Building CDG Zig's booking flow in Compose

The booking flow on CDG Zig used to be ten XML layouts and a stack of fragments. Picking a destination, choosing a vehicle, confirming the fare, watching the driver on the map — every screen was its own little island. There was a lot of duplicated UI code, and the screens people touch on every single trip were slower to draw than they needed to be.

Why we moved

There were two reasons. The first was that the booking flow was the single most-touched surface in the app, and the second was that the new screens we wanted to add (multi-stop rides, fare estimates) did not fit the old layouts at all without forcing a lot of awkward XML into the hierarchy.

Compose was the right call not because XML is bad — it is not — but because state-driven UI maps onto booking state much more naturally when you can read the screen as a function of the booking object.

What I changed

• Replaced the five Fragment classes for the booking flow with a single NavHost and a stack of Composable screens.
• Pulled the shared pieces (date pickers, payment rows, address cards) out into a small library the rest of the app already uses.
• Moved map interactions off MapView and onto the new Compose interop layer, which let us drop the lifecycle glue that used to live in every screen.

The numbers

Metric	Before	After
Screen render (p95)	180 ms	125 ms
Crash rate (per 1k)	0.41	0.33
Lines of UI code	3,200	1,840

The crash-rate number was the part I cared about most. Compose did not magically fix the crashes — the team wrote the unit and UI tests that had been missing, and the team triaged the top Crashlytics issues. What made the change stick was the habit: crash and ANR reports get checked after every release, so a regression surfaces in days instead of weeks.

What I would do differently

I would have started with the design system, not the screens. The first three weeks of the migration were spent building the same buttons and rows that already existed in two other parts of the app. Pulling the shared pieces out first would have cut the migration time roughly in half.