental.
00:04:51.120 When certain situations arise, like needing to save information despite an error, these functions can be useful. However, they can also result in losing sensitive information.
00:05:03.600 Preserving user information is critical for us, especially since losses could amount to tens of millions of dollars per single transaction. Getting this right is absolutely essential.
00:05:25.440 From this, we’ve devised solutions to validate data upfront rather than relying on user input alone—this is achieved through Rails validations.
00:05:31.760 I learned from a colleague, Nick Drivassi, that we could run validations for every model in the database every few hours. Initially, I thought it was a terrible idea.
00:05:53.440 However, it turned out to be a game-changer, and we now validate every model in the database every three hours.
00:06:00.880 This entails pulling production data to validate it against our merged but undeployed code, allowing us to catch future errors before they become problems for users.
00:06:35.520 This proactive approach has been incredibly beneficial, providing us with the lead time necessary to address issues more calmly, rather than in the moment when users often feel frantic.
00:07:11.680 Shifting gears a bit—this concept reminds me of the pre-cogs from the movie 'Minority Report.' In the film, these genetically modified individuals predict murders.
00:07:38.560 While their ability to forecast rather grim events may raise ethical questions, it provokes thought about how we can foresee errors in our own systems.
00:08:02.159 We’ve begun to create long-running jobs to understand the eventualities of errors in our applications, especially with complex transactions that involve multiple parties like the IRS.
00:08:21.680 Managing financial transactions can be a convoluted process, especially considering the communication barrier with banks after a transaction has been initiated.
00:08:50.560 After the transaction begins, responses regarding success are limited. You only know it’s complete if there isn't a failure notice days later.
00:09:16.960 This means we must ensure that everything within our control operates smoothly prior to going live with any updates.
00:09:39.760 Now let’s pivot to recovery from failure, which is actually a fascinating topic.
00:10:02.080 Predictions about errors are interesting and beneficial, yet how we recover once an event occurs is something I believe requires more focus.
00:10:25.679 Every other day in practice, I see that what makes you better is how you respond to failure, including ones caused by you!
00:10:50.800 Mia Hamm once shared that failure happens to everyone, and it’s how you address it afterward that defines growth.
00:11:12.640 Reflecting on the inception of our Rails app, it’s inspiring to remember the early excitement while also recognizing the intricacies involved in payroll systems.
00:11:43.440 Managing a system of maintenance was overwhelming, with one very large method running for a long time, which sometimes led to unexpected errors.
00:12:05.520 When the state becomes uncertain due to an error, regaining stability is challenging and often feels reminiscent of brain surgery.
00:12:34.960 Many underlying dependencies can cause issues, leading us to confront situations we never anticipated with banks, the IRS, and various other integrations.
00:12:51.120 As we developed our API-driven services, it seemed easy to demystify responsibility on components, but once they went wrong, chaos ensued.
00:13:07.360 One particular pain point with the IRS came when the government shut down unexpectedly, leading to delayed faxes—an edge-case we hadn't accounted for.
00:13:23.760 The struggles surrounding the possible failures we face continues to highlight the complexities present in our operations.
00:13:43.760 To reduce the likelihood of complete system failure, we began dissecting existing processes into smaller, manageable pieces.
00:14:01.120 By localizing issues, we can continue payments for users while fixing others, like syncing with external APIs.
00:14:15.960 Our approach led to several new libraries we developed, like Ultramarathon, which helped understand connections and dependencies more clearly.
00:14:35.360 The philosophy guiding various processes is about minimizing interruptions and focusing on granularity, where possible.
00:14:49.920 We also incorporated infrastructure that allows us to recover from partial failures instead of completely snowballing an operation.
00:15:05.920 Now, even if a job fails, we’re able to run other pieces without hassle or halting everything.
00:15:18.080 Additionally, we implemented detailed instrumentation to better understand where issues arise.
00:15:28.560 All this contributes to a more resilient system designed not just to handle growth but to accommodate errors as part of the lifecycle.
00:16:03.440 Accepting that failure is a natural process in software development is critical, and I learned this lesson from my dad.
00:16:33.600 Recognizing that we handle serious matters, the stakes can lead to unnecessary panic; people rely on us to manage their payroll effectively.
00:17:09.440 While we started with a philosophy where failure is unacceptable, it became evident that such pressure isn’t conducive to a healthy work environment.
00:17:38.880 Ultimately, I believe it’s our responsibility to not only avoid failure but also to mitigate its impacts when it does occur.
00:18:06.880 Reflecting on experiences, I feel it’s crucial for our growth and learning to accept failure as part of our evolution.
00:18:43.440 The philosophy to embrace failure allows us to nurture resilience and a culture of learning within the team.
00:19:15.840 If there’s one takeaway today, it’s to recognize that failure is a part of the journey—something we can learn from, rather than fear.
00:19:42.560 So, if your company is sponsoring you for this talk, take that advice with caution!
Explore all talks recorded at RailsConf 2014
+129