For teams
Engineering decisions should name the pager
Architecture arguments are often arguments about taste, settled by whoever has the most conviction or the most seniority. The cost arrives later, at three in the morning, when someone has to run the system the argument produced.
Count the people, not only the servers
A tool's price is the smaller part of its cost. The larger part is the engineering time it absorbs: upgrades, tuning, incidents, and the knowledge that leaves when its one expert does. A system that needs half an engineer forever costs more than most license fees.
Write the decision down
Many teams keep architecture decision records, a short format Michael Nygard proposed in 2011: the context, the decision, its status, and its consequences. The format works because it captures why a choice made sense at the time, which is exactly what is lost when the people who made it move on.
The hardest part is the consequences section, written honestly before the first line of code.
Cheap to undo, costly to undo
Some choices are easy to reverse: a library behind an interface, or a managed service that speaks a standard protocol. Others bind you for years, such as a data model customers depend on or a platform whose features seep into every service. Spend the debate on the second kind and decide the first kind quickly.
An architecture choice in a Decision Room
Each option gets its own Champion, so the plain choice is argued as seriously as the exciting one. The Challenger presses on failure modes and on whether anyone on the team has run this at load. Craftsman walks the migration step by step, from the first dual write to the day the old system is switched off. Grandfather asks what will be hardest to undo in three years.
Vendor benchmarks enter the record as claims, and Inspector marks them unverified until your own test reproduces them.
A queue for 1,500 messages a second
A team of five needs a message queue that handles a peak of 1,500 messages a second. A managed service costs about $2,400 a month, or $28,800 a year. Running their own streaming cluster costs $1,500 a month in machines plus roughly half an engineer to operate it; at $180,000 a year per engineer, that comes to about $108,000 a year. A queue built on their existing Postgres database costs little to run and two weeks to build, but nobody on the team has tested one at this volume. The Verdict favors the managed service on cost and on who gets paged at night, and lists the condition that would change it: a load test showing the Postgres queue holds peak traffic with room to spare.
Questions
- Do engineers need to learn a new tool to use this?
- No. You describe the decision and the options in plain words. The room supplies the structure, and the record reads like a well-kept design document.
- Who makes the architecture call?
- The engineering lead or the team that will own the system. Their Human Decision is recorded with the Verdict, and an Outcome can be added after the first quarter in production.
Further reading
Bring Argumont a Decision worth challenging.
Start with a paid 60 day pilot on your own real Decisions.