Another issue of BWH’s Safety Matters, this time about a prescribing accident. The system seems to have been almost designed to cause this kind of error, so it’s good to hear that a fix was already about to be deployed.

This is a great article on identifying the true root cause(s) of an incident, as opposed to stopping with just a “direct cause”. I only wish it were titled, Use These Five Weird Tricks to Run Your RCA!

Etsy describes how they do change management during the holidays:

[…] at Etsy, we still push code during the holidays, just more carefully, it’s not a true code freeze, but a cold, melty mixture of water and ice. Hence, Slush.

This issue of Production Ready is about battling code rot through incrementally refactoring an area of a codebase while you’re doing development work that touches it.

Shutterstock shares some tips they’ve learned from writing postmortems. My favorite part is about recording a timeline of events in an incident. I’ve found that reading an entire chat transcript for an incident can be tedious, so it can be useful to tag items of interest using a chat-bot command or a search keyword so that you can easily find them later.

The “Outage!” AMA happened while I was on vacation, and I still haven’t had a chance to listen to it. Here’s a link in case you’d like to.

My favorite:

If something breaks in production, how do you know about it?

Weaver is a tool to help you identify problems in your microservice consumers by doing “bad” things like responding slowly to a fraction of requests.

Barclays reduced load on their mainframe by adding MongoDB as a caching layer to handle read requests. What the heck does “mainframe” mean in this decade, anyway?

We’d do well to remember during this holiday season that several seconds of latency in web requests is tantamount to an outage.

Tyler Treat gives us an eloquently presented argument for avoiding writing code as much as possible, for the sake of stability.

A detailed description of Disaster Recovery as a Service (DRaaS), including a discussion of the cost versus creating a DR site oneself. This is the part I always wonder about:

However, for larger enterprises with complex infrastructures and larger data volumes spread across disparate systems, DRaaS has often been too complicated and expensive to implement.

This one’s so short I can almost quote the whole thing here. I love its succinctness:

Regardless of what we discover, we understand and truly believe that everyone did the best job they could, given what they knew at the time, their skills and abilities, the resources available, and the situation at hand.

Just over four years ago, Amazon had a major outage in Elastic Block Store (EBS). Did you see impact? I sure did. Here’s Netflix’s account of how they survived the outage mostly unscathed.

I’m glad to see more people writing that Serverless != #NoOps. This article is well-argued even though it turns into an OnPage ad 3 paragraphs from the end.

What else can we expect from Greater Than Code + Charity Majors? This podcast is 50 minutes of awesome, and there’s a transcription, too! Listen/read for awesome phrases like “stamping out chaos”, find out why Charity says, “I personally hate [the term ‘SRE’] (but I hate a lot of things)”, and hear Conway’s law applied to microservices, #NoOps debunking, and a poignant ending about misogyny and equality.

Microsoft released its Route 53 competitor in late September. They say:

Azure DNS has the scale and redundancy built-in to ensure high availability for your domains. As a global service, Azure DNS is resilient to multiple Azure region failures and network partitioning for both its control plane and DNS serving plane.

This issue of BWH Safety Matters details an incident in which a communication issue between teams that don’t normally work together resulted in a patient injury. This is exactly the kind of pitfall that becomes more prevalent with the move toward microservices, as siloed teams sometimes come into contact only during an incident.

A detailed postmortem from an outage last month. Lots of takeaways, including one that kept coming up: test your emergency tooling before you need to use it.

Next year, SRECon is expanding to three events: Americas, EMEA, and Asia. The Americas event is also moving from Santa Clara to San Francisco, which I, for one, am especially grateful for. The CFP for SRECon17 Americas just opened up, and proposals are due November 30th, so get cracking! I can’t wait to see what all of you have to share!

I have a somewhat dim view of automated anomaly detection in metrics based on my experience with a few tools, but if Datadog’s algorithms live up to their description, they might really have something worthwhile.

When a responder gets an anomaly alert, he or she needs to know exactly why the alert triggered. The monitor status page for anomaly alerts shows what the metric in question looked like over the alert’s evaluation window, overlaid with the algorithm’s predicted range for that metric.

This issue of Production Ready chronicles Mathias Lafeldt’s effort to create a staging environment. I like the emphasis on using an entirely separate AWS account for staging. This is increasingly becoming a best practice.

What’s causing all that API request latency? Here’s an interesting debug run using Honeycomb. Negative HTTP status codes? Sure, that’s totally a thing, right?

I love this idea: Susan Fowler notes that large, complex systems are constantly changing, and this makes reproducing bugs difficult or impossible. Her suggestion is to log enough that you can logically reconstruct the state of the system at the time the bug occurred. This is the same kind of thing the Honeycomb folks are saying: throw a lot of information into your logs, just in case you might need it to debug something unforeseen.

I’ve linked to several posts on Etsy’s Code as Craft blog in the past, and here’s another great one. Perhaps not the typical SRE article you might have been expecting me to link to, but this stuff is important in every tech field, including SRE. We can’t succeed unless every one of us has a fair chance at success.

In 2014, CodeSpaces suffered a major security breach that abruptly ended their business. I’d say that’s a pretty serious reliability risk right there, showing that security and reliability are inextricably intertwined.

Check it out! Catchpoint is doing another Ask Me Anything, this time about incident response. Should be interesting!

My fellow Heroku SRE, Courtney Eckhardt, expanded on a section of our joint SRECon talk for this session at OSFeels. She had time for Q&A, and there were some really great questions!

Mathias rocks it, as usual, in this latest issue of Production Ready.

Netflix has released a new version of Chaos Monkey with some interesting new features.

Scalyr worked with Mathias Lafeldt to turn his already-awesome pair of articles into this excellent essay. He brings in real-world examples of major outages and draws conclusions based on them. He also hits on a number of other topics he’s written about previously. Great work, folks!

How many times have you pushed out a puppet change that you tested very thoroughly, only to find that it did something unexpected on a host you didn’t even think it would apply to? Etsy’s solution to that is this tool that shows catalog changes for all host types in your fleet in a diff-style format.

“We ended up 50 percent over our worst case after day one, we figured this was going to be bad within six hours”

It’s pretty impressive to me that Niantic managed to keep Pokemon Go afloat as well as they did. They worked very closely with Google to grow their infrastructure much faster than they had planned to.

As Spotify has grown to 1400 microservices and 10,000 servers, they’ve moved toward a total ownership model, in which development teams are responsible for their code in production.

Pingdom suffered a major outage during the Dyn DDoS, not only due to their own DNS-related issues, but also due to the massive number of alerts their system was trying to send out to notify customers that their services were failing.

[…] at 19:20 we went to DEFCON 1 and suspended the alerting service altogether.

Here’s Dyn’s write-up of the DDoS.

As they promised, here’s PagerDuty’s root cause analysis from the Dyn DDoS.

This is a pretty great idea. Ably has written their client libraries to reach their service through a secondary domain if the primary one is having DNS issues. Interestingly, their domain, ably.io, was also impacted by the .io TLD DNS outage (detailed below) just days after they wrote this.

StatusPage.io gives us some really interesting numbers around the Dyn DDoS, based on status posts made by their customers.  I wonder, was it really the largest in history?

Here’s a nice write-up of day one of ServerlessConf, with the theme, “NoOps isn’t a thing”.

Not a magic bullet, but still pretty interesting.

[…] it allows you to incrementally replicate live Virtual Machines (VMs) to the cloud without the need for a prolonged maintenance period. You can automate, schedule, and track incremental replication of your live server volumes, simplifying the process of coordinating and implementing large-scale migrations that span tens or hundreds of volumes.

Earlier this year, Australia’s online census site suffered a major outage. Here’s a little more detail into what went wrong. TL;DR: a router dropped its configuration on reboot.

Gov.uk has put in place a lot of best practices in incident response and on-call health.

After extensive rationalisation, GOV.UK have reached a stage where only 6 types of incidents can alert (wake them up) out of hours. The rest can wait until next morning.

Unfortunately, I’m guessing one of those six types happened this week, as you can see in the Outages section below.