You have limited time to deliberate. Which decisions on a new platform deserve it, and which should be made quickly?

What the interviewer is testing

Whether you allocate deliberation by reversibility rather than by how interesting the decision is.

The one-way doors

Data model in a system that will hold millions of records with many consumers. Changing it later means migration, dual-write, backfill and coordinated consumer changes.

Public API contracts with external clients, who upgrade on their own schedule or not at all.

Service boundaries, since redrawing them means redistributing data and ownership.

Identity and authentication foundations, which everything depends on and which are extremely disruptive to change.

Encryption and key management, because re-encrypting at scale is a programme.

Data location and partitioning, particularly where residency applies — retrofitting it into a globally replicated system is close to a rebuild.

Anything involving data that leaves your control.

The two-way doors

Internal library choices. Frameworks within a single service. Deployment tooling. Most infrastructure configuration. Monitoring vendors. Even the cloud provider for stateless compute is more reversible than people assume — it is the data that is sticky.

The move that matters more than classifying

Convert one-way doors into two-way ones. An abstraction over the datastore, a versioned contract, a strangler facade, a pilot in one region rather than a full commitment, a feature flag.

An architect who consistently makes decisions reversible provides more value than one who deliberates well, because the organisation gets to be wrong cheaply — and it will be wrong.

What a strong answer adds

Recording, for each significant decision, what would change your mind. That converts a decision into something reviewable against evidence rather than a position to defend, and it is what distinguishes an architecture that learns from one that accumulates commitments.

Common weak answers

Deliberating over technology choices, which are usually reversible. Treating everything as consequential, which is how organisations become slow without becoming safer.