One-Way Door

Decisions differ enormously in the cost of being wrong, and effort should be proportional to that cost rather than to how important the decision feels. A one-way door is a decision that cannot practically be undone: reversing it means rebuilding, migrating at scale, breaking external consumers, or recovering information that no longer exists.

Everything else is a two-way door, and for two-way doors the cost of deliberation frequently exceeds the cost of being wrong — so the correct response is to try, observe, and adjust.

Why it matters

Both failure modes are common and both are expensive. Organisations agonise over reversible choices — which framework, which library, which internal structure — burning weeks on decisions that could have been settled by a two-day experiment. And they make irreversible choices casually: a data model decided in an afternoon, an identifier scheme chosen by whoever wrote the first migration, a public contract shipped without review.

The distinction is the highest-leverage filter available for allocating architectural attention, and applying it consistently produces both faster delivery and better outcomes on the decisions that matter.

Implementation patterns

• Ask "what would it cost to change this in eighteen months?" rather than "could we change this?" — the second question always answers yes and tells you nothing.
• Classify explicitly in design reviews, and spend the session on the one-way doors. A review that gives equal attention to a library choice and a data model has misallocated the room.
• Identify the closing window. The most important category is decisions that are reversible now and become irreversible with scale or adoption — and the useful question is when that window closes, which is frequently much sooner than assumed.
• Convert one-way doors into two-way doors wherever possible: a boundary around the irreversible component, additive-only evolution of contracts, a proven migration path.
• Capture more information than currently needed. Storage is cheap; information not captured is gone permanently, which makes a lossy data model the archetypal one-way door.
• Delegate two-way doors to the people closest to the work, since escalation makes them slower and rarely better.
• Record one-way doors with their assumptions, since those assumptions are what will later be questioned.

Industry example

The framing is most associated with Amazon's Type 1 / Type 2 decision language, where the explicit warning is that organisations tend to apply heavyweight Type 1 processes to Type 2 decisions as they grow, producing slowness without producing better outcomes.

The pattern recurs in every published architecture reversal. The Prime Video pipeline consolidation was possible because the decomposition was reversible — code and deployment topology, not data. The decisions that appear in postmortems as unrecoverable are consistently the other kind: an identifier scheme that propagated everywhere, a schema that discarded a field someone later needed, a public API shape that thousands of integrators depend on.

Failure scenarios

• A data model that discards information, which cannot be repaired at any price.
• An identifier scheme adopted casually and then embedded in every consumer's data.
• A public API contract shipped without review, permanent the moment someone integrates.
• Missing audit trails where they will later be required, since the history cannot be reconstructed.
• Treating every decision as reversible, and discovering at scale that the migration is a multi-year project.
• Treating every decision as irreversible, producing paralysis and a reputation that slows all architectural involvement.
• Missing the closing window on a decision that was cheap to change last year.

Trade-offs

The classification is a judgement, and it can be wrong in both directions. A decision assessed as reversible can turn out to be load-bearing; one assessed as irreversible can turn out to be straightforward to change once someone tries.

There is also a cost to preserving reversibility: boundaries, abstraction layers, additive-only schemas and extra captured data all have real ongoing costs, paid against an option that may never be exercised. Preserving every option is its own form of over-engineering.

The trade is analysis effort and optionality cost against the risk of foreclosing something valuable. The practical resolution is asymmetric: be generous about preserving reversibility where the cost is low — an extra column, a boundary, an additive change — and accept irreversibility deliberately elsewhere, having said so out loud.

Interview question

"Take our current design and tell me which three decisions in it are one-way doors. For each, tell me what it would cost to reverse in two years, and whether there is something cheap we could do now to make it a two-way door."