There is a lot of discussion lately about how SRE fits into or competes with or whatever-s with DevOps. I’m scheduled to speak on a “SRE vs DevOps Smackdown” panel today here at Innotech Austin, and at the exact same time I see Bridget tweeting Liz Fong-Jones’ slides from Velocity on using SRE to implement DevOps. And the more I think about it, and see what people are doing, the more I’m getting worried.
The Big Lie
Just to get the easily provoked to put up their pitchforks, I don’t dislike SRE and I don’t dislike Kanban. The reason I call Kanban a “Big Lie” is because really doing Kanban correctly and getting the value out of it requires even more discipline that doing something like Scrum. But it looks so close to doing nothing new that many lazy teams out there say “they’re doing Kanban” and by that they mean they’re doing nothing, but they’ve turned on the Kanban view in JIRA for your convenience. They have no predictability, they’re not managing WIP, they’re not identifying bottlenecks – they just have a visible board now and that’s it. I strongly believe from my experience that most teams “doing Kanban” are really doing mostly nothing. There’s articles on this blog about how I make my teams I’m teaching Agile do Scrum first if they want to get to Kanban to build up the required discipline. And I’m not just a crank, David Hawks from Agile Velocity just told our management team the same thing yesterday, which brought this back to mind for me and spurred this article.
Because I’m starting to see the same thing with SRE. It’s not surprising – there was and is plenty of “DevOps-washing” of existing teams out there. Rename your ops team DevOps, done. Well, at least DevOps was able to say “it’s a methodology not a job description or group name stop it” to force deeper thought – it’s why my team at work is the “Engineering Operations” team not “DevOps”, Lee Thompson insisted on that when he set it up! But SRE – yeah, it’s a team just like your own ops team, from an “org chart” viewpoint it looks the same. So doing SRE can – and in many shops does – mean doing nothing new. You just call your existing ops team SRE and figure you’re done.
A brief personal history lesson – my last job before DevOps hit was running the Web Systems team at National Instruments, an ops team. That’s where we agile admins met, Peco and James were both ops engineers on that team! (Karthik was a dev we worked with.) We had smart people and did ops all right. We had automation, monitoring, we had “definition of done” standards for new services. You wouldn’t have to squint too hard to just call that team a SRE team and call it a day. But, I wouldn’t wish that job on my worst enemy. It was brutal trying to do ops for just 4-5 dev teams, and that’s with business support, some shared goals, and so on. Our quality of life was terrible, we weren’t empowered, and no matter how hard we tried, success was always right out of our grasp. When we actually started a team using DevOps thinking at NI after that, the difference was night and day, and we actually began to enjoy our jobs as ops engineers. I would hate for anyone to deceive themselves into thinking they’re getting the goodness they should be able to get from a DevOps/”real” SRE approach while still just doing it the way we were doing it.
I have a friend at a local legal software firm, who told me they’re going through and just renaming all the QA folks to SWET (Software Engineer In Test), whether they can code or not, and all the Ops folks to SREs in this manner. One might be charitable and say they’re leaning forward and they intend to loop around and back that up with retraining or something, but… will they? Probably not, it’s just a rename to the hot new term without any of the changes to help those engineers succeed more in their jobs!
SRE isn’t “an implementation of DevOps” if you just apply it as a name for a hopped-up ops team. Properly understood, it can be an implementation of one of the three parts of DevOps, Infrastructure As Code, Continuous Integration/Deployment, and Site Reliability Engineering. But note that reliability engineering doesn’t start with deploy to production; so much of it is Michael Nygard-esque techniques to write your app reliably in the first place; reliability engineering, in usual DevOps fashion, requires dev and ops work both way back in the dev cycle and out in production to work right. It doesn’t need to be a different team. If it is, and that team doesn’t get to decide if it takes over ops for a given app, and it’s not allowed to spend 50% of its time on reducing toil and you’re not comping SREs like you do dev engineers – it’s not SRE and you’re a liar for calling it SRE. If you don’t keep DevOps principles in mind, you’re just going to get your old ops team with its old problems again.
That’s why SRE is a Big Lie – because it enables people to say they’re doing a thing that could help their organization succeed, and their dev and ops engineers to have a better career and life while doing so – but not really do it. Yes, there have been Big Lies before, which is why I cite Kanban as another example – but even if the new criminal is pretty much like the old criminal, you still put their picture up on the post office wall.
Frankly, anyone pushing SRE that doesn’t put warning labels on it is contributing to the problem. “Well but it mentions in chapter 20 of the second book,” said someone responding to the first version of this article on Twitter. Not good enough. If something you’re selling is profoundly misused it’s your responsibility to be more up front about the issues.
The Little Problems
Now there are legitimate issues to have even with the “real SRE” model, at least the way that it’s usually being described. The Google books kinda try to have it both ways, describing it as an engineering practice (how I describe it above and in the SRE course I did for LinkedIn) and describing it as “a team that works this way.” Even among those not SRE-washing classical ops, the generally understood model is that SRE is a org/job title for a production operations team.
There’s an issue here, the problem of specialization. If you are Google scale, well, then you’re going to have to specialize and a separate ops team makes sense. But – first of all, you are not Google scale. In my opinion, if you are under 100 engineers, you are committing an error by having a separate ops team. You need your product teams to own their products. Second of all – I don’t want to make an enemy of all the lovely Google engineers out there, but is your experience with Google services that they evolve quickly and get better once they go to wide release? It’s not mine. They rot. Have you used Google Hangouts lately without it ending up with cursing and moving off to someone’s Zoom? That kind of specialization still has its downsides in terms of hindering your feedback loops that let you improve (the Second Way). Is SRE just Google-ese for “sustaining?”
I get that the Google folks say they still get feedback and innovation using the SRE model, I’m sure they do and they work hard at that, but that doesn’t change the fact that running a separate ops team is making a deliberate tradeoff between innovation and efficiency. There is no way in that you get as much feedback or improve as quickly with a separate team, you can compensate for it, but you’re still saying “look… Not as important.” Which is fine if that’s your situation, I worked at many companies with 200 abandoned apps in production and you had to do something. But “not getting there in the first place” is better.
Some of the draw of the model, and why Google is highly aligned with it, is Kubernetes itself. k8s is very complex to run drives people back a little bit to the old priest-in-the-tabernacle model of “someone maintains the infrastructure and you write the app and then you have them deploy it,” but now there’s some standards (like deploying as a container) that make that OK – I guess? But if you think reliability, and observability, are the primary responsibility of an ops team that is not involved in constructing the application, you either have deep and profound company standards that allow seamless plugging of the one into the other or you’re fooling yourself. 90% of you are fooling yourself.
At this conference I heard “Service meshes! They get you observability so your devs don’t have to think about it.” Do you not see how dangerous that mindset is?
SRE, as interpreted as “a separate newfangled ops team,” may work for some but you need to be realistic about the issues and tradeoffs you’re making. Consider whether product teams supporting their product, maybe with aid from a platform team making tooling and an enabling/consulting/center of excellence team that can give expert advice? DevOps helped us see how the “throw it over the wall from dev to ops” model was profoundly harming our industry. Throwing over the wall from dev to SRE doesn’t improve that, it’s profoundly regressive. Doing SRE “right” to compensate for this, like doing Kanban right, requires more skill and discipline, not less – be realistic about whether you have Google levels of skill and discipline in your org, eh?
SRE (and Kanban) aren’t bad, they have their pros and cons, but they are easy to “pretend to do” in some minimal, cargo cult-ey way that gets you little of the benefits. And if you think spinning up an ops team and calling it SRE is “an implementation of DevOps” you’ve swallowed the worst poison pill the DevOps talk circuit can deal to you.