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.

I wasn’t able to make it to SRECon17 Americas this year, but it sounds like it was a great time. (day two summary)

My heroine, Julia Evans, gave the plenary session at SRECon17 Americas, all about how to learn how to be an excellent engineer (or really anything!). She proved herself once again not just as an excellent student, but also an inspiring teacher. The best part is that she posted the abstract, slides, and a transcript of her talk shortly after giving it! This is a really excellent resource for folks like me that weren’t there, and I hope more talk-givers will follow her example.

This article is long, but I wish I’d carved out time for it long ago, because it’s really incredible and well worth the read. John Allspaw uses the SEC analysis of the Knight Capital incident as a starting point to introduce and discuss the problems with Counterfactual Thinking (“if the engineer had just done ___, this wouldn’t have happened”).

Rolling back a flawed code release can have significant risk. It doesn’t always fix the problem because the erroneous code may have had effects on other parts of the system. Sometimes, as in the Knight Capital incident, a rollback exacerbates the problem.

This is part two of an annotation of the google SRE book by Stephen Thorne, a Google SRE. Part Three is available too.

Here’s an interesting idea: using metadata about incidents as a proxy for measuring technical debt. PagerDuty goes over the definition of technical debt before diving into measuring it.

GitLab posted an update on “team-member-1”, the engineer that entered the commands that caused their production DB to be erased. I love that they posted this, because I for one was worried about “team-member-1” as a second victim.

During an incident, emotions can run strong. How can we set up incident response so as to provide the best environment for our responders?

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

I love the analogy: you can’t work around a slow drain with a bigger sink.

Stephen Thorne, a Google SRE, annotates the first chapter of the Google SRE book with his personal opinions and interpretations.

The author of this short article starts with the blooper during the Oscars and beautifully segues into a description of techniques organizations can use to halt the propagation of errors.

This webinar looks really interesting, and I’m going to try to see it. It’s about the importance of providing context to incident responders, how much to provide, and how to provide it.

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

When you do as the title suggests, you realize that network partitions go from the realm of theoretical to everyday.

Asana shares their “Five Whys” process, which they use not only for outages but even for missed deadlines. This caught my eye:

Our team confidently focuses on problem mitigation while fighting a fire, knowing that there will be time for post-mortem and long-term fixes later.

Using Route 53 as a case study, AWS engineers explain how they carefully designed their deploy process to reduce impact from failed deploys.

One method to reduce potential impact is to shape your deployment strategies around the failure conditions of your service. Thus, when a deployment fails, the service owner has more control over the blast radius as well as the scope of the impact.

GitHub used a data-driven approach when migrating a storage load from Redis to MySQL. It’s a good thing they did, because a straight one-for-one translation would have overloaded MySQL.

We’ve heard before that it’s important to make sure that your alerts are actionable. I like that this article goes into some detail on why we sometimes tend to create inactionable alerts before explaining how to improve your alerting.

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

Ubuntu backported a security fix into Xenial’s kernel last month, and unfortunately, they introduced a regression. Under certain circumstances, the kernel will give up way too easily when attempting to find memory to satisfy an allocation and will needlessly trigger the OOM killer. A fix was released on February 20th.

Need to tell someone their perpetual motion machine CAP-satisfying system won’t work? Low on time? Use this handy checklist to explain why their idea won’t work.

GitLab seriously considered fleeing the cloud for a datacenter, and they asked the community for feedback. That feedback was very useful and was enough to change their minds. The common theme: “you are not an infrastructure company, so why try to be one?”

If you’ve got a firehose of events going into your metrics/log aggregation system, you may need to reduce load on it by only sending in a portion of your events. Do you pick one out of every N? HoneyComb’s makers suggest an interesting alternative: tag each sampled event you send as representing N events from the source — and N is allowed to very between samples.