The challenge
The company runs a marketplace where travellers compare and book taxi and private-driver transfers across Europe, in several languages, with card payments secured by 3-D Secure and a dashboard for transport providers. Every search depends on a geo service that turns what a traveller types into a precise pickup point.
That geo service was business-critical and fragile. It served a gazetteer of around fifteen million places, but it had no reproducible build, deploys were manual, and a bad release could break search for every customer. At the same time, each lookup that fell through to a paid map API added to a growing monthly bill, and there was no automated way to check that bookings and payments still worked after a change.
What we did
- A service the team can change safely. We reviewed the geo service end to end, fixed correctness and performance issues, and packaged it as a container with a reproducible build on Node.js, PostgreSQL 17 and Valkey.
- Staged deploys with a way back. GitHub Actions builds every change. Merging to staging deploys it, verifies it and runs a smoke test automatically; production deploys keep the previous release ready, so rolling back is one step.
- Database-first lookups. Answering more lookups from the platform's own data before calling the paid map API cut the number of paid place-details calls by about 47% in our measurements.
- End-to-end tests for what makes money. A Playwright suite of twelve flows covers searching, booking, prepaid card payments and the provider dashboard, split into safe checks that can run anywhere and flows that create real test data.
- Errors that become tickets. Sentry captures application errors and opens Jira issues automatically, so problems are triaged instead of lost in logs.
- Search data the team can query. Daily exports of Search Console and GA4 into BigQuery, so marketing and engineering look at the same numbers.
Results
- The geo service moved to the new containerised setup in production and is deployed through the pipeline, not by hand.
- Every change to staging is verified and smoke-tested automatically; production can be rolled back in one step.
- About 47% fewer paid place-details calls in our measurements, which directly lowers the monthly map-API bill.
- Twelve end-to-end flows protect booking and payments from regressions.
- Application errors arrive as Jira tickets with context, instead of being discovered by customers.
Stack
Docker · GitHub Actions · GitHub Container Registry · Node.js · PostgreSQL 17 · Valkey · Playwright · Sentry · Jira · BigQuery
In their words
The infrastructure LL Cloud built handles millions of requests daily with remarkable reliability. Their team operates as a true extension of ours — responsive, proactive and technically excellent.
— CEO, European transfer-booking platform