Decision Library
Choose the architecture your team can run at three in the morning
Architecture decisions are made in a design review and paid for in production for years. Familiar tools win by default and fashionable ones win by enthusiasm, and neither is a reason. A Decision Room compares the real candidates on correctness, operating burden and the price of reversal.
The candidates on the table
Most architecture questions offer four or five honest paths: keep the current system and fix its worst part, extend a tool the team already runs, adopt a managed service, adopt a new platform, or build a narrow component of your own. Each one moves risk somewhere different, from a vendor's status page to your own on-call rotation.
Spend novelty sparingly
In a 2015 essay, the engineer Dan McKinley argued that each company has only a few innovation tokens to spend, and that well-understood technology is usually the better default because its failure modes are known. The point concerns attention more than age. Every unfamiliar component asks the team to learn new ways it can break.
A related habit is the architecture decision record, proposed by Michael Nygard in 2011: a short note for each significant choice, giving the context, the decision and its consequences. A team that keeps them can later tell a mistaken decision from a reasonable one that met bad luck.
Evidence that holds up
Load tests run on your own data shapes, incident histories from teams that run the same system at similar scale, the real number of engineers who can operate it, and the contract terms that govern exit. Vendor benchmarks are marketing until reproduced under your conditions, and a proof of concept is an estimate of production, never production itself.
How the room argues an architecture
Each candidate gets a Champion. The Red Teamer runs the failures that matter for this workload: a region outage, traffic at ten times normal, a corrupted message replayed a thousand times. The Craftsman asks who will be paged and whether they have the skills today, and the Grandfather prices the cheapest way back if the choice proves wrong in eighteen months.
The Verdict names the leading design, the operating assumption it depends on, and the load or staffing figure that would reverse it. The engineering lead records the Human Decision.
What to weigh
Correctness
Whether the design meets the real requirements for ordering, durability and consistency.
Time to value
How soon a working result reaches users with the team as it stands.
Operating burden
Who is paged, how often, and whether they can diagnose the failure without a specialist.
Cost at expected load
Running cost at today's volume and at three times today's volume.
Exit cost
The engineering weeks and data migration needed to leave after two years.
Security surface
New credentials, network paths and dependencies the design adds.
Where it goes wrong
Resume-driven design
A tool is chosen because engineers want it on their record, not because the workload needs it.
The permanent stopgap
A temporary fix ships under pressure and becomes the system for a decade.
Single-vendor gravity
Convenient integrations deepen dependence until leaving costs more than staying.
A queue for order events
A team of five engineers handles about 4 million order events a day and must choose how to move them between services. Option one uses a table in the existing relational database as a queue: no new system, but the database already runs at 60 percent of capacity at peak. Option two is a managed queue service at a quoted $1,100 a month. Option three is a self-run streaming platform, which the team estimates needs half of one engineer's time to operate, roughly $90,000 a year in salary. With weights of correctness 30, operating burden 25, time to value 20, cost 15 and exit cost 10, the managed queue leads. The Verdict states what would reverse it: event volume above 40 million a day, or a need to replay months of history, at which point the streaming platform's case strengthens.
Questions
- Can I paste a design document?
- Yes. Paste the text and list the candidate designs. The room reads the stated requirements, flags any that are missing, such as recovery time or data retention, and argues each design against them.
- Who signs off on the architecture?
- Your engineering lead or design review does. The Verdict is decision support, and the Human Decision with its reasons sits beside it in the Decision Passport.
Bring Argumont a Decision worth challenging.
Start with a paid 60 day pilot on your own real Decisions.