Philip Reynolds bio photo

Philip Reynolds

Director, Engineering @ Workday. All thoughts, opinions expressed are my own

Twitter LinkedIn

Running Cross Functional Service Teams


This is, in written form, something I talked about at Velocity Europe 2014. Based on feedback from the talk, I’ve decided a blog post might be a nice way for people to refer back to this.

The idea of cross functional teams is that the team can be autonomous to do what it needs to do to develop, build, test, qa, deploy, release, monitor & maintain it’s product or service. This means one team which has the entire skillset required to own the full lifecycle of it’s deliverable. You need to either hire or grow a skillset that is often foreign to traditional development teams with respect to areas like deployment & monitoring.

From a technical perspective, if you’ve got a small product (like a small LAMP stack in a startup), then this can be straight-forward. As your architecture grows, you need to figure how to compose your application so it becomes more manageable and maintainable. One popular way of doing this is breaking down your application in to services ala a service orientated architecture.

The topic of microservices has been particularly prominent recently and the principles here align very nicely with microservices if that is how you choose to grow your application.

I’ll continue to use the word “service” as the item to organise around. If you’re small, that service may be your entire application.

This idea is particularly suited towards cloud software where you run your own infrastructure. If you are shipping mobile software to a 3rd party device, or selling on-premise software to your customers this article may not as relevant to you.

Typically, the three main protagonists in any cloud software company for building and running the software are Developers, QA & Operations. Increasingly we start to see Security (SecOps / AppSec / Infosec) as part of this story also. Historically, those organisations have been siloed with their own reporting and management chains. Developers would develop stuff, often times working closely with QA (or not!) and then the proverbial “throw it over the wall” moment would happen as they deliver to Operations. However there has been a tremendous push recently towards “DevOps” led by Patrick Debois. It will be difficult for me to write this without repeating some of the DevOps mantra, but I’ll do my best to try and focus on the specifics of running a cross functional team. Most of the DevOps principles are very abstract and I hope to espouse some more of the concrete benefits I’ve seen.

As we develop more rapidly, deploy more often and maintain ever larger pieces of infrastructure our inefficiencies become exacerbated. The goal we are striving towards is the ability for the team who is building a service to own their own destiny and be autonmous where they need to be.

What’s wrong with the traditional model?

Firstly, most of what I’m defining refers to a specific landscape. These symptoms appear in cloud software companies who have multiple development groups producing part of the product(s) that typically lives on an ever growing and more complex infrastructure. Separately, you have an operations group which is responsible for running the stuff the developers have built. They’re often also responsible for building the stuff that the stuff the developers built runs on.

Over time your operations group grows bigger and bigger. With a bigger application their application knowledge spreads thin. You’ve introduced a throw it over the wall mentality where developers introduce new things that are decided on in a sprint that they don’t understand the full consequences of in production. You’ve got lack-lustre monitoring with an ops team clamouring to figure out what to monitor where. Typically after it goes in to production. Your development teams & managers need qualifications in project management to now manage all of the cross team dependencies when they have new stuff to deploy or have to make significant changes, particularly where there’s deployment impact. Those cross team dependencies often manifest as meetings and as a development manager you’ll find yourself in meetings constantly with managers from other groups talking about what is it you’re putting in to production and how that works.

The key metric for cloud software is delivery. What do we mean by delivery? Running in production. You’ve given developers autonomy over everything that happens before they deliver (build, test etc.), but their delivery channel is non-existant or extremely manual. Delivery becomes a path fraught with peril.

Another interesting part of this is scaling the organisation. As the development & operations organisations grow, resourcing becomes much more tricky. The development organisation often grow independently from the operations organisation. In theory, both could equally outgrow the other but. All anecdotal evidence seems to point to operations as a bottleneck in most companies as opposed to development. It feels like that should only happen some of the time given the law of averages, so something else is at play here. Seems like we’re playing with a stacked deck. Stacked against operations.

New folks in operations start to have increasingly less familiarity with the systems. They don’t have all the tribal knowledge of the older folks and as the system grows it becomes increasingly difficult to understand all of the moving parts together.

The siloing of the teams means often developers and operations folks only ever talk when something new is about to get in to production or something is on fire. Building a relationship in emergencies is certainly a bonding expeirence, but then the only way those relationships get built is if you have a lot of fires. Don’t think that’s something we want to encourage.

Simply, done isn’t done until it’s delivered. If a team can’t own ‘done’, then they don’t own their service. You now have a shared ownership model between operations and development and to coin a cliché, “if everyone owns it, no one owns it”. This manifests itself in many many ways, most of which are covered more appropiately in other devopsy related topics you can read elsewhere.

Embed in the scrum team

You build it, you run it” was the phrase coined by Werner Vogels to describe Amazon’s model. This has more benefits than just empowering the development team. It starts to make operational issues front and center. It forces them to understand their service running in production more but likewise that allows them to make better decisions.

The principle of least effort suggests that once you’ve established an organisation that has now introduced significant friction in to the delivery process, that your developers, who are trying to deliver, will often choose paths that reduce their friction even if that’s not the best path. That can often mean making bad architectural choices rather than going through the slog of introducing more artifacts which complicates deployment and makes them end up in meetings. Suddenly, adding something new to the mix becomes a tradeoff discussion as you talk about the logistics of co-ordinating that.

So, here’s my proposal. Embed all of the knowledge required in the scrum team. Set that as a goal. The scrum team needs to own all aspects of their service.

Delivery is a feature. Upgrade is a feature. Performance is a feature. Monitoring is a feature. Uptime is a feature.

If you do not own those areas in your scrum, then those features are owned outside of your scrum team. You’ve now created dependencies. Uh oh.

Dependency vs Redundancy

There is a natural inclination for every engineer to try to re-use as much as possible. Everything we’re taught in software engineering lends itself to re-using code. Well written, well designed code following DRY principles.

The engineering groups natural inclination is, “Shouldn’t a build team take care of our builds? Shouldn’t a deployment team take care of our deployments? Shouldn’t a monitoring team take care of our monitoring?”. We then centralize those resources. Organisational efficiency at it’s finest. Those are specialized skills that are consolidated within those groups. Easier to hire for. Easier to manage. Dedicated manager. On-paper perfection.

Unfortunately, reality bites and, in most cases, no-one really understands what we’ve created from a project delivery standpoint. First of all, you need to accept that your definition of done, is delivered. Now start to draw the dependency tree of teams that it takes to deliver small features & changes for just one engineering team & you start to see the challenges. That dependency diagram continues to get more and more complicated as the organisation grows. Typically those same groups have multiple other groups that depend on them also. Once you end up in this situation you end up in resource contention hell. A lot of developers can actually relate to this and have experienced this intimately in their job. It’s called Dependency Hell. Who knew, it existed within organisations too!

So, the alternative to dependency is redundancy. We end up with multiple teams potentially inventing the same solution to a similar-sounding problem (e.g. caching services, queueing services). So, the idea is not to deal with redundancy up-front by trying to eliminate everywhere there is redudancy. The idea is, to bludgeon more development metaphors, refactor as and when you see it.

Embrace redundancy! Then figure out how to remove it long term. Don’t make removing redundant services or functions dependencies on delivering. Embrace, potentially, reinventing a slightly less-round wheel. You can get to make it rounder over time.

So, hold on a minute, you say, if we brought this thought to a logical conclusion & every team did everything themselves, they’d run their own IT infrastructure, build data centers, build their own laptops. That’s obviously not a runner. How do we know when to choose redundancy over dependency? Simple answer… when it’s a problem. When it’s stopping you getting things done.

The goal should be to minimize as many external dependencies as possible.

Organisational Growth

One of the side benefits to traditional development teams now owning more of their delivery pipeline is headcount planning.

Often, resources that own different parts of the delivery chain live in different parts of the organisation. When that happens you now have complicated resourcing strategies as different parts of the organisation grow. Inevitably, you end up with capacity constraints somewhere and inevitably those constraints end up on the teams that are the most contended i.e. have the most dependencies.

Owning your build & delivery chain and minimizing those external dependencies makes headcount planning a lot easier. The dev team owns the headcount, one manager makes a decision and a business case. The decision therefore can be made at the closest point to the actual resourcing problem - at the team level.

The people

The last set of benefits I’ll talk about is on the people side.

The first and most important trait of having all of this knowledge on one team is growing a deeper understanding of the system with your engineers. If your delivery chain, build pipeline or your infrastructure is complicated and/or nuanced this becomes more important. Having the engineers understand these areas allows them to dive deep in the belly of the beast, where necessary. Even if you only have a few people who can do that, that knowledge tends to spread when it’s within the team.

Second on my list is autonomy. Suddenly, when you own more of the system, you are free to make different choices. You genuinely become more agile. If you need to upgrade to a newer version of your language framework, test harnass etc. you can do that. Turning on a dime is difficult in any software company, but owning more of your process is the key to make sure you’re not trying to turn the titanic. Co-ordination with other groups is difficult. One of the consequences is that you find over time, people hesitate to suggest things that require large amounts of effort. Now you’re free to try new things, to experiment and iterate towards better solutions for your builds, delivery pipeline etc. When small changes take large amounts of co-ordination and logistics, iterating towards a solution is always painful & risky. This type of perceived red-tape actually starts to penetrate developers well-being. Moaning & griping can often happen as teams complain about how long simple things take to do. Let them own it.

Last is cross training. You’ll find that having e.g. ops and dev folks on the same team means knowledge starts to filtrate between the two groups of people. The developers and ops folks often end up blending in to the same role where it can often become difficult on the surface to even know you have different skillsets within the group.

As we scale, we figure out new and imperfect practices that hamper our ability to grow & sell product. Sometimes these changes are fundamentally different ways to execute and require huge buy-in. Don’t underestimate the sea of change within your organisation required to make this happen.