The questions are fingerprinted to this domain — each one is on the list
because its absence has already cost something written up elsewhere on this
site.
Questions that must have an answer:
- What happens when a supplier returns HTTP 200 with an error body? A swallowed
failure here is how three features once died silently with zero errors
anywhere.
- Is
book idempotent under client retry? A timeout plus a retry must not
become two reservations.
- Which module owns this state, and who is allowed to read it? If the answer
is "everyone", the modules are welded together within six months.
- What is the blast-radius label of the rollout command — read-only, writes to
production data, or deploys? The label goes where the eye lands first, not
three paragraphs in.
- How does this go live across N repos — what is the merge order, and which
MR pairs must land in the same beat? A repo missed is an undercharge that
throws no error.
- Which figures in the document are measured, and which are estimates? Every
number carries its label; an estimate mistaken for a measurement is how a
document loses trust.
- What does the client see while this is failing — an error, or an empty
result that looks like an answer?
Entry criteria — the review does not start until:
- the problem is written down, with the numbers we actually have, each
labelled measured or estimated;
- the failure mode is named — what the user sees when it breaks, not just
what throws;
- the alternatives are real ones stated with their constraints, not strawmen
set up to lose.
Exit criteria — the review is not done until:
- every write path runs dry-run → printed plan → confirmation → apply →
read-back verification, and the tool enforces it rather than the human
remembering to;
- everything that leaves the process — logs, alerts, analytics — has an
allow-list, because egress is a boundary;
- a "what I am not certain of" section exists; it costs a paragraph and is
the difference between a document a colleague can build on and one they
must re-verify from scratch;
- if the change spans repositories, the full list of repos is written and
the merge order settled before anything ships.
There is no count of reviews run here, because none was measured — the
checklist is the deliverable.
Everything above sits on one aggregation layer: ~150 supplier codes on ~90
integrations, across 5 product lines. The map below is structural on
purpose — axes, not a count matrix.
| Axis |
Values |
| Product lines |
Hotel · Flight · Tour · Transfer · Car hire |
| Protocol families |
REST-JSON · SOAP-XML · push-cache · polled availability |
| Offer patterns |
A — one token books N rooms · B — the materialised matrix of valid room-rate combinations |
| Runtimes |
Node.js/TypeScript (the daily stack) · Python (FastAPI adapters) · Go (adapters) |
Deliberately no per-line, per-protocol counts: those live in a supplier
registry, and a number appears on this site only after it has actually been
pulled and measured.