CTO Workbench

Six small decision models for the questions that recur in architecture reviews, each connected to a shared failure-economics layer. Change the assumptions, inspect the technical boundary, then test whether the proposed resilience is justified by the SLO and cost of downtime.

Want to model the dependencies between those boundaries? Open CTO Workbench V2 and simulate failure propagation.

AYUSH / CTO WORKBENCH

Turn architecture assumptions into numbers.

Six small models for conversations that become expensive when they stay vague.

Local calculation
MODEL 01

How much concurrent work must the service carry?

DECISION SIGNAL122

required in-flight capacity

Peak traffic
750 req/s
Base concurrency
90
Load-test floor

Prove at least 122 concurrent requests without violating the latency objective.

FAILURE ECONOMICS

What level of resilience is justified?

Connect this technical signal to the SLO and the cost of downtime.

ANNUAL DECISION+$15,600

net protection value

Allowed downtime / year
8.76 h
Error-budget exposure
$43,800
Avoided failure cost
$30,000
Control cost / year
$14,400
Break-even avoidance
0.72 h / incident
Resilience is economically supported

The modeled avoided failure cost exceeds the proposed control by $15,600 per year.

Use contribution margin, contractual penalties, recovery labour, lost productivity, and customer harm—not topline revenue alone. This is a decision model, not a promise that redundancy will prevent every modeled hour.

Assumptions and limits

Uses Little’s Law with p95 latency as a planning proxy. Real concurrency follows the full latency distribution, traffic correlation, queues, and dependency limits.

Use the result as a question

These models are deliberately transparent. They expose the arithmetic behind a decision, then state the assumption that most often breaks in production. A green capacity signal is not an investment decision. The Workbench also calculates permitted downtime, error-budget exposure, avoided failure cost, annual control cost, and the break-even recovery improvement.

Useful follow-up questions are concrete: Does the p95 hide a dangerous tail? Can every replica fill its connection pool at once? Are failures correlated? Can the queue drain while new work still arrives? The answer should become an observable production invariant.

Continue the investigation

Have a recurring engineering decision that deserves a tool? Send it to me.

>