CSE Roadmap

Core tier · about 4 hours · 8 parts

Part 6 of 8 · 3 min read

Teams and ownership

The style only works if the organisation matches it.

BM gives a whole chapter to this, and the reason is Conway's law, which page 3 introduced: a system's design copies the communication structure of the organisation that builds it. (A quick reminder: four groups on a compiler give a four-pass compiler.)

BM shows how this plays out. A single team in one place talks in small, frequent steps. That matches the fine-grained calls inside one service. If two teams in two time zones both change one service, talk becomes slow and coarse, and each person feels less safe about a simple refactoring. BM then describes what happens. Either people find ways to cut the cost of coordination, or they stop making changes, and that is how large, hard-to-change codebases arise. In one project BM saw, two sites that shared a service slowly split the work, and the code ended up with a coarse interface exactly where the teams met.

BM's advice has three parts.

  1. Give each service to one team, ideally in one place. This is service ownership: the team may change anything inside the service, provided it does not break its callers.
  2. Let that team own the full lifecycle: requirements, building, deploying and running. A team that must run its own code writes code that is easy to run. There is nobody to throw it over the wall to.
  3. Keep teams small. BM cites Amazon's two-pizza rule: no team larger than two pizzas can feed. Amazon built the tools behind Amazon Web Services partly so these teams could work alone. Netflix, as BM tells it, built small independent teams from the start so the services would end up independent too. This is the inverse Conway maneuver from page 3: design the organisation for the architecture you want.

Service ownership

Service ownership means one team is responsible for a service from change to production, and may restructure its inside without asking others.

Example: the payments team picks a new database for its service, and the shop team learns about it only from the release note.

BM also names why teams end up sharing a service: it is too hard to split, or the team is organised by feature and each feature touches many services. Both are warnings. Shared ownership gives back the coordination cost that the style was meant to remove.

A worked example (made up to show the idea). A company has 12 developers in three teams, 4 each. They choose 30 services. Each team now owns 10, which is more than a team of four can run well. The fix is not to ask for more people but to merge to about 9 services, 3 per team. The service count is set by the people available, not the other way round.

Check yourself: 1. What does Conway's law predict for one service owned by two teams that rarely meet? 2. Why does BM want the team that builds a service to run it too? 3. In the made-up example, why is merging services a better fix than hiring?