In February of 2016, a metal hospital gurney was inadvertently wheeled* into an MRI room, resulting in a costly near-miss accident. Brigham and Women’s Hospital posted about the mishap on their Safety Matters blog and also released a Q&A with their chief quality officer about their dedication to an open and just culture.

If an employee at Brigham makes a mistake that anyone else could make, we will work on improving the system, rather than punishing the employee. We believe that in every circumstance involving “human error” there are systemic opportunities for mitigating reoccurrence.

* Yes, I used the passive voice on purpose. See what I did there?

Sometimes logs help us prevent outages or discover a cause. But raise your hand if you’ve seen logging cause an outage. Yeah, me too.

Traditionally, auditd, together with Linux’s system call auditing support, has been used as part of security monitoring. Slack developed go-audit so that they could use system call auditing as a general monitoring tool. I can think of plenty of outages during which I’d have loved to be able to query system call patterns!

Dropbox has some pretty complex needs around feature gating. Apparently existing tools couldn’t satisfy their use case so they wrote and released a tool with sophisticated user segmentation support.

Why depend on fallible QA testing to spot regressions in a web UI? Computers are so much better at that kind of thing. Niffy spots the pixel changes between old and new code so you can see exactly what regressed before putting it in front of your users.

In this beautifully-illustrated article, Stripe engineer Jacqueline Xu explains how Stripe safely rolled out a major database schema upgrade. Many code paths had to be refactored, and they took a methodical, incremental approach to avoid downtime. Thanks to this article, I now know about Scientist and can’t wait to use it.

Speaking of Stripe, they have another polished post on how and why to add load shedding to your API.

Scientist is such an awesome idea. The idea is to try out a new code path and see if it produces the same result as the old code path. It only returns the new code path, so you know you can safely prove to yourself whether the new code path is safe before exposing users to it.

I’m including this article at least in part due to its mention of the February S3 outage. AWS had difficulty reporting the outage on its status site because of a dependency on S3.

Conway’s Law is extremely important to us as SREs. As we can see in the example of Sprouter, a poor organizational structure can produce unreliable software. My fellow SRE, Courtney Eckhardt, loves to say, “My job is applying Conway’s Law in reverse.”

The big story this week is the release of the inaugural issue of Increment, a newsletter by Stripe, edited by Susan Fowler. They bill it as “A digital magazine about how teams build and operate software systems at scale” and the first issue, dedicated to on-call, certainly delivers. Below, I’ll include my short take on each article in the issue.

Increment interviewed over thirty companies to build a picture of the common practices in incident response. I’m actually pretty surprised to hear that “it turns out that they all follow similar (if not completely identical) incident response processes”, but apparently the commonalities don’t stop at just process:

Slack and PagerDuty appear to be two points of failure across the entire tech industry

Bonus content: Julia Evans shared her notes on Twitter.

Next up, Increment addresses the dichotomy of ops teams versus developers on call for their code. It turns out that the latter practice is more prevalent than I’d realized.

After laying a solid groundwork of suggestions for avoiding burn-out in on-call, this next Increment article raises a really important point: on-call affects people differently based on privilege. Example: single parents are going to have a much harder time of it.

[…] if you set up an on-call rotation with a schedule or intensity that assumes the participants have no real responsibilities outside of the office, you are limiting the people who will be able to participate on your team.

Remember a couple of months back when GitLab live-streamed their incident response? Increment caught up with their CEO to give us this in-depth interview about their radical transparency.

Increment shares tips and key practices for setting up on-call, targeted to companies of size ranges varying from 0-10 employees all the way up to 10000+.

Increment rounds out their issue with advice in the form of quotes from six of the companies they interviewed.

The other big news of the week is the official launch of Honeycomb.io. If you haven’t had a chance to check it out yet, here’s an introduction, and you can also sign up for a free one-month trial.

This article is about the risks of automation. While automation can reduce risk by making errors less likely, it also disengages human operators from what’s actually happening, meaning that they’re less likely to catch and correct problems.

The author spent seven months sifting through, categorizing, and documenting over 1700 production incidents. The result was impressive: a massive improvement in the SRE team’s incident response process and documentation. It’s got me wondering if we can do something similar at $JOB.

Thanks to Steven Farlie for this one.

A danger of a microservice architecture is that one failing service can affect those that depend on it, even indirectly. The Netflix API handles over 10000 requests per second, and it was carefully designed to avoid the case where a slow dependency breaks unrelated requests.

Without taking steps to ensure fault tolerance, 30 dependencies each with 99.99% uptime would result in 2+ hours downtime/month (99.99% * 30 = 99.7% uptime = 2+ hours in a month).

Nuclear Family is an interactive play in which the audience is presented with critical decisions as the characters move inexorably toward a nuclear plant disaster. The goal is to demonstrate local rationality, the principal that people make the best decision they can with the information they have at hand — even if in retrospect that decision led to an adverse outcome.

Last year, PagerDuty moved toward giving developers operational responsibilty for the systems they create. The really cool thing about their transition is that they have hard stats on reduction of incidents, decrease in MTTR, and increase in changes deployed to production.

This post is primarily a new feature announcement, but the intro section is just awesome. I love the idea of designing a system with empathy for your future self that will be on call for it.

A short but enlightening blog post on designing systems to degrade gracefully.

when weird stuff happens, make sure it doesn’t cause harm you didn’t expect or plan for.

I hope you’ll enjoy reading this debug session as much as I enjoyed co-writing it. My former co-worker and I did some serious digging to get to the bottom of an unexpected EADDRINUSE that caused a production incident.

Full disclosure: Heroku is my employer.

Distributed filesystems provide high availability by duplicating data. In this research paper, the researchers created errorfs, a FUSE plugin that passes through a backing filesystem but introduces single-bit errors. Result: almost all major distributed filesystems missed the error, resulting in corruption.

The part I like most about this article is the emphasis on the difference between DR and HA.

Full disclosure: Heroku, my employer, is mentioned.

The S3 outage a month ago is a great reminder that chaos experiments are useful not just for taking down parts of our own infrastructure, but also simulating the failure of external dependencies.

There are several core HumanOps principles, but the most important one to remember is that human health impacts business health.

It’s about time that we recognised that engineers are humans who get stressed and need downtime and that there are strong business as well as social reasons why these needs should be met.

Impressively quickly, USENIX has posted the videos from SRECon17 Americas! I’ve linked to a post by Woodland Hunter, whose review of SRECon I featured here two weeks ago, with links to the talks he reviewed and more.

• MTTR ZERO: one weird trick solves all the problems – VictorOps
• PagerDuty is proud to announce the launch of our “OnCallClub” program

The first article is published by my sponsor, VictorOps, but their sponsorship did not influence its inclusion in this issue.

PagerDuty theorizes that if developers don’t trust the incident response process, they’ll fear outages and thus be less productive. Proper incident management eases that fear so that they feel safer deploying code.

This article could be titled, “Use these three wacky tricks to reduce your pages by 100x!” In all seriousness, the methods described are aggregation (group related alerts), routing (sort alerts by team), and classification (page-worthy alerts versus warnings).

This article is published by my sponsor, VictorOps, but their sponsorship did not influence its inclusion in this issue.

A nice primer on using tc to induce latency, which is really important when testing the resiliency of systems to network instability. Thanks, Julia!

Here’s the second half of Stephen Thorne’s commentary on “Embracing Risk”, the third chapter in Google’s SRE book.

As your company grows in infrastructure size, number of employees, load, and other areas, how do you change your incident response to cope?

Look, a new newsletter about monitoring! I’m really excited to see what they have to offer.

And another new newsletter! Like Monitoring Weekly, I’m betting this one will have a lot of articles of interest to SREs.

VictorOps held a webinar last Thursday to present and discuss the concept of context in incident management. Just paging in a responder isn’t enough: we need to get them up to speed on the incident as soon as possible. Ideally, the page itself would include snapshots of relevant graphs, links to playbooks, etc. But if we’re not careful and add too much information, the responder is overloaded by a “sandstorm” of irrelevant data. “time to learn” — post incident learning careful of info overload in presenting context with pages

This webinar was created by my sponsor, VictorOps, but their sponsorship did not influence its inclusion in this issue.

Here’s the next in Stephen Thorne’s series of commentary on chapters of the SRE book. I like that Google makes an effort not to be too reliable for fear of setting expectations too high, and they’re also realistic in their availability goals: no end-user will notice a 20-second outage.

Writing an API, system, server or really anything people might make use of? Don’t make the default timeout be infinite.

PagerDuty really has been churning out excellent articles in the past couple of weeks. [Spoiler Alert] The five things are: internal communication, monitoring, a public status site, a support ticket system, and a defined incident response procedure.

Keep on rockin’ it, PagerDuty. This time they identify common problems that hinder incident response and give suggestions on how to fix them.

The author reviews their experience at SRECon17 Americas, including interesting bits on Julia Evans, training, recruiting, and diversity.

I love that the ideas we’re talking about regarding human error apply even to commercial cannabis growing.

Sadly, little is known about the nature of these errors, mainly because our quest for the truth ends where it should begin, once we know it was a human error or is “someone’s fault.”

The newer and shinier the technology, the bigger the production risk.

In other words, software that has been around for a decade is well understood and has fewer unknowns.