The interesting bit in this story is that upgrading to 5.7 requires a full table rewrite (<tt>ALTER TABLE</tt>) for any table that has time-related columns. Their initial test-run took months and still hadn’t finished.

AdStage made the move from Heroku to running their service directly on EC2, and in this article they explain why and how.

We were officially only getting about 2 ECUs per dyno, but the reality was that we were getting something closer to 6 since our neighbors on Heroku were not using their full share. This meant that our fleet of AWS instances was 3 times too small, […]

Language Warning: contains the word “sexy” used to describe new or interesting technology.

Full disclosure: Heroku, my employer, is mentioned.

I’ve featured many articles from Mathias Lafeldt as part of his series, Production Ready. Now that he’s moved to Gremlin Inc (a SaaS helping customers run chaos experiments), Mathias reintroduces the history and theory of Chaos Engineering.

The folks behind Mail.ru implemented their own master-master replication system on top of Tarantool, a DBMS I’d never heard of. Their implementation is based on some details of their use-case that may not apply more broadly, but the design discussion is interesting nonetheless.

Facebook rewrote their tool, OnlineSchemaChange in Python (from the original PHP). OSC is a tool for doing DDL in MySQL without downtime.

The original open sourced OSC was more like an engine than a tool. Users needed to write PHP code wrapping to run the schema change, and, with PHP becoming less popular in the operations world, OSC.php wasn’t widely adopted by the community.

From PagerDuty, an article on the incident management data to gather, how to gather it, and how to analyze it.

A basic introduction to structured logging, including rationale on why you’d want to use it. With infrastructures growing more and more complicated, I find structured logging indispensable in keeping everything up and running and debugging difficult problems.

For the network nerds, Facebook details their new inter-datacenter network topology.

New in the latest version of Elastic Stack (think ElasticSearch, Logstash, Kibana, etc) is built-in anomaly detection using machine learning, based on technology from Prelert (acquired by Elastic in 2016). “Machine Learning” — they might as well say it’s powered by “Lasers™”. If you try this out and have any success, please write up your results and send me a link!

GitHub has released OctoDNS, their tool for synchronizing DNS across multiple providers. Shortly after the Dyn outage last fall (covered here), they still only had one DNS provider (source: direct observation). I suspected that this may have had to do with complication in keeping records synched across two providers – perhaps that’s why they created OctoDNS.

Bolt is Netflix’s “event driven diagnostic and remediation platform”, although it actually seems like a full-blown remote execution system for large fleets of servers.

A Google SRE takes us through their first on-call shift including running incident command for a production incident. I like the emphasis on a blameless postmortem.

Pete Shima received some questions about onboarding SREs, and lucky us, he decided to answer them publicly. My favorite section is the one about connecting a new SRE to people across the company. I find that solid connections to folks in various positions are vital to getting my job done well. Thanks to Pete for the SRE Weekly mention!

Salesforce has a humongous infrastructure, and they needed a tool to help visualize data from lots of monitoring systems. They created Refocus to serve that need, and they open sourced it. They had three goals: gather data from all of the monitoring systems, on-board new services quickly, and visualize data in a way that makes sense for each service.

Full disclosure: Salesforce (parent company of my employer, Heroku), is mentioned.

Tcpdump is a critical tool for debugging thorny network issues. Julia Evans created a new zine to help you learn the basics, although if her other zines are any indication, even a pro may learn a new trick or two. The zine is $10 now and will be available for free at some point in the future.

Turns out that sharks are a reliability risk. And not just those WFLB.

From their Global Developer Survey, GitLab learned that it’s common for developers to release code before it’s production-ready in response to organizational pressures.

Code released before it’s ready might be good for meeting deadlines, but that’s about all it’s good for.

Here’s a pretty excellent analysis of why adopting the cloud can be difficult for banks. Just skip past the bit with the incorrect uptime calculation, since four nines of uptime actually equates to about 53 minutes’ downtime per year, not 9 hours.

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.