The away team model at Amazon (2022)
pedrodelgallego.github.io
pedrodelgallego.github.io
You’re not supported but being expected to deliver by your “home” team. Meanwhile the away team has no obligation to help you. Can spin their thumbs and give you three day turn arounds on code reviews only to move the goal posts and/or hold you to some unattainable standard. For example “all integration tests need to pass” while the integration tests haven’t even STARTED correctly in over a year.
Not to mention the increased visibility and communication that comes with now effectively having 2+ managers you have to report to and keep in the loop.
FUCK. THE. AWAY. TEAM. MODEL.
I’ve lived through it and it seriously burned me out harder than anything else in my life.
I think you meant the host team.
The author of this piece is an idiot who has drunk the Kool aid and has very clearly never performed away team work.
I have personally wasted a year of my time on away team work, have supported several away teams, and most importantly: operated systems/built on top of stuff built by away teams. All of this at Amazon, specifically.
Having personally experienced every engineering facet of away teams, I feel as qualified as anyone to say this: away teams are a stupid concept that hurts everyone except management.
You spend way more time to deliver a product you don't really understand and will have a hard time spinning on promos, which will almost certainly be a nightmare to maintain and operate, all because your employer is too cheap to allocate headcount for surprises and your team took a dependency on another team years ago so they have you by the balls.
It is easy to slap together crap and toss it over the wall as good enough, and then, the "hone team" is stuck with supporting it.
For example: Imagine building the feature on a O(n^2) data structure, it'll work fine to start, but long term, we'll need a O(log(n)) structure. I wonder who has to do the work on the "slow performance" bug, and data migration.
These concerns aren't abstract. They are very concrete. There will be landmines left behind. Even when we try not to leave them behind we do as developers. Without responsibility... it would be even worse.
This is also why I don't like having one team do all the Greenfield work, and one do all the fixing.
The away teams should be writing “userland” code within the home teams well defined guardrails. The home teams act as OS devs for their service mostly write “kernel” code for their service that empower the away teams.
Usually the “home team” are functional specialists. They are working on some piece of infrastructure or maintaining a service. The “away” team OTOH is building a feature and needs to touch a few dozen codebases
It should be obvious which code is kernel and maintained by the service team, and which by the away “app” team. The away teams should be maintaining, on-call, and responsible for their code running within the context of the home teams service. Which hosts many such “apps”
https://softwaredoug.com/blog/2023/07/08/why-search-orgs-fai...
In big orgs, if you have time to set up baby fences (or had them from the start) you have enough free time to build the feature yourself.
No one who hosts an away team is prepared to do so
For example, this worked really well for us at Shopifys Discovery team at for moving fast on search relevance and recommendation systems. To clarify how to build and test search experiments really fast. I would say it’s the default model/culture at Shopify.
(Not away team per-se but more the “OS” model)
I misunderstood your comment, I thought you were suggesting the home team build the guard rails in response to another team's request. I don't disagree that the pattern you're describing can be effective if the home team was always expecting other teams to work in their codebase.
But that's not what the away team model is, the away team model is for unexpected cross team work. E.g. if the guard rails are too restrictive and the home team doesn't have time to expose the new functionality to "userland" code so the other team needs to work with "kernel" code to do so.
What I suggest I think is different than away team (which feels a bit like swoop in and crap on host teams codebase). Just saying I prefer it to away team.
Fundamentally it’s just a hard problem most large orgs don’t do that well
Plus, you lose a ton of time implementing features outside your area.
Your away team didn’t know that those 5 functions had to happen in order, due to a timing issue nobody could figure out, so your code to change how a panel animates is now breaking unrelated calculations somewhere and causing data loss. You blame the home team for having bad architecture, but then learn a different group of three different away teams set that up and nobody can figure out how to clean it up because of the ripple effect bugs. There be dragons
Source: was sr engineer at Amazon
The airline business model is therefore maximizing the percentage of time an airplane is in the air generating cash.
Boeing responds to this by expending a lot of design effort towards minimizing required maintenance, and making maintenance quick and easy to do.
But, sometimes, an airplane develops a fault that the airline mechanics cannot fix. Enter the AOG team. The AOG team is composed of the best engineers and mechanics, and they are on a moment's notice to fly anywhere and get that AOG back in the air ASAP.
I heard Microsoft has a similar team to go out and fix a customer's software. It's one reason Microsoft is able to sell expensive enterprise software in the face of free software. It's worth the money corporations pay to Microsoft to "come out and fix it!"
About 20 years ago, they were called Rangers, modeled after the Army Rangers, soldiers who could venture out into enemy territory with no support and accomplish missions out there.
About 15 years ago, the program evolved into Microsoft Certified Masters, and it included both internal staff and specialized outside consultants. (I was one of about 200 MCMs worldwide for SQL Server.)
If they called themselves SEALs though, I’d be laughing my ass off.
however giving enough details might be prohibited by NDAs and whatnot, or even render companies as toxic/disfunctional so nobody will ever join them :)
it's a farce. Interesting idea in theory, terrible in practice. The away team incentive is "launch as fast as possible at all costs" and that goes about as well as you'd expect.
We constantly deal with shit code and loose rules. Given that many away teams work in different time zones AND Amazon gives them carte blanche to get their work done, the host team can't manage them effectively.
So yeah. On one hand management is like "push back! Make them follow your standards!" while also saying "Don't dedicate time to away teams, you have your own deliverables."
It's a major drain on the home team and can cause a lot of unnecessary stress.
These opinions are my own, obviously, lol.
* You have to work with an, at best, passive team, at worst, actively hostile team. * You need to work against very tight deadlines and a lot of pressure in a codebase that you know nothing about, and need to ramp up in quickly.
* You spent weeks (months?) developing expertise in a system you will probably never work again.
It is a terrible idea, but to be honest I can't think of another way to deal with the problem that it is trying to solve (mismatch between agendas of different teams).
Since away teaming is an accepted technique, it's usually considered as a way to solve these issues.
I guess we would need to completely give away the concept, and just accept that some thing won't get done unless they are important enough to keep two distant teams aligned (in other words, they are important enough for someone high in the hierarchy).
Open-source model works for libraries, or small changes in service code (where the maintenance burden is trivial). But it doesn’t work for complex services with high maintenance burdens
This comes at the cost of much slower progress.
The worse case scenario is when the scenario you given: the "pet project," where a team goes nuts on a codebase they aren't maintaining, in which case it's effectively out-sourcing maintenance and tech debt.
Ideally, the home team leadership has a lot of leverage to push back on work by the away team, which means away team work is slow but at least it gets done.
I saw no sign at all of the away team compromising on quality to meet deadlines.
It's worked for the last year or two semi-effectively, but I am loving this discussion for reminding me of the downsides/risks.
- Use as much as you can of what they already have.
- Talk to them, but don't waste their time.
- Follow their patterns.
- Understand at the end of the day its their domain.
- Choose your advice carefully.
- Try to throw a little extra gratitude their way. Organizational visibility, code reviews, documentation, mentorship, whatever feels appropriate and not over-stepping.
- Put extra effort into handoff documentation.
> ...it should not be the first option for developing necessary changes to a system dependency. Away team work is complex and inefficient.
Is very telling. I've worked in a similar model, and I'd say this is basically the important part for everyone wondering how well it works.
The main thing to note is that the arrangement should be of last resort, when the crushing bureaucracy means it's either away teams, or nothing at all.
> Away teams accelerate software delivery [...]
So while it is compley and inefficient, it also accelerates delivery?
Also from the comments here I get that away teams are mostly stuck in meetings. So I find it interesting to use away teams as a solution where crushing bureaucracy seems the main issue.
Does it work? Technically, yes. Is it sustainable for the engineers on both sides? Not in my eexperience.
>> An away team is a self-sufficient engineering team that works on code owned by another team (the host team) to deliver features they are dependent on.
This is fundamentally wrong. The point of the away team is to get support from the host team as they work. It is not self sufficient by nature. Otherwise, this is just a team randomly contributing to a codebase you own.
As someone who made a career out of being loaned out to other teams as an 'away resource' at Amazon, I've struggled to find a similar level of mental stimulation after leaving for other Big 4 companies. Being a loaner engineer meant something new to work on every few months, no maintenance or oncall burden, making lots of new connections, and just generally there being always something interesting/intense to do.
Also, if you're looking for a thrill, I would highly recommend going after cloud-based infra providers. There's never a shortage of unique problems there.
As for my fellow engineers- I don’t know anyone else who did this regularly so I’m not sure
Since these are Amazon employees, I would expect them to experience the downsides of being a consultant, such as alienation, and the downsides of belonging to a product organization, such as extreme pressure.
> As organizations grow, engineers’ code intertwines with other engineers’ work, making it difficult for teams to complete their features independently.
The current consensus in the industry is that it should be the opposite. To grow in a healthy way, teams should be as decoupled as possible.
> The away team model is a powerful and flexible mechanism for removing dependencies between teams. However, it should not be the first option for developing necessary changes to a system dependency.
Did I miss something? You introduced this team model because there was a dependency, but you never removed it. This post is full of contradictions.
It's just another article by a manager expounding the virtues of a managerial fad to score points in his org. Sounds great to his peers, sounds like a nightmare to ICs.
> It doesn’t affect ownership. After the away team finishes its changes, the service ownership remains with the host teams. The away team model also reflects how Amazon weakly assigns code ownership to service teams.
The article quotes Martin Fowler to back this up:
> Weak code ownership: Developers are allowed to change modules owned by other people. Module owners are expected to take responsibility for the modules they own and keep an eye on changes made by other people.
However, this quote seems taken out of context. It is about individual developers making changes. When you have an entire team doing this for (months?) at a time, can maintainers really keep up with the changes and support them in perpetuity in addition to their other responsibilities?
2. Make sure the amount of process matches the amount of complexity. If the change is very simple or only a few lines, make it trivial to submit code changes that will pass review. (So have really good automated style checks/test suites.) If the change is complex, make sure you’re doing design reviews before implementation, to ensure the away team doesn’t waste time with bad implementation strategies
Clear goals and acceptance criteria.
So one management takeaway is this type of work seems ideal for highly competent outsourcing.
Offer a bonus for early delivery, and hire a small army of cheap out of school eager beaver BAs to just go around frantically documenting all the tech debt being incurred.
https://en.wikipedia.org/wiki/Brooks%27s_law
Also, out of curiosity, does AMZN have red shirts on the away teams?
Per that wikipedia article, Brook's law is more about trying to use more people to make a late projects faster, than projects in general starting from scratch. Knowledge transfer (KT) mid-project, and knowledge transfer as the project inevitably evolves in design/implementation is the challenge due to the need for N people to communicate which can lead to n-squared communication paths which fails to scale.
Brook's law also doesn't really apply if you have some segmentation so the "Away" team and "Home" team have semi-clear boundaries of responsibility allowing them to work independently. The architecture of whatever you are working on has to have some internal boundaries where the "Away" team works on something that generally doesn't require change (or awareness of changes) from the "Home" team and vice-versa. It is presumed in this Home/Away scenario that the "Home" platform/system is fairly stable for the most part and the "Away" team is just plugging in features/new-capabilities and there is some degree of organization and pattern in the core system already.
Ability to scale with multiple Away teams is based on 1) the degree of that segmentation, and also 2) any reduction you can make in amount of upfront KT needed and judgement-based decisions that otherwise require interaction along the way and during the later review stages.
The KT and communications along the way and review process will still cause some Amdahl's-law like situations in overall scalability but you can avoid the "n-squared" Brooks phenomena.
In my experience, the problem with the "Home"/"Away" approach I've seen so far (from a manager/architect perspective more than IC) is that Away teams tend to be project based, so you spend a lot of time KT'ing between Home and Away, and when the project end, that KT goes to waste and is lost. If you have a steady stream of incremental work done by away teams, management should arrange to create a partner permanent team divided along some architectural boundary ideally (make the org match the architecture) or if needed, along some different mission statement. This is analogous to what happens when you have employee turnover on your team.