SH-02PPlate 1
Payment & Travel Platforms
Sanjow
Production work at Sanjow on payment and travel platforms that several brands share, serving live customer traffic. My part: I built most of the check-in service and of the online ticket shop. On the shared platform API, as one of its main developers, I work on check-ins, payments and bookings, customer messages and database speed.
- Type
- Payments / SaaS
- Status
- Private
- A–A · 1When a check-in failed, two parts of the system each retried it three times, so one booking could be tried nine times. Now only one part keeps count: three tries.
- A–A · 2Every check-in attempt now leaves a record, even one the server cuts short, so a system in trouble can never look idle.
- A–A · 3A check-in counts as done only when every passenger has a boarding pass that we opened and checked. A "done" on screen is not enough.
- Front-ends: Check-in funnel, one deploy, every brand · 8 locales, contributor; Newer checkout funnels, a public login-link proxy, gated at the API, contributor; Member zones, single sign-on across the group, contributor; Tickets storefront, seat maps · 3-D Secure, main developer; Back-office, the operations console I joined in 2022, contributor.
- API: Platform API, Express · Prisma · Zod → OpenAPI, one of the main developers.
- State: PostgreSQL, 101 models, hand-written SQL migrations; Redis · BullMQ, 51 queues · 50 scheduled jobs; Object storage, verified documents.
- Outside: Check-in workers ×12, main developer; Payment gateways, card tokenisation · 3-D Secure; Mail · SMS · voice, webhooks in, messages out; Travel inventory · eSIM · top-ups, partner APIs; LLM, three API modules, a fallback pair in the check-in workers.
Section A–A, one check-in, from click to boarding pass. The path runs Check-in funnel, then Platform API, then Redis · BullMQ, then Check-in workers ×12, then Object storage.
- Problem
- Several brands sell payment and travel products on one stack, with live money moving through it.
- Constraint
- Every outside provider is reached through one shared service, each website serves every brand from the same code, and monitoring is not optional.
- Decision
- Websites that adapt to each brand's look, with one shared service behind them. That service is monitored, and its changes must pass automatic tests.
- Outcome
- In production, serving live customer traffic. The architecture is the drawing above; there are no public stills.
General notes
- 01Each outside service we connect to has its own folder, with a single way in.
- 02Every change to the database structure is written by hand, and its file records what was measured in production and how to check it can be undone.
- 03When the system cannot tell what state an account is in, it assumes the costly case: a failed lookup no longer shows an account as brand new.
- 04Alarms go off at levels taken from real traffic.
- 05Better a link that fails than a customer landing on another brand's site.
- 06Before the seat map was designed, four small tests recorded how the supplier's tool really behaves.
- 07Before release, every new safety check is deliberately broken to prove its test catches the failure, then restored.
- 08An hourly database search went from 818 ms to 3 ms, thanks to an index built while the database stayed open for reading and writing.
Bill of quantities · As-built 2026-11
| Item | Quantity |
|---|---|
| Prisma modelsgit show origin/master:prisma/schema.prisma | grep -c '^model ' | 101 |
| BullMQ queuesgit show origin/master:src/app.ts | grep -c 'new Queue(' | 51 |
| Scheduled jobsgit grep 'scheduleLocked(' origin/master -- src/app.ts | wc -l | 50 |
| Check-in workersawk over the exported worker map in src/workers/ac.ts, origin/master | 12 |
| Locales, one front-endthe locales array in i18n/config.ts, checkin-ai origin/main | 8 |
Stack used
Next.jsReactTypeScriptExpress.jsPostgreSQLPrismaBullMQRedisMantineOpenTelemetrySentryVitestAWS S3